New to AI in public services, or using this site for the first time. No technical knowledge needed.
Start with: The four principles and the 2-minute review triage
Then: Glossary and FAQ, transparency
Public edition · one portal · open to everyone
One trusted route to discover, assess, implement, govern and get value from AI in public services. Governance follows what the AI can see, create and do, whichever supplier made it.
Four non-negotiable principles
Spotlight · new in 3.3.0
The agent governance playbook is now part of this portal
Chapters on criticality and tiers, registers and transparency records, oversight, cost and value, people and adoption, a live-pilot method and the delivery documents behind it, with one glossary and one contents drawer for everything.
Task-first entry points, because people arrive with a decision to make — not an interest in our information architecture.
Choose a level, then a route below. Every route and task is marked with its level.
New to AI in public services, or using this site for the first time. No technical knowledge needed.
Start with: The four principles and the 2-minute review triage
Then: Glossary and FAQ, transparency
Using, buying or approving AI in your work, and need to show it is governed.
Start with: H0–H5 review model and the evidence pack
Then: Transparency record, resources, signals
Building, assuring, integrating or reusing this guidance, including in code.
Start with: Toolkit and the metadata schema
Then: Authority tiering, open data, the repository
"I don’t want to get this wrong."
Plain-language guidance, safe defaults, confidence prompts, buddying and clear role limits.
Take the 5-minute triage, then use the declaration generator.
Journey: Human review triage → Evidence pack → Code of practice
"Give me the fast, correct route."
Direct links to standards, patterns, conformance packs, test scripts and evidence automation.
Jump to the toolkit and copy the metadata schema.
Journey: Toolkit → Metadata schema → Tools
"Show me what to challenge."
Source-linked evidence, issue triage, risk decisions and assurance history.
Use the H-level table and reviewer checklist.
Journey: Reviewer checklist → Authority tiering → Signals
"What am I signing?"
Risk overview, approval gates, residual risk, maturity dashboard and public transparency records.
Generate the transparency record and check the MoSCoW roadmap.
Journey: H0–H5 review model → Transparency record → Roadmap
"What will I be assessed against?"
Published expectations, evidence formats, accessibility and security obligations, exit terms.
Filter resources by procurement and security.
Journey: Resources (procurement, security) → Catalogue metadata → Assurance
"How is AI being used on my case?"
Plain-English explanation, published records, how to challenge a decision and get support.
Read the transparency and redress commitments.
Journey: Transparency → Glossary and FAQ → Contact
"What is changing, and does my data estate let us govern agents?"
Weekly signals with sources, the agent registry pattern, catalogue metadata for AI assets and the review model.
Start with this week's signals.
Journey: Signals → Agent registry (TK-03) → AI asset metadata → Review model
Answer six questions and get an H0–H5 level, the reviewers required and the evidence to keep.
3 minutesBeginnerGenerate a copy-ready record of sources consulted, checks made and who approved what.
5 minutesIntermediateFlag protected characteristics, access needs, assisted digital and exclusion risks early.
10 minutesIntermediateProduce an ATRS-shaped record describing the tool, its purpose, data, oversight and risks.
5 minutesAdvancedScore harvest-now-decrypt-later risk against NCSC migration milestones.
ReferenceIntermediateSearch 68 tiered laws, standards, guidance documents and tools.
Not sure whether to use AI at all? Ask three questions first: is there a simpler non-AI way to meet the user need; can you explain the decision to the person it affects; and could you defend this use at a public inquiry? If any answer is no, stop and seek advice.
Contents · 76 chapters in 12 parts
The agent governance playbook takes a team from “can I use AI for this?” to a governed, monitored and eventually retired AI agent. Open a part to see its chapters, or use the contents drawer.
Part 1 of 12 · 5 chapters
The guidance for deploying AI agents in the public sector, sequenced into one route — and the operating capability that turns experimentation into governed services. Crown-copyright sources are the government's own; every illustrative figure is labelled.
Eight stages from mandate to governed operation. The last stage feeds the third: every review either re-authorises the agent, changes it, or ends it.
Usage and cost Who uses it, how often, what it spends
Accuracy and drift Evaluation results against the baseline
Risk and incidents New risks, near misses, escalations
Value realised Benefits claimed against benefits evidenced
Review or retire Re-authorise, improve, restrict or decommission
The hierarchy of AI-related guidance a public-sector team works inside. This guide is a proposal: it interprets and sequences the guidance above it, and it does not replace or override any of it.
OECD AI Principles, EU AI Act risk tiers, ISO/IEC 42001 management standards.
Data Protection Act 2018 and UK GDPR, Equality Act 2010, Procurement Act 2023, the AI Opportunities Action Plan.
AI Playbook for the UK Government, Service Standard, Technology Code of Practice, ATRS, Data Ethics Framework, NCSC secure-AI guidelines.
Organisational AI policy, data classification, security standards, delegated authority and spend controls.
AI Capability Governance and Implementation — 69 chapters, assessment tools and register schemas that turn the guidance above into delivery.
Proposal
Agent register entries, ATRS records, assurance evidence, instruction blocks and cost models.
Every source in this app lands on the same point. The AI Playbook makes skills and expertise a principle in its own right. The Service Manual builds the multidisciplinary team into the Service Standard. NCSC wrote its secure-AI guidelines for data scientists, developers, managers, decision-makers and risk owners together. The roster is in Teams; the readiness barometer is in the Observatory.
Responsible AI · start here
Start from your task. This part of the playbook routes you to the proportionate tool, template, assessment or support. Governance follows what the AI can see, create and do, not which company made it.
Quick route selector · five questions
Your route
The route appears when every question has an answer. “Not sure” is a valid answer.
A first sort, not an approval. Any route other than an approved tool leads to a registered use case.
Generative AI predicts plausible output. It can be fluent and wrong at once, so outputs need checking in proportion to what depends on them.
A product can use several models from several providers, sometimes chosen automatically. The model can change without the product name changing.
Grounding feeds organisational content to the model. Whatever the index can reach, the AI can surface, including titles, filenames and metadata.
An agent plans steps and uses tools to act: send, update, create or call other agents. Acting changes the risk boundary.
Oversight works only if the reviewer has the competence, the evidence and the authority to challenge, override or stop.
The same model reached directly, through a cloud platform or inside another supplier’s product is a different governance object each time.
All colleaguesFind guidance, approved tools and practical help
Leaders and managersMake informed decisions, manage risk and value
Risk, security and complianceAssure, monitor and advise
IT and digitalEnable, integrate and manage
Data and informationEnsure data is ready and well managed
Commercial and procurementManage suppliers, licences and costs
Accessibility and inclusionSupport all users and workplace adjustments
Policy, strategy and transformationDrive adoption, value and continuous improvement
Real-world outcomes · turning responsible AI into a stronger public sector
Microsoft Copilot, Anthropic Claude, GitHub Copilot, Gemini and future tools sit as product overlays beneath one common control model. Role routes for every audience are in How to use this guide.
Your IT service desk, then the product’s support route.
The AI assurance team owns screening and the decision register.
Report through the incident route on the Operate page. Treat data exposure as a security incident.
Your line manager and workplace adjustments service.
Named contacts are set by each organisation when the guidance is deployed.
Your organisation’s AI policy and acceptable use.
The AI Playbook for the UK Government.
UK GDPR, DPIAs and international transfers.
Threat modelling, secure configuration and incident response.
Retention and disposal of prompts, outputs, transcripts and logs.
WCAG 2.2 AA for AI interfaces and outputs.
Supplier assurance, subprocessors, exit and service commitments.
Algorithmic Transparency Recording Standard for public-facing tools.
Twelve core artefacts and seven implementation artefacts.
Overlays for each supplier and deployment route.
Start here · orientation
87 chapters in 12 parts. Nobody needs all of it. Pick a reading mode, then a route.
12 parts in delivery order: decide, build, assure, operate, fund, the people and pilot material, then the deployment part and the reference. Roughly a day's read.
10 routes below, each an ordered set of chapters for one job. Twenty to forty minutes.
Use the launchpad or search: rate readiness, tier an agent, price a job, build an instruction block, check exit readiness.
Each route is an ordered set of chapters. Click any step to jump there.
Get data ready, keep metadata and indexing honest, govern connectors and treat AI outputs as records.
Find an approved tool, get a route for your task, and know where to get help.
Decide where AI is worth the risk and the spend, and be able to answer for it.
Adopt agents without destabilising a service people depend on.
Frame the problem, prove the value, keep the team moving.
Build well-grounded agents and prove they work.
Apply proportionate control and see the evidence behind it.
Fund adoption honestly and keep exit possible.
Build capability that outlasts the pilot.
Use the guide during a real deployment and record what happens.
What the guidance says, what this guide adds, and the ten capability domains it all rests on.
Whether to do it at all: evidence quality, use-case selection, criticality and the assurance tier that follows.
Capability class — generative, agentic or declarative — then instructions, architecture, grounding and the autonomy boundary.
The trust cycle, the operating model, risks and mitigations, registers, transparency records and recovery.
Life after approval: the portfolio view, the agent lifecycle, public-service impact and readiness tracking.
The full investment line, unit economics, supplier requirements and exit.
Teams, roles, pathways, inclusion and the organisational adoption programme.
How this guide is being tested in a real deployment, and what would justify a permanent toolkit.
Requirements, architecture, design, theory of change and the research base for the guide itself.
Glossary, thesaurus, generated knowledge graph, taxonomies, ontology, site map, guided tours and platform links.
The part that turns the guide into a deployment: fifty-four real services placed on volume and criticality, a generator that derives the whole governance set for any case, four agents carried end to end — one in each quadrant — and an explicit list of what this guide still does not cover.
Devolved and international comparators, and every source behind the content.
Verified Present in a surfaced source.
Interpretation A conclusion drawn across sources.
Proposal Recommended future state, not adopted policy.
Illustrative Looks like data, carries no measured source.
It does not override departmental policy, professional advice or formal approvals.
It is not the register of record for any live deployment.
It holds no personal, case or OFFICIAL-SENSITIVE data; assessments stay on your device.
It cannot issue an approval — the accountable owner does that.
The playbook
The AI Playbook for the UK Government (opens an external website) (GDS, 10 February 2025) supersedes the Generative AI Framework and covers all AI: machine learning, NLP, computer vision, generative and agentic systems. The ten principles are paraphrased below; open each for its first moves.
Playbook principle 5: manage the whole life cycle, including the unglamorous parts. Each stage has a gate; nothing moves right without passing it.
Stage 1
Assemble the multidisciplinary team; screen for DPIA and equality impacts
Gate · A named SRO and a written problem statement
Stage 2
Legal, ethics and commercial shape the approach, not the finished thing
Gate · DPIA drafted; risks owned; oversight designed in
Stage 3
Gate · Evaluation results reviewed by the whole team
Stage 4
Gate · Assurance sign-off; transparency record live
Stage 5
Gate · Monitoring dashboards live; exit plan on file
The domains
Cyber, data, digital, technology, and the legal-ethical frame around them. Each domain has its own authoritative guidance; a deployment is ready only when all five say so. That is why the team is multidisciplinary.
Cyber
NCSC Guidelines for secure AI system development cover secure design, development, deployment and operation. Agreed with CISA and 21 international agencies; aimed at any system that uses AI, built or bought.
Secure by Design and the Government Cyber Security Standard apply to AI services like any other, plus AI-specific threat modelling: prompt injection, data poisoning, model extraction, over-permissioned agents.
NCSC's principles for the security of machine learning translate the threats into controls. Walk before you run with agentic AI.
Data
The Data Ethics Framework guides appropriate, responsible data use across the public sector: transparency, accountability, fairness.
UK GDPR and the DPA 2018 require a DPIA before high-risk processing. The ICO's AI guidance, with a statutory AI code of practice in development, is the regulator's view of what good looks like. Cross-border processing (for example a US-processed model behind a UK product) needs a Transfer Risk Assessment as well as SCCs.
The Algorithmic Transparency Recording Standard is mandatory for central departments and arms-length bodies: a published record for every in-scope algorithmic tool.
Model quality is data quality: govern provenance, representativeness and retention for training and grounding data alike — retrieval systems inherit the permissions, quality and currency of the estate.
Digital
The Service Standard and Service Manual set the operating model: start with user needs, work in multidisciplinary teams, iterate in the open. AI services are still services.
The AI Playbook and the AI Knowledge Hub supply the patterns, case studies and communities; the PM's AI Exemplars programme is the test-and-learn engine for scaling what works.
Accessibility is law, not polish: WCAG 2.2 AA under the Public Sector Bodies Accessibility Regulations 2018. AI interfaces (chat, voice, summaries) need their own accessibility evidence.
Technology
The Technology Code of Practice governs the technology choices: cloud first, interoperability, open standards, sustainability, exit.
Compute and infrastructure are national policy: AI Growth Zones and a compute roadmap from 21 ExaFLOPs (2025) towards 420 by 2030 under the AI Opportunities Action Plan.
Buy well: use Crown Commercial Service AI routes, insist on evaluation access and model-change notice in contracts, and design the exit before signing the entry.
Legal & ethics
The Equality Act 2010 and the Public Sector Equality Duty apply to automated and AI-assisted decisions. Public-law duties (fairness, rationality, giving reasons) still bind when a model is in the loop.
Explainability has a manual: 'Explaining decisions made with AI', by the ICO and The Alan Turing Institute — and explanations shown to people should reference functional needs, never a diagnosis or label.
Public trust is earned in the open: engage affected people early, say clearly where AI is used, and give a route to challenge and redress.
Part 2 of 12 · 6 chapters
The evidence
What varies is method, not the tool. Uncontrolled self-report gives the highest figures; comparison-group econometrics gives mid-range; counterfactual-free evaluation of a short pilot gives the lowest. Role mix is the second driver.
26
minutes / day
20,000 staff, self-reported
60
minutes / week
3,000 licences
19
minutes / day
3,549 staff, econometric
0
robust evidence
1,000 licences
Never average these numbers. The spread is evidence about evaluation design. Sources: GDS findings report (Jun 2025); HMRC phase 3; DWP evaluation (Jan 2026); DBT evaluation (Aug 2025).
AI surfaces what a user may already reach. Permission drift becomes an incident on day one.
Makes scaling UNSAFE
The transparency standard records algorithmic tools. It is not an operational register of agents.
Makes scaling UNGOVERNABLE
Credits are billed per feature, not per conversation. A builder's design choice sets a budget holder's cost.
Makes scaling UNBUDGETABLE
Workflow redesign, local support and output verification happen but rarely appear as funded lines.
Makes scaling UNSUSTAINABLE
Between deployment and disclosure sits a gap no current UK instrument closes. The ATRS is a transparency mechanism, not an inventory: it does not capture team-shared agents, orphaned agents, data-source approval, consumption cost or retirement.
21%
of bodies responding to the NAO had an AI strategy
NAO · Mar 2024
74
use cases deployed, reported at the same date
NAO · Mar 2024
transparency records: Jan 2025 → end-2025
PAC · OECD.AI
28%
of central government IT is legacy
PAC · Mar 2025
A finding about where to deploy, not whether.
saved per day — GDS cross-government trial, 50 organisations
72% said good value; 58% would not return to working without it
slower — METR randomised trial: experienced developers, mature code
20% felt faster — the same developers, in the same trial
Benefit is largest on well-scoped, greenfield and boilerplate work; smallest or negative on mature, complex codebases. Developer self-assessment is unreliable in both directions. Join the AI Engineering Lab (OCTO, GDS) before procuring independently — the evidence you generate is reusable.
Contrary evidence, recorded deliberately.
of cross-government trial users noticed no clear time savings (GDS, Jun 2025)
of Australian participants reported the tool added time (DTA, Oct 2024)
of Australian managers could not confidently identify AI outputs (DTA, Oct 2024)
of councils were not ready on workforce skills and expertise (LGA, Jun 2025)
Canada names a further risk no UK instrument addresses: skill atrophy and automation drift. Skill maintenance is a prerequisite for meaningful human oversight, not an optional extra (TBS Canada, May 2026).
Twenty sources in one AI research draft. What checking found:
7
verified precisely
6
verified in substance
4
not verified as cited
1
could not be located
None of the unverifiable entries was obvious: each was correctly formatted, plausibly dated, attributed to a real institution. Treat AI-produced research as a hypothesis-generating artefact, never an evidence base — and keep the verification log with the output.
Verified fact — supported by official, regulatory, audit or primary sources
Organisation-reported — reported by the organisation itself, not independently measured
Supplier claim — authoritative for product terms; interested-party evidence for value
Interpretation — analytical judgement drawn from cited evidence
Recommendation — action proposed, not established practice
Unresolved — evidence incomplete, conflicting or arrangement-specific
Use cases & evidence
Eighteen categories, from productivity tools to enterprise agents — supporting people, processes and public outcomes. Use AI and agents responsibly to enhance the work of public servants and improve outcomes for citizens.
1. Knowledge retrieval & insight Find, summarise and synthesise information from trusted sources
2. Content & correspondence Draft, edit and refine high-quality content and communications
3. Contact centres & helpdesks Support staff and citizens with accurate, consistent responses
4. Casework & service operations Streamline case handling, triage and workflow automation
5. Citizen services & engagement Deliver simpler, faster and more accessible services
6. Meetings & collaboration Capture, summarise and turn meetings into actions and outcomes
7. Data analysis & reporting Analyse data, create summaries and produce management information
8. Software engineering & IT operations Accelerate development, testing and secure operations
9. HR & people management Improve people processes and the employee experience
10. Finance & procurement Enhance financial planning, analysis and supplier management
11. Education, training & skills Personalise learning and improve knowledge retention
12. Programme & project management Plan, track and report with intelligent support
13. Health & care Support clinical and administrative workflows safely and securely
14. Policing, justice & public safety Improve investigations, evidence handling and public protection
15. Regulatory & compliance Monitor compliance and produce regulatory outputs
16. Policy & strategy development Scenario analysis, option generation and evidence synthesis
17. Communications & media Create campaigns, content and prepare responses
18. Research & horizon scanning Scan trends, research and emerging opportunities
Secure access to trusted data & knowledge AI agents & orchestration (autonomy with oversight) Human oversight & accountability Privacy, security & data protection Quality assurance & model evaluation Observability, monitoring & auditability Integration with systems & workflows Change management & adoption FinOps & cost management
Data quality & governance Information management Access management Standards & interoperability Skills & culture Ethics & responsible AI
A consistent way to describe, evaluate and share AI use-case evidence: what the capability does, who it helps, how it works, the value it delivers, and the evidence that supports it. Evidence first. Value always. Capture consistently, share openly, decide confidently.
1. Basic information Use case name · Short description · Unique ID · Date created · Last updated · Author / owner
2. Problem & purpose Problem / opportunity · Purpose & objectives · Success criteria
3. Users & context Primary user groups · Who benefits · Operating context / environment · Geography / organisation
4. Solution & capability AI capability / agent type · Key features / functions · How it works (high level) · Human oversight points
5. Data & information Key data sources · Type of data · Data sensitivity · Information governance
6. Implementation & integration Platform / technology · Systems / tools integrated · Deployment model · Date went live / expected · Resource requirements
7. Outcomes & value Quantitative outcomes (measured) · Qualitative outcomes (observed) · Value delivered (how & for whom)
8. Costs & financials Cost model (licence, usage, etc.) · Estimated / actual cost · Cost per user / transaction · Ongoing cost considerations · Funding source / owner
9. Risks & assurance Key risks · Mitigations in place · Assurance / evaluation approach
10. Evidence & validation Evidence summary · Evidence sources · What evidence is available (performance results, user feedback, evaluation report, audit / review, other)
11. Limitations & lessons Known limitations / constraints · Lessons learned / recommendations
12. Transferability & reuse Potential for reuse (high / medium / low) · Reusable artefacts available · Transferability notes
13. Lifecycle & review Lifecycle stage (ideation, pilot, live, scaled, retired) · Review frequency · Next review date · Sunset / retirement plan
Live / in use Pilot / trial Planned Retired / stopped Unknown
Sharing the full picture — successes, failures and emerging ideas — so we learn faster and build smarter. Transparent evidence helps the public sector avoid repeating mistakes and scale what works. The cases below are illustrative.
Worked — repeatable success
Evidence shows value, quality and user benefit.
Helps advisers quickly find accurate, impartial information during calls.
Failed — learned & corrected
Initial attempt did not deliver — lessons captured and applied.
Automated extraction from clinical documents produced high initial error rate.
Unproven — promising, not yet validated
Early potential identified — evidence still limited.
Summarises meeting decisions, actions and sends updates to owners.
Extended pilot, independent evaluation, costing and governance review before scaling
Share openly, including failures Ground in authoritative sources Test with real users Measure impact and value Capture lessons and publish patterns Reuse and continuously improve
Each pattern is a pipeline with human oversight built in — not a single tool. Use them as the starting shape for a new use case.
Transforming conversations into decisions, actions and organisational memory
1. Meeting planning 2. Consent & capture 3. Transcript creation 4. AI processing & summarisation 5. Decisions & actions 6. Risks & insights 7. Distribution & integration 8. Organisational memory
Delivering trusted, accurate answers grounded in authoritative sources, with human expertise in the loop
1. Define need 2. Retrieve 3. Evaluate & ground 4. Generate response 5. Human review 6. Deliver answer 7. Feedback & learn 8. Maintain sources
Secure. Quality. Production-ready. Human in control.
1. Discover & plan 2. Architect & design 3. Build & code 4. Test & validate 5. Review & assure 6. Deploy & release 7. Operate & monitor
Smarter oversight. Better decisions. Greater impact.
1. Capture & ingest 2. Analyse & understand 3. Plan & optimise 4. Monitor & track 5. Decide & recommend 6. Collaborate & execute 7. Learn & improve
Decide and prioritise · public sector case library
Sourced cases from central government, councils, the Scottish, Welsh and Northern Ireland governments and health bodies. Each carries its source and an evidence grade. Supplier-published figures stay graded as reported.
Central government · Generative Corroborated
A package of AI tools for civil servants: consultation analysis, parliamentary search, meeting transcription, document summarisation and legal research.
Governance lesson: Consult was tested on past consultations before live use.
Central government · Generative Verified
Extract digitises planning documents. Early testing completed a one-to-two-hour manual task in under a minute.
Governance lesson: Check extraction against the source before it feeds planning decisions.
Devolved: Scotland · Generative Verified
The first live test of Consult used a Scottish Government consultation, followed by several other government consultations.
Governance lesson: Compare AI theming with human analysts before relying on it.
Local government · Generative Verified
Minute, a transcription and summarisation tool, was trialled in frontline local services in July 2025.
Governance lesson: Settle consent and retention for recordings before a trial starts.
Local government · Generative Verified
Piloted AI recording and summarising of assessments to free social workers for direct care.
Governance lesson: A DPIA examined data flows, storage and access. The privacy notice was updated and practitioners get consent to record.
Local government · Generative Reported
After 29 practitioners piloted the tool, more than 1,000 frontline staff use it. An evaluation reported seven hours a week saved on average.
Governance lesson: Figures are supplier-published. Validate locally before using them in a business case.
Local government · Generative Verified
A privacy notice for Magic Notes in adult social care sets out purposes, legal basis and rights.
Governance lesson: A published, service-level privacy notice is a reusable transparency artefact.
Local government · Generative Corroborated
Piloted AI to improve the quality and speed of recording assessment conversations.
Governance lesson: Evaluate quality of the record, not only time saved.
Local government · Declarative Verified
Set up a Data & Analytics Board to oversee data and AI governance, alongside a council-wide generative AI policy and staff training.
Governance lesson: Data and AI governance sit under one board.
Devolved: Scotland · Agentic Verified
Published 20 March 2026. An eight-layer “AI Stack” and a data matchmaking pilot for trusted public-sector datasets by 2027.
Governance lesson: Public-sector AI use is made visible through the Scottish AI Register.
Devolved: Scotland · Generative Verified
A suite of AI guidance and resources published on the Digital Scotland Service Manual in August 2026.
Governance lesson: Guidance lives in the service manual teams already use.
Devolved: Wales · Generative Verified
AI adoption across health, education and local government, with a commitment that AI works equally well in Welsh and English.
Governance lesson: Public bodies assess risks to vulnerable groups and consider offline users.
Devolved: Wales · Generative Reported
A baselining exercise found public bodies duplicating AI work. The office now coordinates adoption and says not every pilot should go live.
Governance lesson: Duplication is a governance failure; a shared register and reuse check fix it.
Devolved: Wales · Agentic Reported
A briefing on generative AI in Welsh public services covers inaccurate outputs, discrimination, privacy and the added security and accountability challenges of AI agents.
Governance lesson: Agent autonomy brings separate security and accountability questions.
Health · Generative Reported
A pilot assessing whether AI scribing reduces clinicians’ administrative workload.
Governance lesson: Pilot and evaluate before scaling clinical documentation tools.
Devolved: Northern Ireland · Agentic Verified
Consultation from 12 August to 7 October 2026. Eight principles include human oversight, accountability and redress, and data governance.
Governance lesson: Proposes human oversight teams, a cross-sector AI governance board and inventories of AI projects.
International comparator · Generative Corroborated
Advises against free generative AI tools because information given to them could be used for training.
Governance lesson: A clear position on consumer tools reduces shadow AI.
17 cases in the library. Gathered from public sources on 19 October 2026. Re-check each source before citing it.
Responsible AI · assess a use case
As AI can see more, do more and act more independently, the assurance required rises. Criticality & gates scores the service impact; this chapter sets the route from what the AI itself can do.
Class
Use organisational information
Retrieves and works with data from organisational sources and external systems.
Example: Search SharePoint or analyse internal data.
The capability classes say what the AI can do; the capability type says how it is built. Both set the assurance.
goes deeper on choosing between them.
Creates content in response to a prompt: text, summaries, images or code. A person decides what happens next.
Configured, not coded. Purpose, instructions, knowledge and permitted actions are declared; the platform supplies the model and orchestration.
Plans and takes actions across tools under delegated authority, choosing its own path to a goal, sometimes on a schedule.
| Class | Usual capability type | Minimum assurance |
|---|---|---|
| Ask | Generative | Review the output. |
| Create | Generative | Review the output; label AI involvement. |
| Connect | Generative, grounded · or Declarative with knowledge sources | Test what the index and connectors expose; review sources. |
| Act | Declarative with actions · or Agentic | Approval before action; log every action; least-privilege connectors. |
| Delegate | Agentic | Full agent route: identity, tool permissions, approval points, behavioural tests. |
| Persist | Agentic, scheduled or proactive | Specialist assurance: defined oversight, monitoring and a kill switch. |
Output quality, grounding sources, data entered, transparency labels.
The declared configuration: instructions, knowledge sources, actions, owner, version and publishing scope. Reviewed like code.
The whole human-agent system: identity, tools, approval points, escalation, logs, cost caps and the kill switch.
Assurance route
Start with the most the tool will do in this use case, not the most it could do.
0 of 60 answered · 0 concerns · 0 not known
Answers stay in this browser. “Concern” and “Not known” items carry into the decision record as conditions or open questions.
Proceed within the documented scope and controls.
Proceed once named conditions are met; conditions are tracked.
Controlled authority for a limited population, time and data set.
Narrow the users, data or actions permitted.
Extend a pilot to gather the missing evidence.
Do not proceed. Record the reason and lessons.
Criticality & assurance gates
Evaluate across ten dimensions to identify the agents that need stronger governance, assurance and operational support. Score each 1 (very low) to 5 (very high); the total sets the criticality tier and the action.
How many people use it and how often?
Who relies on it? Internal teams, partners, other organisations, or the public?
How important is it to achieving business outcomes or service delivery?
Does it influence, recommend or make decisions that have significant consequences?
What is the sensitivity, value and regulatory significance of the data it uses or produces?
How independently can the agent act? Human-in-the-loop versus autonomous.
How easy is it to reverse an error or unintended action?
How integrated is it with other systems, processes or agents?
What are the legal, regulatory, security and reputational risks?
Could failure cause harm, loss of trust or impact citizens?
CRITICALITY SCORE
— —
0 of 10 dimensions scored
12–20 LowMonitor, light governance, owner accountable
ModerateStandard governance, regular review and quality checks
HighEnhanced assurance, monitoring, human oversight required
Very highEnterprise governance, formal support, continuity and audit
LowPersonal research assistant used by creator only
ModerateTeam policy search agent used by a department
HighCase triage agent influencing case allocation and actions
Very highCitizen service agent that makes determinations or commitments
Assurance increases with audience, autonomy, information sensitivity, criticality and consequence. Pick a tier to see the gate required before use.
AudienceWho is affected? InformationWhat data? AutonomyWhat can it do? CriticalityHow essential? ConsequenceWhat is the impact?
Tier 2 · Departmental
Audience Department / multiple teams
Autonomy Acts with human approval
Information sensitivity Moderate
Criticality Medium
ASSURANCE GATE — REQUIRED BEFORE USE Formal assessment (risk & data review)
Oversight & accountabilityBusiness owner + data/info owner
Evidence & loggingAudit log · decisions recorded
Review frequencyMonthly
A holistic, cross-disciplinary approach ensuring AI systems and agents are safe, secure, reliable, fair — and deliver public value.
Security Privacy & data protection Data & information governance Ethics & fairness Transparency & explainability Human oversight & accountability Reliability & safety Evaluation & testing Auditability & evidence Risk management Legal & compliance Accessibility & inclusion
1. Plan & designDefine purpose, risks, requirements and success measures
2. Build & deployDevelop, test, assure and prepare for secure deployment
3. Operate & monitorMonitor performance, risks, usage and emerging issues
4. Review & improveLearn from evidence, update controls and improve continuously
Agents must operate within clearly defined accountability and oversight boundaries, with named owners and transparent decision trails. Clear accountability at every step; human judgement at the right points.
1. Source approvalInformation is approved for use by the agent.
2. Configuration reviewAgent design and guardrails are reviewed and authorised.
3. Output reviewOutputs are checked for quality, risk and compliance.
4. Decision authorisationFinal decision is made or authorised by a person.
5. Outcome monitoringOutcomes and performance are monitored and reviewed.
6. Independent assuranceIndependent review, audit and redress are available.
Clear accountability Better decisions Reduced risk Greater trust Stronger public value
Maturity & assurance · illustrative frameworks
Two illustrative frameworks from the capability report: a seven-stage maturity model for organisations, and six assurance tiers for agents. Neither scores named organisations; both are diagnostic tools to calibrate against evidence. Over-control halts adoption; under-control creates unmanaged exposure.
Observable evidence at stage 2
Bounded pilots; named sponsors; temporary funding; largely self-reported benefit evidence.
Named owners, repeatable gates, measured outcomes — not licensed tools.
Named accountable owners, clear outcomes, risk appetite and decision records.
Role-based training, protected learning time, inclusive adoption measures, staff consultation.
Register, architecture, approved platforms, integrations, autonomy level, observability.
Authoritative sources, metadata, permissions remediation, retention, provenance.
Threat model, DPIA where required, evaluation evidence, incident route, audit logs.
Licensing clarity, usage data, audit rights, exit, substitution, support terms.
Forecasts, showback, anomaly alerts, unit economics, value attribution.
Reusable artefacts, supersession, contribution workflow, evidence of reuse.
Bounded autonomy and recoverability adopted from Government of Canada agentic-AI guidance (May 2026).
Personal assistantDrafting, summarisation and retrieval for one accountable user
Acceptable-use policy, permissions hygiene, human verification, retention and transparency guidance.
Personal or team agentBounded knowledge and repeatable team tasks
Registered owner, approved sources, test prompts, sharing gate, usage telemetry.
Departmental reusable agentMultiple teams or a repeatable departmental process
Business, technical, information, risk and support owners; service level; change control; retirement plan.
Operational agentActs on records, workflows or line-of-business systems
Distinct identity, least privilege, action logging, rollback, incident response, continuous monitoring.
Citizen-impacting agentInfluences access, rights, safety, services or public communication
Impact assessment, legal basis, accessibility and equality assessment, transparency, redress, independent assurance.
Autonomous / multi-agentDelegated goals, tools, sub-agents or long-running actions
Sandboxing, constrained authority, short-lived credentials, kill switch, behavioural monitoring, dependency mapping.
What may this agent do without asking, and what can be undone when it goes wrong?
Part 3 of 12 · 8 chapters
Responsible AI · implement AI
Each stage answers one question, produces one artefact and passes one gate. Select a stage to filter the templates. The Agent journey chapter follows a single agent through the same path.
Explore Assess and design Deploy and enable Operate and evolve
Key gates
Stage 1
Stage 2
Stage 4
Stage 6
Stage 9
Stage 10
Stage 12
| Portal section | Key outputs | Primary roles |
|---|---|---|
| Start here | Opportunity brief · initial questions · support and contact | All colleagues |
| Find an AI tool | Catalogue search · tool comparison · reuse decision | Users · IT and digital |
| Assess a use case | Risk classification · data and privacy assessments · supplier and model review · approval record | Project team · risk, security, privacy · data and information |
| Implement AI | Design record · pilot plan · evaluation report · deployment checklist | Product owner · technical team · users and change leads |
| Operate and monitor | Usage, cost and outcome data · incident and risk records · change-impact assessments · continuous improvement | Service owner · operational support · governance and assurance |
| Learn and reuse | Lessons learned · updated guidance · patterns and templates · retirement record | All teams · community of practice · centre of excellence |
T1 · Core · stage 1
T2 · Core · stage 2
T3 · Core · stage 3
T4 · Core · stage 4
T5 · Core · stage 4
T6 · Core · stage 4
T7 · Core · stage 5
T8 · Core · stage 6
T9 · Core · stage 7
T10 · Core · stage 9
T11 · Core · stage 11
T12 · Core · stage 12
A1 · Added · stage 9, 10
A2 · Added · stage 5, 10
A3 · Added · stage 5, 9
A4 · Added · stage 6, 7
A5 · Added · stage 4, 7
A6 · Added · stage 5, 9
A7 · Added · stage 11
Design and build · capability classes
Three ways of using the same underlying models, with different failure modes and therefore different controls. Most real deployments combine all three, which is exactly why the class needs naming before the controls are chosen.
Produces content, analysis or options on request. Nothing changes until a person acts on the output.
You specify: A prompt, grounding sources and an output format.
Drafting, summarising, translating Analysis and option generation Search and question answering Code and test generation
Primary failure Plausible but wrong output accepted without review. Tier tendency Usually low to moderate: consequence sits with the person who acts on the output.
Controls that carry the weight
Takes actions across tools and systems under delegated authority, deciding its own path to a goal.
You specify: A goal, tools, permissions, guardrails and approval thresholds.
Triage and routing Multi-step workflow execution System updates within limits Orchestration across services
Primary failure Correct-looking action taken outside intended scope, or an irreversible action taken without approval. Tier tendency Moderate to high: the agent itself changes the world, so authority and reversibility dominate.
Controls that carry the weight
You state the outcome, rules and boundaries; the platform resolves how they are met and keeps behaviour inside the declaration.
You specify: Instructions, knowledge sources, permitted actions and constraints — as configuration, not code.
Configured assistants defined by a manifest Policy and rules expressed as configuration Low-code flows with declared triggers Templated agents published to a team
Primary failure The declaration drifts from intent, or nobody owns the specification that is now running in production. Tier tendency Varies with what is declared: the specification is the control surface, so review it as you would code.
Controls that carry the weight
Generative Agentic Declarative
What you supply Prompt and sources Goal, tools and authority Instructions, sources and constraints
Who chooses the path The person The agent, within guardrails The platform, within the declaration
What changes Nothing until a person acts Systems and records directly Whatever the declaration permits
Primary failure Wrong content believed Wrong action taken Wrong specification running
Assurance emphasis Output quality and provenance Permissions, approval and rollback Change control over the declaration
Register fields that matter most Sources, evaluation, review point Permitted actions, authority level, escalation Version, owner, constraints, publication scope
Three questions. The answer sets the controls that matter most — not whether the work is allowed.
The declaration sets purpose, sources, permitted actions and limits — and is the artefact under change control.
Inside those limits, the model drafts, summarises and reasons; evaluation and provenance apply here.
Only declared actions run, each with its permission, approval threshold and audit trail.
Record the class in the register alongside purpose and authority. It tells a reviewer which controls to look for before they open the evidence — and it makes the difference between an agent that drafts and an agent that acts visible without reading the configuration.
Skill studio · modular instruction block
A reusable way to brief any assistant: seven parts, a directives layer that always applies, validation that is always on, and a library loop so good prompts get kept, shared and versioned — not rediscovered.
Skill focus — pick any
System directives — always applied
Be accurate & truthful Be helpful & relevant Be safe & secure Be fair & inclusive Cite sources when possible Declare limits & assumptions
Assembled block
SYSTEM DIRECTIVES: be accurate and truthful; be helpful and relevant; be safe and secure; be fair and inclusive; cite sources when possible; declare limits and assumptions.
Validation — always on · 0 of 8 passed
Governance & safety rails
Confidentiality & privacy. respect data boundaries and permissions
Security & compliance. follow policies, standards and legal requirements
Trust & transparency. show reasoning, sources and confidence levels
Fairness & non-discrimination. avoid bias; promote equity and inclusion
Human oversight. AI supports; humans decide and take accountability
Monitoring & audit. logs, reviews, feedback and continuous improvement
Risk & harm prevention
Misinformation Privacy breach Overreliance Bias & exclusion Unsafe content Hallucination
DEIA considerations
Accessibility by design Inclusive language Cultural & regional awareness Multiple perspectives Bias detection & mitigation Cognitive load & readability
Learning & improvement loop
Apply the prompt in real work
Evaluate output and experience
Improve prompt and instructions
Add to library; share learnings
Retire, update or version prompts
The reusable library holds role-based, task-based and domain prompts, scenario templates, anti-pattern examples — tested and versioned, with supersession history.
Success signals
Useful Trusted Inclusive Accessible Traceable Actionable Impactful
Principles at the core: human-centred · inclusive by design · transparent · accountable · reliable · respectful · secure · sustainable · equitable.
Responsible AI · agents and skills
What starts an agent, how it plans, what it can reach, where people stay in control, and what it leaves behind. Skill studio builds the instruction block; Autonomy boundaries sets the limits.
What starts the agent?
How it thinks, plans and decides
What it can reach and do
Where people stay in control
What it delivers
Who the agent is and what it can reach
Classification and retention
Activity, actions and decisions
Budgets, limits and alerts
Risks, DPIA and incident response
Usable for all
Environmental impact
Stop and close safely
A shared skill or instruction changed without review changes every agent that loads it.
Proposed
Documents, web pages and emails can carry instructions that redirect the agent.
Proposed
Agents acting through shared or delegated contexts can reach more than the requester could.
Reported
Permissions alone may not stop titles, filenames and metadata surfacing through the index.
Reported
Agents calling agents need identity, approval points and automated records at each hop.
Reported
Agent activity logs may not show enough to reconstruct what happened.
Reported
1. Test before deployment
Behaviour, safety and access
2. Monitor in operation
Usage, outcomes and incidents
3. Learn and improve
Feedback and evaluation
4. Reassess on change
Model, tools, data and permissions
5. Retire when no longer needed
Revoke access, close data, capture lessons
A personal prompt stays personal. Once a prompt, skill or agent instruction is shared, it needs an owner, a version, a test set, security checks, a cost profile and a retirement date, and it goes in the prompt and skills register.
Deterministic alternative test
If a rule, form, flow or script can do the job reliably, use it. Generative AI earns its place where judgement over unstructured material is the actual task.
Design and build · connectors, MCP and integrations
MCP servers, connectors, plugins, add-ins and APIs turn a generative tool into one that can reach, and change, other systems. Each needs an owner, a scope and a log.
MCP serverAn open-protocol endpoint that exposes tools and resources to any MCP-capable client. Tools can be discovered at runtime, so capability can change without the agent being republished.
ConnectorBrings an external system’s content into an AI tool’s index or grounding, such as Copilot connectors.
Plugin or actionA packaged custom action, such as an MCP or API plugin in a declarative agent, that lets an agent read or change data.
Add-in or appAn extension that runs inside a host application, such as an Office add-in or a chat app.
API and webhookDirect integration through REST APIs, OpenAPI descriptions or event webhooks.
Integrator or gatewayA platform that brokers many connections, applies policy and logs traffic, such as an MCP gateway or integration platform.
Graded claims from supplier documentation. Where a route is unresolved, confirm it with the supplier before connecting.
Microsoft 365, Copilot and Copilot Studio Verified
Declarative agents call MCP servers and REST APIs through MCP and API plugins. Copilot Studio adds MCP servers through an onboarding wizard, and servers can be registered in Agent 365.
Controls: Tools are discovered at runtime unless pinned in the manifest. Register MCP servers in Agent 365 and review each tool’s read and write scope.
GitHub Verified
GitHub maintains a remote MCP server for repositories, issues and pull requests. Microsoft’s walkthrough wraps it as a plugin in a declarative agent.
Controls: One OAuth app per organisation, repositories scoped, no write tools without approval, and no secrets in prompts.
Corroborated
The official Rovo MCP server reached general availability on 4 February 2026. It is cloud-hosted and authenticates with OAuth 2.1 or API tokens.
Controls: Admins choose which AI tools and domains may connect, destructive tool groups stay off until enabled, and every tool call is written to the audit log. Atlassian Cloud only, not Data Center.
Anthropic Claude Reported
Claude connects to Microsoft 365 and other systems through connectors built on MCP. Its Microsoft 365 connector reads Outlook, Teams, SharePoint and OneDrive and writes to the last two.
Controls: Set connector write scope per user, keep an admin allow-list, and record each connector in the data-source register.
No verified documentation linked yet.
Google Workspace and Gemini Unresolved
Gemini works with Workspace content and connected data. Confirm the current MCP and extension routes with Google before use.
Controls: Confirm data location, admin controls and audit logging for each extension.
No verified documentation linked yet.
Slack Unresolved
Slack can be a source or a destination for agent work. Confirm the official MCP or app route and its data residency before connecting.
Controls: Scope channels, keep private channels out without approval, and align retention.
No verified documentation linked yet.
Service and line-of-business systems Proposed
ServiceNow, Salesforce, Dynamics 365 and case-management systems publish connectors, APIs or MCP servers. Each is a new data path and a new action surface.
Controls: Treat any write tool as Act class, require approval before change, and log to the incident trail.
No verified documentation linked yet.
Model Context Protocol, the standard Corroborated
An open standard that lets an AI client call tools and read resources on an outside system through one interface.
Controls: Prefer official servers, pin tools where possible, review tool descriptions for injected instructions and route calls through a gateway that logs them.
A generative chat tool with a read connector is Connect class. Add a write tool and it is Act class; let it run on a schedule and it is Persist. Reclassify whenever a connector or tool is added, and record it in the change register.
Architecture & quality
A three-tier canon governs everything that influences how something is built — prototypes explicitly included, because most principle breaches enter the estate through prototypes that were never intended to survive, and then did.
T1-01 Secure by Design T1-02 CAF & GovAssure T1-03 Technology Code of Practice T1-04 GDS Service Standard T1-05 AI Playbook T1-06 Data protection & public records T1-07 WCAG 2.2 AA
Tier 1 conformance is binary: the principle applies and is met, or a recorded exception exists. No informal partial conformance.
Test: Can each element be traced to a named authoritative owner outside the central platform?Looks compliant, is not: "We will federate later, once everything is in one place."
Test: Can a user discover an asset exists without thereby gaining access to it?Looks compliant, is not: Endpoint or credential details published as discovery metadata for convenience.
Test: Can any automated component write to the record without recorded human approval?Looks compliant, is not: "The user can always review it" — optional review presented as a control.
Test: Take any output at random: can you reconstruct inputs, owner, decision and evidence without asking a person?Looks compliant, is not: Evidence linked at closure rather than at capture.
Test: Is there a recorded reuse assessment that predates the build decision?Looks compliant, is not: Reuse assessment performed after the build decision, to justify it.
Test: Is accessibility evidence produced before release? Does every status carry a non-colour cue?Looks compliant, is not: "We will run an accessibility audit before go-live" as the only accessibility activity.
Test: Attempt to breach the control through the normal interface. Does it succeed?Looks compliant, is not: A mandatory gate with no exception path — guaranteeing informal workarounds.
Prototype promotion gate. A prototype may not be promoted toward live use until Tier 1 and Tier 2 conformance is assessed and recorded. Record intended lifespan and disposal point at creation; record deferred principles; re-assess them at the gate. Human-in-the-loop applies in full even at prototype stage — autonomy introduced early is rarely removed later. Promotion is a decision, not a drift.
From the AI Engineering Lab's reference implementation: the transferable innovation is not any one use case — it is the operating model around it. Specialist agents routed by an orchestrator; a challenger agent that reviews outputs; a compliance agent; assurance gates executed in the workflow; immutable audit trails; human approval before anything writes to the record. The same pattern serves capability management, catalogue operations, workforce planning and architecture governance.
Specialist agents Orchestrator routing Challenger agent Compliance agent Assurance gates in-workflow Immutable audit trail Human approval to write Escalation paths
A gate, not a form: run at the end of every iteration. Tier 0 is the brake — the only tier that can argue for less; a veto is the instrument doing its job. Scores are capped by evidence strength, unscoreable items go on an open register published with the score, and demo drift — a level-3 prototype someone useful quietly relies on — is the most common governance failure in this class of work.
Quick index — ten dimensions, 1–5
Score the dimensions for a reading
The ladder: Tier 0 proportionality gate → Tiers 1–6 architecture compass → Tier 7 self-* autonomic → Tier 8 human & institutional sustainability. Evidence-rung caps; accessibility floors at human testing, not automated checks; an empty removal ledger three iterations running is itself a Tier 0 veto. Dimensions here are representative — attach the full template for real assessments.
Knowledge foundation
Transforming conversations into trusted, discoverable knowledge that drives action and public value. Capture once, reuse everywhere — the seven-step pipeline and the record it produces.
Discussions and insights Questions and challenges Decisions and agreements Actions and owners Shared links and resources Demos and use cases Risks and issues Ideas and opportunities Lessons learned
Easier to findSearchable, organised knowledge at your fingertips
Avoid duplicationLearn from others' experience and mistakes
Stronger decisionsBetter informed choices backed by evidence
Faster deliveryReuse patterns, guidance and assets to go faster
Assured and trustedQuality, governance and provenance you can rely on
Greater public valueBetter services, outcomes and trust for citizens
Twelve components that turn a conversation into a reusable organisational asset. Session record · community owned · governed and secure · discoverable and reusable · continuously improved.
| 1. Session details Title & description · Date, time, duration · Community / forum · Participants & roles · Host / facilitators | 2. Recording & transcript Recording link · Transcript / captions · Time-stamped version · Accessibility info | 3. Summary & key points Executive summary · Key discussion points · Important context · What was agreed | 4. Questions & answers Questions raised · Answers provided · Outstanding questions · Follow-up needed |
| 5. Demonstrations & artefacts Demos & walkthroughs · Screens / diagrams · Code / configs · Resources shared | 6. Decisions & agreements Decisions taken · Agreements reached · Scope confirmations · Changes approved | 7. Actions & owners Action items · Owners · Due dates · Status / progress | 8. Linked resources & references Related publications · Previous sessions · External references |
| 9. Classification & sensitivity Information classification · Sensitivity level · Handling instructions · Access permissions | 10. Evidence & insights Evidence referenced · Impact / value assessed · Metrics mentioned · Outcomes expected | 11. Lessons & insights Key takeaways · Update history · Supersedes / superseded · Recommendations | 12. Review & lifecycle Review cycle · Update history · Supersedes / superseded · Retention |
VerifiedIndependently confirmed
Reported / emergingReported by a source, not yet validated
In progress / unvalidatedBeing worked through
Contradictory / needs reviewSources disagree
Outdated / supersededReplaced by newer content
Good knowledge is captured once, curated well, reused often and improved continuously — and its status is always visible.
Create & capture 1. Capture Capture knowledge, evidence and experiences from sessions, projects or operations.
Create & capture 2. Classify & tag Add metadata, taxonomy, topics, sensitivity and links to related assets.
Curate & govern 3. Curate & quality Check accuracy, completeness, relevance and quality. Remove duplicates.
Curate & govern 4. Validate & approve Subject matter and governance review. Approve for use and assign lifecycle status.
Share & use 5. Publish & discover Publish to trusted repositories and make discoverable through search, portals and APIs.
Share & use 6. Use & reuse Apply knowledge in work, decisions and services. Share feedback and outcomes.
Share & use 7. Review & update Review on a regular cycle or when triggered by change. Update content, links, owners and status.
Share & use 8. Sunset, supersede or archive Retire outdated content, mark as superseded or archive with retention and audit trail.
Draft Newly captured. Not yet reviewed.
Reviewed Peer or owner reviewed. Accuracy and relevance checked.
Validated Approved by subject matter and governance for use.
Published Officially published and visible to intended audience.
Superseded Outdated or replaced by newer content. Retained for reference.
Archived No longer active. Retained in accordance with policy.
Information is scattered across teams, tools and tenancies, which causes duplication, inconsistency and poor discoverability. Federate rather than migrate: connect the sources, add a metadata and governance layer, and put one discovery layer on top.
Chats & calls SharePoint sites Personal OneDrives Email & attachments Departmental file shares GitHub repositories Slack / other channels Knowledge hub Spreadsheets & lists Recordings & transcripts Documents & guidance External links & references
Unified discovery layerSingle search & navigation · Smart filters & facets · Recommendations & related content · Personalised views
Metadata & governance layerCommon taxonomy · Metadata standards · Classification & sensitivity · Ownership & accountability · Lifecycle management
Federated source connectorsSource systems remain in place — connect, do not migrate
Find fasterOne place to search across the estate with better results
Trust & assuranceClear ownership, quality signals, classification and provenance
Reuse & collaborateEasily share, reference and reuse patterns, guidance and code
Learn & improveInsights and lessons flow back to improve practice continuously
Deliver more valueBetter decisions, less duplication, faster delivery and stronger outcomes
High-quality information in, trusted retrieval and reasoning, accurate answers or actions, human review, better sources. Every link has a quality gate.
Answer accuracy rate Citation coverage User satisfaction score Time to answer Re-use rate of content Feedback captured Source improvements
Work down the questions in order. The first "yes" is your answer — the goal is the right information, to the right people, at the right time, securely and efficiently.
Yes → Use existing authoritative location — keep content in place, ensure agent uses existing permissions, maintain single source of truth, lower duplication and maintenance risk.
Yes → Grant access via existing permissions — respect least privilege, use groups and roles already in place, review access periodically.
Yes → Adjust permissions or create a secure location — create dedicated site or library if needed, configure granular permissions, review with the information owner.
Yes → Copy or move content with strong controls — risk of duplication and divergence, increased maintenance and update process, maintain links back to source of truth, confirm retention and records handling.
Information governance — Is the content authoritative, current and controlled?
Duplication risk — Will copying create multiple versions or conflicting answers?
Maintenance — Can the team keep copies current and in sync with the source?
Licensing & cost — Do users have the right licences? Any connector or storage costs?
Records & retention — Where is the record of business? How long must it be kept?
Security & risk — Does the agent minimise risk of overexposure or data leakage?
Connecting people, communities, data, decisions and capabilities to power trusted AI — and a common language layer that improves responses, saves time and builds organisational memory.
Connects what sits apartLinks people, data, agents and outcomes
Improves discovery and reuseFind what already exists before building
Provides context for AIGrounded, related, permissioned context
Supports governance, assurance and valueOwnership, lineage and measurable benefit
1. Organisational prompt library Curated collection of high-quality prompts · Categorised by task, role and domain · Example prompts and use guidance
2. Prompt lab & testing Experiment and refine prompts safely · Compare results and measure quality · Promote best prompts to the library
3. Business glossary & acronym dictionary Common definitions and terminology · Acronyms, abbreviations and plain-government terms · Versioned and community-maintained
4. Domain & policy knowledge packs Topical guidance, policies, frameworks · Approved content summaries for AI use · Content packs for specific domains
5. Technology-agnostic prompt patterns Reusable patterns that work across tools and models · Task-based templates and structures · Adapts to different AI platforms and models
1. Create & contribute Users contribute prompts, definitions and knowledge
2. Curate & review Experts review for quality, accuracy and safety
3. Publish & tag Approved items are tagged and published to the layer
4. Use & apply AI tools and users discover and apply in their work
5. Feedback & improve Results and feedback feed back to improve content
Responsible autonomy boundaries
Autonomy is a privilege, not a default. Human oversight increases with risk, impact and consequence — and the human role is named at every level, not implied.
Human controlPeople remain accountable
Proportionate by riskMatch autonomy to impact
Transparent & explainableMake actions understandable
Safe by design & defaultGuardrails built in
Continuously monitoredDetect, learn and improve
LOW AUTONOMY HIGH AUTONOMY
Level 2 of 6
Agent recommends options or actions for human consideration.
User reviews and decides
GUARDRAILS (ALWAYS ON) ESCALATION & OVERRIDE
Part 4 of 12 · 8 chapters
Responsible AI · govern and assure
Consistent principles, proportionate assurance and reusable components. Evidence and transparency records sit with the controls because they are how each control is shown to work.
Capability-based classification sets the assurance depth.
Readiness, metadata, indexing, provenance and retention.
Threat model, DPIA, transfers, subprocessors and DLP.
Reviewer competence, evidence access, override and stop.
Accessibility of AI, AI for accessibility, alternatives and adjustments.
User-visible cost, caps, alerts, budget owners and value criteria.
Supplier evidence, usage indicators and proportionate use.
AI involvement statements, labelling and audit evidence.
Control effectiveness, incidents and reviews.
Change watchlist, reassessment triggers and retirement.
| Role | Accountable for |
|---|---|
| Senior responsible owner | Risk acceptance, sponsorship and the decision to proceed or stop. |
| Service owner | Day-to-day operation, monitoring, support and retirement. |
| Technical owner | Configuration, identity, permissions, release and change. |
| Commercial owner | Contract, licences, consumption, supplier commitments and exit. |
| Assurance owner | Assessments, evidence, exceptions and reviews. |
| Budget owner | Caps, alerts and spend against value evidence. |
| Agent, prompt or skill owner | Version, tests, approved users and withdrawal. |
Unmanaged external use falls when asking for an approved alternative is quicker than working around the rules.
Keep clear labels on AI-sent messages and disclose material AI involvement. Removing a label is prohibited where transparency is required.
What AI did in producing a document, decision or message.
The sources an output drew on, so a reader can check them.
Who checked the output, against what, and when.
How AI informed a decision and who made it.
Provenance and review of AI-generated code.
What an agent did, when and why.
Configuration, approvals and test results for assurance.
These sit alongside the guide's own conventions (verified, interpretation, proposal, illustrative). They grade claims about products and suppliers, many of which come from community calls.
Confirmed against authoritative supplier documentation or a live configuration.
Reported by more than one independent source or session.
Raised in a community call or channel. Not yet confirmed.
A design recommendation in this portal, not a finding.
An open question. Do not rely on it until answered.
Assure and govern · data and information governance
Whatever an AI tool can reach, it can surface, summarise and act on. Data and information governance sets what it should reach, and treats what it produces as records.
Ownership and accountabilityEvery source has an information asset owner; the senior information risk owner accepts residual risk.
Classification and handlingClassification travels with content into prompts, indexes and outputs.
Data protectionUK GDPR lawful basis, DPIA, transfers, rights and service-level privacy notices.
Quality and authorityGrounding sources are authoritative, current and fit for the decision they inform.
Metadata and discoverabilityTitles, filenames and metadata can surface through AI. Discovery is not access.
Records and retentionPrompts, outputs, transcripts and agent logs are records with retention rules.
Access and permissionsLeast privilege, oversharing clean-up, delegated and shared contexts tested.
Sharing and reuseData-sharing agreements, data boundaries and trusted reuse such as data matchmaking.
| Stage | Data and information control | Where it is recorded |
|---|---|---|
| 1 Discover | Identify the information assets involved and their owners | T1 Opportunity Brief |
| 2 Check | Reuse existing datasets, catalogues and indexes before creating new ones | T2 Duplication Check |
| 3 Classify | Classification, personal data and special category data | T3 Classification |
| 4 Assess | DPIA, lawful basis, transfers and subprocessors, data quality and authority | T4 Data Readiness · T5 Supplier |
| 5 Design | Metadata, index scope, permissions, connectors and retention rules | T7 Agent Design Record · data-source register |
| 6 Pilot | Minimisation, test data, consent and privacy notices | T8 Pilot Plan |
| 10 Operate | Access reviews, DLP, records of AI outputs and agent logs | Monitoring and assurance record |
| 11 Reassess | New sources, connectors or locations trigger review | T11 Change-Impact Assessment |
| 12 Retire | Disposal, records transfer and revoked access | T12 Retirement Record |
Retain where they shaped a decision or a published output; organisational prompts are versioned products.
Treat as drafts until a person verifies them; keep the verified version with an AI involvement statement.
Consent notice, retention period and deletion route set before recording starts.
Keep for audit and incident review; owner can read them.
Derived copies of source data; inherit the source classification and retention.
Test sets and results are evidence for the decision record.
Data Ethics Framework ↗ (opens an external website)UK government framework for using data responsibly in policy and services.
Government Data Quality Framework ↗ (opens an external website)Principles and practices for assessing and improving data quality.
ICO guidance on AI and data protection ↗ (opens an external website)How UK GDPR applies to AI, including fairness, transparency and DPIAs.
Northern Ireland draft AI Strategy ↗ (opens an external website)Data governance is one of the eight proposed principles.
Scotland’s AI Strategy 2026–2031 ↗ (opens an external website)Data framed as a public asset, with a data matchmaking pilot.
Kingston Council ran a DPIA on data flows, storage and access before going live with AI note-taking, and updated its privacy notice. Hartlepool publishes a service-level privacy notice for the same kind of tool. Belfast City Council put data and AI governance under one Data & Analytics Board.
Architecture of trust
A continuous cycle that governs AI agents as digital workers to deliver trusted public services and outcomes. The agent at the centre is purpose driven, accountable, supervised, transparent, secure and ethical. Trust at every step: right agent, right authority, right outcome.
Human accountability Safety, security and privacy Lawful, ethical and fair Transparency and explainability Inclusive and accessible Proportional and risk-based Evidence based and measurable Sustainable and value for money
Service Standard Technology Code of Practice AI Playbook & Responsible AI Data & AI Ethics Framework Security & Privacy Standards ATRS & Transparency
Better servicesSimpler, faster and more responsive
Trusted decisionsFair, consistent and well governed
Efficient operationsAutomate safely, reduce burden
Empowered workforceAugmented capability, not replaced
Improved outcomesBetter impact for citizens, business and communities
Stronger public trustAccountable, transparent and responsible AI use
Security by design Protect systems, data and users
Privacy by design Respect and protect personal data
Human in control Meaningful oversight at all times
Impact aware Consider equality, ethics and public impact
Transparent Open about how and why agents are used
Measured & monitored Track performance, risk and value
Continuously improved Learn, adapt and evolve responsibly
Enterprise AI governance operating model
A layered model for governing AI agents as digital workers to deliver trusted public services. Each layer answers one question — from why we do it, to the impact we create together.
Set direction, objectives and principles aligned to mission, public value and national priorities.
Establish policies, roles, responsibilities, decision rights and governance structures.
Apply mandatory standards and guidance (Service Standard, TCoP, AI Playbook, Responsible AI, data and security standards, ATRS).
Identify, assess, mitigate and monitor risks across the AI lifecycle. Assure compliance, safety, ethics and performance.
Provide secure, scalable, reliable platforms, data foundations, tools, shared services and reusable components.
Design and deliver AI-enabled services and solutions that solve real problems and meet user needs.
Managed AI agents with defined purpose, capabilities, authority, guardrails, tools, data access and human oversight.
Civil servants, partners and operational teams who work with AI agents, provide feedback and maintain oversight.
Better services, fairer decisions, improved efficiency and outcomes for citizens, businesses and communities.
People & capabilityRoles, skills, training, culture and change enablement.
Security & privacySecure by design, privacy by design, data protection and resilience.
Ethics & inclusionFairness, equality, accessibility, human rights and inclusion.
Transparency & ATRSTransparency records, explainability, traceability and open publication where appropriate.
Data & knowledgeQuality data, governance, lineage, provenance and knowledge management.
Technology & architectureInteroperability, open standards, cloud first, modular and secure architectures.
Continuous improvementMeasure, learn, iterate and improve services, models and operations.
Key risks & mitigations matrix
Proactive risk management so AI agents operate safely, lawfully and effectively as digital workers. Human accountability remains non-delegable. Residual ratings assume the controls listed are in place and effective.
Residual: Low
Residual: Low
Sensitive or personal data is exposed, misused or shared inappropriately.
Breach of privacy law, fines, individual harm, loss of trust.
Residual: Medium
Users over-trust agent outputs and fail to challenge or validate them.
Poor decisions, unintended harm, reduced critical thinking.
Residual: Medium
Residual: Medium
Agent outputs unfair, discriminatory or disproportionate results.
Inequitable outcomes, legal challenge, reputational harm, loss of legitimacy.
Residual: Medium
Agent exploited, manipulated or used maliciously by attackers.
System compromise, data theft, service disruption, fraudulent actions.
Residual: Low
Agent performance declines over time as data, context or patterns change.
Reduced accuracy, poor outcomes, inefficiency, user dissatisfaction.
Residual: Low
Unclear ownership, missing controls or weak governance arrangements.
Unclear responsibility, compliance breach, regulatory scrutiny, loss of trust.
Human accountability People remain answerable.
Lawful & ethical use Compliant, fair and responsible.
Transparency & explainability Understandable actions and outputs.
Security & privacy by design Protect data and systems.
Inclusive & equitable Designed for all users and groups.
Measure & improve Evidence based continuous improvement.
Monitor & adapt Stay responsive to change.
Document & record Keep clear, complete and auditable records.
A single source of truth for AI agents: discoverable, governed and actionable. The four-group record above is the minimum; this is the full blueprint, with an illustrative value for each field.
WhyVisibility, control, safe adoption and reuse WhoOwners, users and support WhatPurpose, function and behaviour WhereData sources, systems and environment HowControls, assurance and monitoring Cost & valueUsage, cost and outcomes
1. Identity Agent ID (unique) · Name · Description · Version · Status e.g. AG-2026-015
2. Purpose & function Purpose · Primary functions · Typical tasks · Intended outcomes e.g. Summarise policy documents and answer internal questions
3. Ownership Business owner · Creator / builder · Technical owner · Information owner · Risk owner · Service owner e.g. Policy Team Lead
4. Users & audience Primary users · User groups · Audience size · Geographies / organisations e.g. 250 staff in 2 departments
5. Platform & model Platform (e.g. M365, Copilot Studio) · Model(s) used · Model version · Region / residency e.g. GPT-4o via Microsoft 365 Copilot
6. Data & information sources Source systems · Information assets · Grounding method (RAG, connectors) · Data sensitivity e.g. SharePoint libraries, Policy DB, HR system
7. Access & permissions Access model · Authentication · Permissions set · Least privilege confirmed e.g. Entra ID group based access
8. Sharing & distribution Sharing scope (personal / team / department / cross-gov) · Channels / locations · Public-facing? Y/N e.g. Teams tab, shared with 3 teams
9. Risk & criticality Risk category · Criticality (low/med/high) · Impact areas · Consequences e.g. Medium — operational impact
10. Controls & assurance Guardrails / controls · Human oversight points · Testing completed · Assurance status e.g. Human review before action
11. Usage & performance Usage level · Key metrics · Quality indicators · Last reviewed e.g. 1,250 queries / month, 92% helpful
12. Cost & consumption Licensing / pricing model · Tokens / credits used · Estimated monthly cost · Cost owner e.g. Pay-as-you-go, £320 / month, IT
13. Integrations & dependencies Connected systems · APIs / connectors · Dependencies · Third parties e.g. Graph API, Power Automate
14. Monitoring & observability Logging enabled · Monitoring tools · Alerts & thresholds · Incident contact e.g. Purview logs, alerts via Sentinel
15. Change & release Change process · Release history · Last updated · Next review date e.g. v1.2 — 12 Aug 2026, review 12 Nov 2026
16. Support & operations Support model · Support owner · SLA / hours · Escalation path e.g. IT Service Desk P1–P3, 24×7
17. Benefits & value Expected benefits · Measured outcomes · Baseline / target · Value owner e.g. Saves 10 hrs/week, faster responses
18. Retirement & archiving Retirement criteria · Data retention · Archive location · Decommission plan e.g. Retire if inactive 180 days
Active In review Planned Retired Draft
By owner By department By status By risk By platform By data sensitivity
Policy & standardsAligned to organisational AI policy, standards and acceptable use.
Approval & publishingClear gates for publication and sharing at each tier of use.
Ongoing reviewRegular reviews of risk, performance, cost and relevance.
Discovery & reuseEasily discoverable to avoid duplication and share good practice.
Audit & accountabilityTraceable ownership, decisions, changes and evidence.
Integrated, not duplicated. Each serves a distinct purpose and links to the others for discovery, control and reuse — over one set of shared foundational services. Single source of truth · integrated, not duplicated · accessible and discoverable · governed and accountable.
Maintain authoritative inventory of information assets for governance, compliance and risk management.
Policies, standards & procedures Documents & records Data sets & databases Forms & templates Knowledge bases Reference data Expert directories
Owner & accountability · Classification & sensitivity · Retention & disposal · Legal & regulatory · Quality & governance
Make data easy to find and understand to promote reuse, reduce duplication and improve decision-making.
Data sets & data products APIs & data services Metrics & KPIs Data dictionaries & models Knowledge bases Lineage & relationships Usage & popularity
Data definition · Owner & steward · Lineage & relationships · Quality & freshness · Access & usage
Register and govern AI agents and agentic capabilities to ensure safe, accountable and effective use.
Agents & agent teams Purpose & use cases Intended users & audience Data sources & tools Risk & criticality Controls & assurance Lifecycle & status
Agent owner & creator · Purpose & scope · Data sources & tools · Risk & criticality · Lifecycle & review
Identity & access management Security & privacy controls Metadata & standards (terms, definitions) Integration & APIs (data & systems) Audit & monitoring (usage, changes) Quality management (validation, rules) Search & discovery (federated)
1. Create / ingest — New information, data or agents are identified and registered
2. Link & relate — Relationships are captured between assets, data and agents
3. Use & leverage — People and agents find and use trusted sources and capabilities
4. Monitor & assure — Activity, quality, risk and compliance are monitored across the landscape
5. Improve & evolve — Feedback drives better sources, smarter agents and stronger controls
Real-time visibility of agents, usage, performance, risk and value — insight into every agent, control where it matters, confidence in outcomes. These are the panels the register needs to earn its keep.
Agents by status Active · In review · Retired
Usage over time Total agents against active agents, last 30 days
Top agents by interactions Ranked list with trend sparklines
Risk distribution Very high · High · Medium · Low
Agents by owner type Individual · Team · Department · Enterprise
Spend by agent Cost per agent, highest first
Quality & reliability Success rate · Response time · Feedback score · Error rate
Sharing & discovery Shared · Public to organisation · Cross-organisation
Alerts & needs attention Agents with no active owner · Breached sharing thresholds · Not reviewed in 60+ days · Rising error rate
Recommended actions Assign ownership and confirm purpose · Review sharing scope · Confirm continued value and risks · Validate cost and usage patterns
Owner Department Agent type Platform / tool Data sensitivity Risk level Lifecycle status Tags / capability Geography
Internal ATRS registry · AI agent transparency record
Based on the principles of the Algorithmic Transparency Recording Standard. Purpose: to ensure transparency, accountability and responsible use of AI agents by recording key information about their purpose, authority, data, risks, controls and lifecycle. The registry integrates with enterprise systems for identity, security, monitoring, risk, audit and knowledge management.
Lifecycle stage (proposed, pilot, live, restricted, suspended, retired)
Central inventory of AI agents Transparency for oversight Evidence for audits & assurance Risk management & reporting Informed decision making Public sector standards compliance Learning & improvement
Record quality: keep the record accurate, up to date and complete. Review at defined intervals or when significant changes occur.
Failure & recovery loop
Turning errors into stronger systems and better outcomes. Nine stages, and the ninth feeds the first: every failure is a source of learning. Work a live issue through the loop and record where you are.
TransparencySurface issues quickly AccountabilityOwn and act on problems LearningCapture and share lessons Continuous improvementStrengthen continuously ProportionateRespond to risk and impact
Stage 1 of 9
Identify that something has gone wrong.
Errors, unexpected outputs User feedback Monitoring & alerts Quality checks
Inaccurate or incorrect outputs
Outdated or incorrect data sources
Prompt or model misbehaviour
User confusion or misuse
System or integration failure
Policy or compliance breach
Part 5 of 12 · 6 chapters
AI governance control tower · illustrative dashboard
Govern. Assure. Monitor. Improve. Every figure below is illustrative, showing the dashboard a governed portfolio produces — last updated 20 May 2025 09:30.
128
↑ 12 this month
87
↑ 7 this month
45
↓ 3 this month
5
— No change
12
↑ 2 this month
116
↑ 10 this month
Proposed 8 6%
Live 87 68%
Pilot / testing 45 35%
Suspended 5 4%
Retired 12 9%
Assured 72 56%
Partially assured 28 22%
Not assured 18 14%
Not required 10 8%
High risk 12 9%
Medium risk 34 27%
Low risk 62 48%
Very low risk 20 16%
12 agents are high risk and require action.View risk register →
Citizen services 32 25%
Corporate services 24 19%
Policy & strategy 18 14%
Operations 16 13%
Finance 14 11%
HR & people 12 9%
Other 12 9%
AGENT NAME SERVICE / DOMAIN PURPOSE OWNER LIFECYCLE RISK ASSURANCE ATRS RECORD LAST REVIEWED NEXT REVIEW
Policy & Strategy
Supports policy analysts by summarising key documents and drafting briefing notes.
Sarah Johnson (Director)
15 May 2025
15 Aug 2025
Citizen Services
Prioritises and routes incoming casework to the right teams.
Mark Patel (Head of Ops)
10 May 2025
10 Aug 2025
Finance Anomaly Detector Finance Detects unusual transactions and supports investigation. Alex Kim (Finance Lead) Live Medium Assured ATRS-00028 08 May 2025 08 Aug 2025
HR Policy Advisor HR & People Provides guidance on HR policies and employee queries. Lucy Bennett (HR Manager) Pilot Low Assured ATRS-00061 12 May 2025 12 Aug 2025
Corporate Services
Assists IT staff with troubleshooting and knowledge retrieval.
Tom Edwards (IT Manager)
18 May 2025
18 Aug 2025
Comms Drafting Agent Corporate Services Drafts communications and public messages. Nina Shah (Comms Lead) Suspended High Partially assured ATRS-00017 02 May 2025 Review pending
46Data sources connected
27High sensitivity data sources
96%Agents have named owners
100%DPIA completed (in scope)
94%Security reviews up to date
85%Agents have defined success metrics
65%Agents have sustainability assessment
Responsible AI · operate and improve
Monitoring, incidents, the change watchlist, reassessment and retirement, run as one learning loop. The Portfolio control tower shows the estate view; this chapter defines what feeds it.
| Domain | What it shows | Fed by | Act when |
|---|---|---|---|
| Adoption and usage | Who uses which capability, and how often, by group. | Tool, use-case registers | Uptake is uneven across groups or roles. |
| Cost and consumption | Spend by tool, team and use case against caps. | Tool, benefits registers | A cap or alert is hit, or cost grows faster than use. |
| Outcome quality and value | Verified outcomes, not prompt counts or time saved alone. | Benefits register | Value evidence is missing at review. |
| Correction and rework | Review effort and how much output is edited or discarded. | Benefits register | Rework cancels the time saved. |
| Accessibility and inclusion | Barriers reported, adjustments supported, alternatives used. | Incident, use-case registers | A barrier is reported without an alternative. |
| Risk, incidents and near misses | Incidents by tool, model, use case and control. | Incident register | Repeat incidents on one control. |
| Sustainability indicators | Supplier environmental evidence and usage indicators. | Tool, supplier registers | Evidence is missing or usage is disproportionate. |
| Control effectiveness | Whether controls worked when tested or triggered. | Incident, review registers | A control failed or was bypassed. |
| Change and deprecation exposure | Live use affected by upcoming supplier changes. | Change, model registers | A deprecation lands inside the review window. |
| Support demand and recurring issues | Tickets and repeat questions by tool. | Support and service records | The same issue recurs across teams. |
No figures are shown. Each domain is defined so live data can be connected without redesign.
A new, changed or automatically selected model.
Partial or full
New subprocessor, location, terms or ownership.
Full
New sources, connectors or classifications.
Partial or full
Wider access, new identities or delegation.
Full
New tools, actions, schedules or agent links.
Full
Changed terminology, layout or defaults.
Guidance update
New pricing, metering or consumption pattern.
Partial
New law, policy or guidance.
Partial or full
Microsoft IgniteSeveral preview items in the service atlas are likely to change. Refresh entries after the event.
Proposed
Staggered Copilot interface rolloutUsers see different interfaces at once. Keep guidance by release channel and show preview versus general availability.
Reported
Default-enabled featuresNew features can arrive switched on. Check admin defaults after each release.
Reported
Claude models reached through MicrosoftConfirm subprocessor terms and whether processing sits inside the data boundary.
Unresolved
Microsoft UK data processing commitmentConfirm in your admin centre before relying on it.
Unresolved
Indexed metadata exposureTitles, filenames and metadata can surface despite permissions. Re-test indexing on change.
Reported
Cowork cost and securityConcerns raised about credit consumption and security. Set caps before any pilot.
Reported
1. Pilot
2. Evidence
3. Decision
4. Operation
5. Monitoring
6. Incident, change or learning
7. Reassessment
8. Retire
Turning experience into impact
Better decisionsBased on real outcomes Stronger controlsInformed by actual risks Greater valueFor users and public services More inclusiveIdentifying and removing barriers Smarter spendingUnderstanding cost and benefit Shared learningReusable patterns and guidance
Report anything that went wrong or nearly did: a wrong output that was acted on, an exposed file, an unexpected action, a cost spike, an accessibility barrier. Record:
Agent journey
An agent that helps one person and an agent that decides a citizen's case are not the same thing, and should not carry the same controls. Eight stages, each with a gate question, the control that applies, the role accountable and the outcome expected.
Stage 1 of 8 Personal / low risk
Security & identity — least privilege, access control and monitoring
Creator — builds and tests the agent; owns initial purpose
Clear purpose and value Right controls at the right time User engagement and feedback Sustainable cost model Continuous learning and improvement Strong community and reuse Better public value outcomes
A managed progression that balances innovation, control, risk and value. Agents evolve from personal productivity tools to trusted, supported services — and ownership moves with them.
1 Idea & personal experiment Individual creator Low risk · personal
2 Share & team pilot Team data owners Low to moderate risk
3 Managed team agent Named team owner Moderate risk
4 Departmental capability Departmental owner Moderate to high risk
5 Operational service Service owner High risk
6 Enterprise service Business / enterprise owner High risk · high impact
7 Cross-government capability Cross-government sponsor & owners High impact · strategic
A continuum of assistance and autonomy — from individual productivity to enterprise agentic services. Purpose: enable people to work better today, orchestrate intelligent workflows, and deliver trusted, efficient, inclusive public services.
Tier 2 of 7 · Microsoft 365 Copilot
AI in your apps to summarise, draft, analyse and explore
TYPICAL USERSIndividual
AUTONOMY LEVELHuman-guided
SCOPEPersonal
EXAMPLESWord drafting · Excel analysis · Meeting summary
GOVERNANCEOrg policies
How a small, helpful agent becomes business critical without visibility or control. Small agents can create big value — and big risk. Visibility, ownership and governance early in the journey prevent unmanaged tools from becoming unmanageable dependencies.
1 Personal experiment Low visibility An individual creates an agent to solve a personal problem or save time.
2 Shared with team Low to moderate visibility Shared with colleagues or a small team to help with common tasks.
3 Wider adoption Rising visibility More teams discover it and start using it organically.
4 Dependence builds Rising to high visibility Critical work starts to rely on the agent and its outputs.
5 Business critical High visibility Used across the department or service. Business processes become dependent on it.
6 Invisible criticality Gap in visibility No formal ownership, limited documentation, unknown controls, unknown costs.
7 High risk exposure Critical · unmanaged Risk of failure, data breach, inaccurate decisions or service disruption.
8 Event of failure Loss of trust and value Agent breaks, data changes or owner leaves — causing a real operational impact.
1. Discover early Agent inventory · Usage monitoring · Sharing visibility · Community reports
2. Require description Purpose & use case · Data sources · Support owner · Review dates
3. Assign ownership Business owner · Technical owner · Information owner · Risk owner
4. Assess & classify Risk & impact · Data sensitivity · Criticality score · Assurance needs
5. Control & assure Approved sources · Guardrails · Human oversight · Testing & validation
6. Monitor & review Usage & performance · Cost & consumption · Accuracy & quality · Incident tracking
7. Promote or remediate Move to managed platform (e.g. Copilot Studio) · Or retire / remediate
Shadow IT Inconsistent quality Lack of documentation Knowledge silos Unknown cost Compliance breach Loss of public trust
Agent employment lifecycle
Agents go through a managed lifecycle — from intent to retirement — with controls, oversight and accountability at every stage. People remain accountable: agents augment human capability, they do not replace judgement, responsibility or accountability.
Identify the need, purpose and expected outcomes.
Key controls
Assess capability, risks, data needs and alignment with policy and standards.
Key controls
Configure the agent, set permissions, guardrails and connect data and tools.
Key controls
Operate in a controlled pilot with close human oversight and limited scope.
Key controls
Agent operates within its delegated authority to support people and processes.
Key controls
Regularly review performance, accuracy, impact, risks and value delivery.
Key controls
Update knowledge, models, prompts and controls based on learning and feedback.
Key controls
Restrict or pause the agent if risks increase, rules are breached or performance drops.
Key controls
Retire the agent when no longer needed or superseded. Revoke access and archive records.
Key controls
Retain records for audit, transparency and learning in accordance with retention rules.
Key controls
Continuous monitoring Human oversight Audit logging Policy compliance Security Privacy Safety
An agent may be able to do many things, but can only do what it is authorised to do. Capability is about potential; authority is about permission. Capability + authority + context = action — only when all controls are satisfied can the agent act.
Technical potential and skills: the full range of capabilities, knowledge, tools and actions the agent is built or enabled to perform.
Natural language understanding Tool use & API integrations Reasoning & problem solving Workflow orchestration & scheduling Data retrieval (RAG) Analysis & visualisation Code generation & automation Communication (drafting, email, chat) Content creation & summarisation Learning from feedback
Bridging capability to appropriate action.
Policies & rulesOrganisational policies, legal requirements and ethical guardrails define boundaries.
Delegated authorityExplicit permissions granted for specific tasks, data and systems.
Conditions & controlsContext, thresholds, approvals and risk checks must be satisfied.
Human oversightPeople remain in control of decisions and can intervene or override.
Monitoring & auditActions are logged, monitored and reviewed for assurance and compliance.
Permitted actions in the right context, within defined boundaries, for specific purposes and users.
Draft policy briefingApproved for Policy Team (sensitivity: OFFICIAL)
Send standard email responseApproved with template and review
Summarise documents up to OFFICIALApproved with data classification filter
Update internal tracker or systemApproved with role-based permissions
Escalate issue to human reviewerRequired when confidence below threshold
Reduces risk Limits exposure by ensuring agents cannot act beyond intent or policy.
Improves accountability Clear ownership and delegated authority for all actions.
Supports compliance Meets legal, regulatory and ethical obligations.
Enables agility Adjust permissions without rebuilding capabilities.
Builds trust Transparent controls and oversight build confidence.
Adapts to context Decisions and permissions can vary by role, risk and situation.
Public service impact cascade
Seven links from capability to legitimacy — each with the elements that must hold and the risks that arise if they do not. Better governed AI agents lead to better decisions, better services, and better outcomes for everyone.
The agent has defined capabilities, knowledge, tools and guardrails within its delegated authority.
Key elements
Risks at this stage
Civil servants interact with the agent, using its insights, suggestions or automation to support their work.
Key elements
Risks at this stage
The agent's output informs or automates parts of operational decisions and actions within the service process.
Key elements
Risks at this stage
Operational actions produce internal service results such as case resolution, guidance issued, or processing completed.
Key elements
Risks at this stage
Citizens, businesses and communities experience the service through interactions, outcomes and communications.
Key elements
Risks at this stage
Better outcomes are achieved for people and society, improving trust in services and contributing to public value.
Key elements
Risks at this stage
Transparent, accountable and responsible use of AI strengthens public trust and democratic legitimacy.
Key elements
Risks at this stage
Strong governance & ownership Policies & ethical use Security & privacy by design Quality data & provenance Monitoring & auditability Human oversight at every step Transparency (ATRS) Performance measurement Continuous improvement
Observatory & barometer
The observatory tracks published signals of UK public-sector AI adoption; the model is the OECD.AI Policy Observatory, applied at home. Each figure carries its publisher and date. The barometer is your own assessment, stored on this device only.
OECD.AI · 2026
ATRS transparency records published. Mandatory reporting for central departments and ALBs since 2025.
Civil Service · 2026
252
ideas entered in the Civil Service AI & Data Challenge 2026 — a 160% increase, in its fourth year.
DSIT analysis · 2026
AI investment pledged into the UK since January 2025.
delivery.ai.gov.uk · 2026
UK AI compute in ExaFLOPs: 2025 actual to 2030 commitment (from 2 in 2024).
DSIT analysis · 2026
200k
People studying AI-related higher-education programmes in the UK.
Heriot-Watt & GoLLM · 2025
Scottish councils piloting or using AI. The leading UK nation for local-government adoption.
Rate each dimension 0–4. Saved locally on this device; nothing leaves the page.
0 of 8 dimensions rated
Part 6 of 12 · 4 chapters
Cost, value & investment · illustrative figures
A continuous cycle that connects business value, cost efficiency and responsible use. Align, optimise, govern — then deliver and reinvest. Figures on this page are illustrative: replace them with your own contract and telemetry data.
| 1 Forecast & plan Understand demand, use cases, models and unit economics. | 2 Allocate & budget Allocate costs, set budgets and guardrails. | 3 Observe & measure Monitor usage, costs, quality and performance in real time. | 4 Analyse & evaluate Analyse spend, quality, outcomes and business value by use case. |
| 5 Optimise & improve Optimise models, workloads, prompts and architecture. | 6 Report & transparency Report on spend, value, carbon and compliance. | 7 Reinvest or rebalance Reinvest in high-value initiatives or rebalance priorities. | 8 Govern & control Review policies, controls and risks. Adapt and evolve. |
Consumption costsTokens, credits, API calls, capacity, compute, storage, network
Technology costsModels, platforms, connectors, licences, infrastructure
Adoption costsTraining, change, champions, workflow redesign, support
Data & integration costsData preparation, quality, integration, access and governance
Assurance & risk costsSecurity, privacy, testing, human review, audits, incident response
Operational costsMonitoring, tooling, run support, maintenance, model updates
Transition & exit costsMigration, parallel run, data portability, vendor switching
Better public servicesImproved outcomes for citizens
Productivity gainsTime saved, efficiency improved
Quality & consistencyFewer errors, higher quality outputs
Risk reductionStronger controls, fewer incidents
Transparency & trustClear oversight, explainable decisions
Sustainable impactLower waste, optimised resources, lower carbon
Business owners Product & delivery Finance / FinOps Engineering & data Security & privacy Procurement & commercial Operations & service management
Total cost per interaction = token cost + infrastructure + tool and API cost + human effort + overhead. Multiply by volume for the period, and set it against the value delivered. The starting values are the source model's worked example.
Input tokens 2,000 tokens @ £0.002 / 1K £0.0040
Output tokens 1,000 tokens @ £0.006 / 1K £0.0060
Tool / API calls External calls and connectors £0.010
Infrastructure Compute, storage and network allocation £0.0050
Human effort 5 minutes @ £1 / min £5.00
Overheads Operations, support and monitoring £0.500
Total cost per interaction £5.53
× 1,000 interactions in the period £5,525.00
Total value £20.00
NET VALUE £14.48 RETURN 262%
Value to cost ratio 3.6× · value across the period £20,000.00. Human review time is usually the largest single line — design it down before changing model.
User or system initiates a task
Tokens processed by the model
What drives the cost of the interaction
Quality of output and outcome achieved
Human time to review, edit or redo
Value delivered to the organisation
Right modelMatch quality needs to the most cost-effective model
Prompt engineeringOptimise prompts to reduce tokens and improve outcomes
Context managementRight-size context, use retrieval and caching
Routing & policiesRoute to the best model or tier based on cost and risk
Caching & reuseReuse responses, embeddings and results
Evaluation & guardrailsMeasure quality and set thresholds to avoid waste
FinOps controlsBudgets, alerts, showback / chargeback and governance
Cost per 1,000 tokens Cost per interaction Cost per outcome Quality score Rework rate Time saved ROI / value per £ Budget adherence Token utilisation efficiency
Fair, transparent, actionable. Allocate costs across the organisation to drive accountability, optimise spend and demonstrate value. Without a named owner, a layer is not allocated — it is absorbed.
OWNER CTO / enterprise Direct allocation by actual cost
OWNER Platform / centre of excellence Cost pools and apportionment drivers
OWNER Product / service owner Showback or chargeback to products
OWNER Team lead Team budgets and workload tagging
OWNER Use case owner Usage-based allocation (tokens, calls, transactions)
OWNER Business / service Unit cost per user or beneficiary
Licences, tokens and compute are usually budgeted and tracked, and are only a fraction of the total. The hidden lines below the waterline determine real value and long-term success.
Data & information
Integration & engineering
Security & compliance
Adoption & change
Assurance & evaluation
Operations & support
FinOps & management
Governance & risk
Workforce & roles
Sustainability
Transition & exit
Value realisation
Total cost of AI = technology + data + people + process + assurance + operations + commercial. Budget for the full cost. Manage it actively. Realise the value.
ComplexityMany moving parts across technology, data and people
ChangeWorkflows, roles and ways of working must adapt
Risk & assuranceHigher risk requires stronger controls and evidence
Scale & usageMore users, data and interactions increase ongoing costs
GovernancePolicies, oversight and reporting add essential investment
AI adoption is a long-term investment across technology, people, process, data and assurance. Sustainable value comes from funding the whole system, not the licence.
1. Discover & plan Identify needs, use cases and business value. Build the investment case.
2. Pilot & experiment Test, learn and validate value, risk and cost. Measure and refine.
3. Scale & adopt Roll out to teams and services. Build capability.
4. Operate & optimise Monitor usage, value and cost. Optimise models and processes.
5. Sustain & evolve Continually improve, adapt and reinvest for future value.
Central funding Departmental budgets Programme funds Shared services Outcome / value based funding Public-private partnerships
Adoption rate Usage & engagement Quality & accuracy Cost per transaction Time saved Outcome & impact
FinOps & models
AI cost is a portfolio, not a licence line. You are billed per token, in and out — roughly four characters or three-quarters of a word each; an A4 page is 600–800 tokens. Micro-costs stack when an agent runs on repeat. Report by service, product, agent, owner, model and cost driver; showback before chargeback; no expenditure classification without finance-led assessment.
Licences Credits Model calls Compute Storage Connectors Monitoring Evaluation Adoption & training Remediation Security Human review Support
Cost: a per-user licence ($30 per month list) covers Chat and Copilot in the Office apps. Cowork, Code, Autopilot and frontier models consume Copilot Credits at $0.01 each on pay-as-you-go. Agent security for custom agents needs Agent 365.
Residency: follows the tenant's Microsoft 365 geography. Web grounding, subprocessor models (Claude, Grok) and Code apps on Managed Runtime need separate checks.
Cost: a seat per user, plus usage at token rates by model (Opus 5.5 lists at $4 input / $20 output per million tokens). Model routing, effort settings and prompt caching are the main levers.
Residency: set by deployment route. Anthropic-processed with a US-only option, or in-region through Bedrock, Google Cloud or Foundry. Connectors carry the content they fetch to the model.
Illustrative list prices from the "Where your AI spend goes" explainer (2025), GBP per 1,000 tokens — verify current pricing before any decision. Non-English output can cost more: the same words take more tokens.
GPT-4o class (OpenAI) Balance of quality and tone £0.017per run £1.71per month
Claude Haiku class (Anthropic) Cost-efficient, high-volume work
£0.0043per run
£0.427per month
Claude Opus class (Anthropic) Deep, high-context threads
£0.026per run
£2.56per month
Mixtral class (Mistral) · cheapestUltra-low cost; needs prompt tuning
£0.0032per run
£0.320per month
Gemini Pro class (Google) On par with GPT-4o by usage tier
£0.017per run
£1.71per month
Design levers before model choice: reuse outputs, cache repeated answers, trim inputs, use plain logic before calling a model.
A 10% findability gain on ~1.8 hrs/day spent searching. The big lever — and it can cover the licence bill 2.5–8× over ("earned allocation").
Storage cost ~£1.5k / yr Effectively £0 marginal under bundled M365. Symbolic.
Modest; aligns with Greening Government; curbs future growth.
Interrupted every ~2 minutes; 48% find work chaotic and fragmented. The human case.
Rules that hold: require total-cost and portability evidence before scaling any tool or agent · per-agent consumption caps and anomaly alerts · unit economics per production agent published within 12 months · value measured through the evidence chain: activity → output quality → workflow change → organisational outcome → public outcome.
Commercial barriers · illustrative default scores
Score each barrier by its impact and its likelihood in your context. Risk score is impact × likelihood; the band follows the published scale. The defaults are the sector-level assessment — change them and the bands move.
4 barriers are critical and 8 high in this scoring. Address these before scale.
BARRIER IMPACT LIKELIHOOD I×L RISK TYPICAL CURRENT STATE
Licensing clarity Unclear boundaries between included features, credits, capacity and pay-as-you-go.
Cost & usage transparency Limited access to usage data, telemetry, forecasting and billing detail.
Price predictability Usage-based pricing makes budgets volatile and hard to forecast.
Supplier / platform lock-in High switching costs, proprietary connectors and formats.
Telemetry & audit rights Lack of contractual rights to monitor, audit and export operational data.
Data ownership & IP Uncertainty over prompts, outputs, data use, training and generated code.
Data residency & jurisdiction Data and model processing across jurisdictions may create compliance risk.
Service level & support Unclear SLAs, responsibilities and support across multiple products and models.
Interoperability & portability Limited open standards, export options and easy integration with other systems.
Commercial complexity Complex contracts, overlapping costs, multiple components and change over time.
Small body disadvantage Smaller organisations lack commercial leverage and specialist expertise.
Exit & transition risk Uncertain exit costs, data extraction, migration and parallel running.
Bands are computed from the published impact × likelihood scale. Two rows on the source board carry a stated risk level that does not match its own band scale; the computed band is shown here.
Match the approach to the need, risk, control, value and total cost. Decide wisely, start small, scale with confidence — and revisit the decision as needs, risk and technology evolve.
1. Define the needWhat problem are we solving?
2. Assess contextUsers, data, risk, regulation, criticality
3. Evaluate optionsUse the criteria to compare
4. Decide & planSelect approach, plan delivery
5. Review & adaptMeasure outcomes, learn and adjust
Moderate complexity / risk
Use existing platform features or pre-built agents with configuration.
Common needs, moderate risk, fast value
Microsoft 365 Copilot, pre-built agents, Copilot Studio templates, AppSource solutions
Strategic value Risk & assurance Control & flexibility Total cost (TCO) Time to value Reusability & scale AVG Personal configuration 2 2 4 5 5 1 3.2 Use platform 3 3 3 4 4 3 3.3 Configure / extend 4 3 4 3 3 4 3.5 Build custom 5 4 5 2 2 3 3.5 Share / buy 4 4 3 4 4 5 4.0
Very high (strong) High (good) Medium (acceptable) Low (weak) Very low (poor)
1. Monitor usage, value and risk 2. Evaluate outcomes and costs 3. Learn and improve approach 4. Adapt or reinvest
A continuous loop that turns real-world experience across government into stronger commercial engagement, better terms and improved public value. Open once, use many times.
Inputs
Departments & public bodies Devolved governments Local government & combined authorities Health & care Police, justice & emergency services Education & skills Regulators & public agencies Communities of practice & forums Audits, reviews & external reports Financial & FinOps data
Outputs
Better contracts & terms Cost transparency & predictability Stronger controls & governance Interoperability & portability Improved supplier performance Shared patterns & reusable assets Evidence-based policy & standards Better outcomes for citizens
Evidence led Real experience, not assertion
Collaborative & inclusive Across departments, nations and sectors
Fair & proportionate Terms matched to risk and scale
Transparent & accountable Open about costs, terms and outcomes
Open once, use many times Publish so others do not repeat the work
Continuously improve The loop does not stop at signature
Dependency, lock-in & exit
AI-enabled services rely on a chain of suppliers, platforms, models, data sources and infrastructure. Visibility of these dependencies is what makes risk management, resilience, portability and value possible.
AI-enabled public service / use case
Knowledge assistant Case summary Policy drafting Meeting intelligence Citizen support Code assistant Workflow automation Analytics & insight
Applications & platforms
Microsoft 365 Copilot Copilot Studio / Agent Builder Teams, SharePoint, Outlook Power Platform (connectors, apps, automate) Foundry / Azure AI services Third-party applications Custom applications
Models & AI capabilities
Microsoft models OpenAI models Anthropic models Google models Open source models Specialist / domain models Future / emerging models
Cloud & infrastructure
Microsoft Azure (compute, storage, networking) Azure regions & availability zones GPU / compute capacity providers AI services (API management, safety, content filters) Identity & access (Entra ID) Monitoring & observability Backup & disaster recovery
Data & information
SharePoint libraries Teams & collaboration content Line-of-business systems Open data & external sources Reference data & master data Vector stores & indexes Logs, telemetry & feedback
Third parties & suppliers across the ecosystem
Model providers Cloud provider Software vendors API / service providers Data providers Integration partners Support & operations Consulting & implementation
Platform dependencyReliance on specific platforms and ecosystems
Model dependencyReliance on model providers and availability
Data dependencyAccess to quality, timely and trusted data
Infrastructure dependencyCompute, storage, network and region reliance
Supplier & contractual dependencyThird-party services, SLAs, pricing and contractual terms
Operational dependencySkills, support, monitoring and continuity
Better risk management Resilience & continuity Cost optimisation Informed supplier strategy Sustainable innovation Public value protection
Avoid unintended dependency. Preserve portability across data, models, agents, integrations and knowledge, so you can adapt, switch or exit when you need to.
PortabilityKeep your data, prompts, models and agents portable
TransparencyKnow what you use, where it is, and the terms
InteroperabilityUse open standards and avoid proprietary traps
Plan aheadDesign exit and migration paths from day one
Control costsUnderstand true switching and transition costs
ContinuityMaintain service and compliance through change
Technology dependencies
Data & knowledge dependencies
Model & AI dependencies
Operational dependencies
Commercial dependencies
Organisational dependencies
Rate how prepared you are to exit. The path above is the work; this is the readiness.
Data Metadata & schemas Documents & content Prompts & instructions Agents & workflows Configurations & settings Integrations & APIs Access & permissions Audit logs & evidence Knowledge & context
Part 7 of 12 · 11 chapters
The team
The Service Manual's multidisciplinary team, extended for AI. Not every role is full-time, but every role is engaged from discovery, not summoned for sign-off. "Critical at" marks where each role's absence hurts most.
Accountable and delivery leadership
Accountability, benefits, risk appetite. The name on the assurance forms.
Critical at · Discover → Operate
Problem framing, value, prioritisation of the backlog.
Critical at · Discover · Design
Cadence, dependencies, unblocking the team.
Critical at · Build · Deploy
Fit with the wider portfolio; stops duplicate AI pilots.
Critical at · Discover · Deploy
Data and AI engineering
Modelling choices and honest evaluation design.
Critical at · Design · Build
Pipelines, integration, MLOps, guardrails in code.
Critical at · Build · Operate
Data quality, provenance, pipelines that last.
Critical at · Design · Build
System design, interoperability, exit routes.
Critical at · Design
Risk, rights and control
Threat modelling and NCSC-aligned controls.
Critical at · Design · Deploy
DPIA, lawful basis, records of processing.
Critical at · Discover · Design
Public law, IP and copyright, contracts.
Critical at · Design · Deploy
Bias, participation, societal impact.
Critical at · Discover → Operate
Users, domain and money
Real user needs, and real harms, observed rather than assumed.
Critical at · Discover · Operate
The end-to-end journey, including the failure paths.
Critical at · Design
Casework truth, edge cases, what the data actually means.
Critical at · Discover → Operate
Procurement routes, vendor terms, ethical parity for third parties.
Critical at · Discover · Deploy
Skills pipeline. Recruit and upskill through the Government Digital and Data profession's capability framework, Civil Service AI training and the cross-government AI community. Playbook principle 9 makes this the deploying organisation's responsibility, not the vendor's.
Multidisciplinary AI service team
Collaborating to design, build, govern and operate trustworthy AI services that deliver public value — through design and plan, build and enable, deploy and operate, measure and improve. The service at the centre is defined by user needs, public value, trust, safety and outcomes.
User-centredDesign with and for users at every stage.
InclusiveAccessible, equitable and easy to use.
Trust & safetyBuild security, privacy and ethics in from the start.
CollaborativeWork together, share ideas and challenge constructively.
Agile & iterativeLearn quickly, adapt and improve continuously.
Outcome focusedMeasure what matters and deliver public value.
Roles & pathways
Role-specific learning journeys to build skills, confidence and impact at every stage of adoption. Pick the role you are supporting; the pathway shows what capability looks like at each of four levels.
YOUR JOURNEY
Understand AI basics and opportunities
Apply AI tools and techniques with guidance
Design, adapt and lead AI solutions
Shape strategy, standards and innovation
Learning & developmentRole-based curricula, microlearning, labs, communities of practice
Tools & platformsApproved tools, secure sandboxes, reusable components
Guidance & standardsPolicies, playbooks, architectures, guardrails and templates
Measuring progressSkills assessment, certification, impact and value metrics
Recognition & careerCareer pathways, communities, badges and incentives
Role relevant Right skills for the job
Practical Learn by doing in real work
Progressive Build from foundations to mastery
Trusted Safe, secure and responsible AI use
Measurable Evidence of competence and impact
Inclusive Accessible and supportive for all
AI augments people, it does not replace them. Roles evolve to focus on judgement, creativity, oversight and continuous improvement.
ROLE TRADITIONAL FOCUS AI ENABLES AI-ENABLED FUTURE FOCUS IMPACT
Policy adviser
AI summarises evidence, suggests options, drafts content, finds patterns
Better policy quality and impact
Contact centre adviser
AI handles routine queries, retrieves answers, suggests next best action
Improved citizen experience
Software engineer
AI generates code, finds bugs, recommends tests, documents changes
Faster delivery, higher quality
Data analyst
AI prepares data, generates insights, builds visuals, detects anomalies
Stronger insight and decisions
HR advisor
AI answers FAQs, guides processes, drafts documents, checks compliance
Better employee experience
Risk manager
AI identifies risks, analyses scenarios, monitors controls, suggests mitigations
Stronger risk management
Success depends on people, process redesign, culture, leadership and trust — supported by responsible AI and strong governance.
AI workforce capability academy
Five learning pathways across four stages. Build confidence. Apply skills. Drive impact. Lead change.
FoundationUnderstand the basics BuildDevelop practical skills AdvanceApply and lead LeadExpertise and influence
Set direction and provide oversight
Foundation AI essentials for leaders Opportunities, limitations · Risks and public value · Responsible AI principles
Build AI governance & accountability Oversight responsibilities · Risk appetite and decision rights · Human control and assurance
Advance Strategic AI leadership Portfolio management · Measuring outcomes and impact · Ethics, bias and inclusion
Lead Shaping the AI agenda Leading culture change · Stakeholder engagement · Policy influence and advocacy
Design and deliver AI-enabled services
Foundation AI for product & service teams User needs and problem framing · AI use cases and value · Ethical and inclusive design
Build Designing AI-enabled services Journey mapping with AI · Prototyping and evaluation · Managing suppliers
Advance AI product management Prioritisation and roadmapping · Data-informed outcomes · Scaling AI responsibly
Lead Leading AI transformation Cross-team leadership · Change management · Driving adoption and value
Build and run secure, reliable AI agents
Foundation AI & data fundamentals Data types and quality · Intro to ML, LLMs and agents · Privacy and security basics
Build Building AI agents Prompt engineering · RAG and tool integration · Testing and evaluation
Advance Advanced engineering Agent orchestration · Model evaluation & guardrails · Red teaming and robustness
Lead Technical excellence Architecture leadership · Innovation and emerging tech · Open standards and reuse
Assure, audit and manage risk
Foundation AI governance basics Policies, standards and roles · Risk types and controls · Transparency and ATRS
Build Assessing and controlling risk DPIA for AI systems · Bias, fairness and inclusion · Security threat modelling
Advance AI assurance & audit Assurance frameworks · Monitoring and evidence · Incident and issue management
Lead Assurance leadership Independent challenge · Audit strategy and reporting · Regulatory horizon scanning
Operate, monitor and support users
Foundation Working with AI tools Everyday AI use · Human oversight basics · Data handling and security
Build Operating AI services Monitoring and logging · Incident reporting · User support and feedback
Advance Optimising operations Performance analysis · Continuous improvement · Knowledge sharing
Lead Operational excellence Service reliability leadership · Automation and productivity · Building a learning culture
Responsible AIEthics, fairness, transparency
Security & privacySecure by design, data protection
Human oversightAccountability, challenge, escalation
Data literacyQuality, lineage, classification
Accessibility & inclusionDesign for everyone, equitable outcomes
Continuous learningIterate, adapt and improve
Instructor-led workshops E-learning modules Hands-on labs & sandboxes Communities of practice Certification & badging Playbooks & job aids
Aware Practitioner Expert Leader
Capability & maturity · illustrative frameworks
Seven levels, from ad-hoc use to adaptive capability. Choose a level to see what it looks like across the dimensions and the twelve capability domains — then record the evidence you could actually show.
Defined governance and controls. Repeatable patterns, registered use cases and measured pilots.
| Leadership Accountable owners | Governance Controls, risk tiers, assurance gates | People & skills Role-based skills |
| Data & knowledge Catalogued, governed access | Technology Reference architecture, model choice | Value & outcomes Measured value from pilots |
Strong Independently validated
Moderate Documented internal evidence
Emerging In pilot or partially proven
Weak Anecdotal or not evidenced
Not assessed No evidence available
AI strategy aligned to business objectives and risk appetite
Role-based skills framework and capability plans
Use case portfolio managed with benefits tracking
Target architecture, standards and integration patterns
Governed data assets, metadata and lineage
Security-by-design; identity, access and monitoring in place
Responsible AI framework and assurance process
Managed suppliers, contracts and risk assessments
Cost visibility, allocation and showback
Defined operating model and SLAs
Benefits tracking; outcomes measured
Communities of practice and curated knowledge
Policies & strategies Architectures & standards Training records Use case business cases KPIs & dashboards Audit & assurance reports Lessons learned & patterns
Current against required capability, scored 5 (critical gap) to 1 (strong, sustain). Illustrative sector-level assessment — replace with your own to prioritise investment. Filter to one role group to read a single column.
| CAPABILITY DOMAIN | Executives & leaders | Business & service owners | Product & delivery managers | Data & info professionals | Engineers & technical teams | Security & risk professionals | Commercial & procurement | Finance & FinOps | Frontline staff & practitioners |
| Strategy & leadership Vision, governance, value | 2 | 4 | 4 | 4 | 3 | 3 | 3 | 4 | 4 |
| People & change Adoption, culture, engagement | 3 | 3 | 3 | 3 | 4 | 4 | 3 | 3 | 4 |
| AI & data literacy Understanding AI, data and information | 3 | 4 | 3 | 2 | 3 | 3 | 4 | 3 | 5 |
| Technology & engineering Tools, platforms, integration | 3 | 3 | 3 | 3 | 2 | 3 | 4 | 4 | 4 |
| Data & information Quality, access, governance | 4 | 4 | 4 | 2 | 3 | 3 | 3 | 3 | 4 |
| Security, privacy & risk Threats, controls, compliance | 3 | 3 | 3 | 4 | 3 | 2 | 3 | 3 | 4 |
| Responsible AI & ethics Fairness, transparency, human oversight | 3 | 3 | 4 | 3 | 3 | 3 | 3 | 3 | 4 |
| Commercial & procurement Contracts, suppliers, value | 3 | 3 | 4 | 3 | 3 | 3 | 2 | 4 | 4 |
| Finance & FinOps Cost management, allocation, value | 3 | 4 | 3 | 3 | 3 | 3 | 3 | 2 | 3 |
| Operations & service management Monitoring, support, continuity | 3 | 4 | 3 | 3 | 3 | 3 | 3 | 3 | 3 |
| Measurement & outcomes KPIs, impact, benefits — not scored on the source board | — | — | — | — | — | — | — | — | — |
Critical gap High priority
Major gap Act soon
Moderate gap Plan to address
Minor gap Monitor
Strong / no gap Sustain
AI & data literacy for frontline staff
Data & information for executives, business owners and product managers
Technical teams are relatively strong but need continued investment in security by design
Role-based learning pathways Communities of practice and peer learning Targeted coaching and enablement Tools, templates and reusable assets Time, budget and leadership support
Inclusive & accessible adoption
AI tools should empower everyone by offering choice, clarity and support. The spectrum below runs from empowered and confident to apprehensive — and inclusion is not one size for all: listening, choice and support help everyone succeed with AI.
Sees AI as a productivity booster and enjoys exploring new tools.
Interested and willing to try, but wants reassurance and guidance.
Wants to benefit from AI but needs help understanding how and when.
Feels unsure or overwhelmed by change, complexity or risk.
Does not see value or may distrust AI or the organisation's intent.
Choice & controlLet users choose how AI helps, when and in what way
Clear & simpleUse plain language, short explanations and examples
Flexible interactionSupport different workflows, speeds and preferences
Neuroinclusive designReduce cognitive load, provide structure, minimise distractions
Accessible by defaultWCAG compliant, multimodal, and assistive-tech friendly
Safe & trustedProtect data, explain limits, and make AI behaviour predictable
Human supportEasy access to help, feedback channels and champions
Leadership sponsorship Digital & AI champions Inclusive communication Role-based learning paths Protected learning time Local support networks Feedback & continuous listening Measure & adapt for all users
Inclusion is not one size fits all. Listening, choice and support help everyone succeed with AI — and inclusive adoption empowers people, strengthens trust and delivers better public services.
Working together across communities to share knowledge, avoid duplication and build trusted, reusable AI and Copilot capabilities. The weekly cross-government call is the open forum at the centre.
Hands-on development, experimentation and engineering excellence.
AI Engineering Lab Data Science & AI Community Low Code Community Microsoft 365 Technical Community Prototyping and Sandboxes
User enablement, organisational change and the digital workplace.
Digital Workplace Community Change & Adoption Network AI Champions Network Learning & Capability Network Comms & Engagement Network
Data governance, catalogues and information management excellence.
Data Management Network Data Catalogue Community Information Governance Network Metadata & Standards Community Knowledge Management Community
Ethics, AI governance and assurance across the public sector.
Cross-Government Data & AI Ethics Community AI Assurance Network Security & Privacy Community Accessibility & Inclusion Network
Policy, standards, guidance and compliance.
Government Digital Service (GDS) DSIT AI Policy & Standards Government Analysis Function Identity & Access Community Service Design Network
Commercial strategy, vendor engagement and cost insight.
Commercial Community FinOps Community Supplier Engagement Community Contract & Licensing Community Procurement Policy Community
Cyber security, risk management and resilience.
Cyber Security Network Risk & Assurance Community Resilience Community
Adoption, change, champions and enablement.
Copilot Weekly XGov Call Community of practice sessions Show and tell forums
Central government departments & ALBs Devolved administrations Local government & combined authorities Health & care network Education network Policing & justice network International & supranational links
Share once, use many timesPublish once so others do not repeat the work
Avoid duplication and reworkFind what exists before building it again
Consistent standards, safer adoptionCommon expectations across organisations
Better value and lower total costAggregate demand and reuse assets
Faster innovation with confidenceLearn from those who have already tried
Better outcomes for citizensThe point of all of it
Journeys · adoption operating model
A system of people, processes and support, not a rollout. Leadership sets direction. Business owns outcomes. Communities enable change. Data and evidence drive continuous improvement.
Inputs
Strategy & priorities User needs & frontline insights Data, content & tools Policy & regulatory requirements Evidence & external intelligence Budget & investment
Outcomes
Better public services Improved user experience Lower cost, higher productivity Stronger compliance and trust Sustainable and adaptable Measurable public value
1. Observe usage, quality, cost, risk 2. Analyse results, feedback, insights 3. Learn what works, what doesn't 4. Decide prioritise actions and changes 5. Act improve, enable, remove barriers 6. Repeat continuous improvement
Leadership Set direction, sponsor change and unblock
Business owner Define outcomes, prioritise use cases and realise value
Product / service manager Manage lifecycle, measure value and iterate
Champion network Advocate, share learning and support peers
Data & information professional Ensure quality, access, lineage and governance
Security & privacy Manage risk, protect data and users
Commercial & FinOps Manage contracts, licensing, cost and value
Change & adoption team Drive engagement, training and communications
Technical team Build, integrate and operate solutions
Frontline users Use, feed back and help shape improvements
The person-level journey
Six stages from awareness to community contribution — a progression, not a course. The continuous improvement loop runs across all six.
Stage 1 of 6 All staff
Build awareness of AI basics and potential in public services.
Leadership and sponsorship Clear guidance and guardrails Accessible tools and environments Role-based learning and support Time to learn and experiment Recognition and career pathways Feedback and user research Measurement and value realisation
Copilot programme · agent rollout
The opportunity is significant, but ownership, risk and consistency matter.
More than simple prompts, built well, they can support processes, provide advice, surface information and shape decisions.
Every agent needs a clear purpose, an accountable owner and active lifecycle management. Risk management is critical.
Through a culture of safe experimentation, coupled with close monitoring, we will find opportunities to deliver the best agents, while avoiding “agent-aggedon”.
Innovate confidently. Manage risk. Deliver value. Together.
Delivering
Public Beta and measured communications ramp up gives the organisation space to learn about the technology, ownership and responsibility before broader scale.
Copilot Agent Builder is one of many business tools
Assurance requirements vary by risk and operational impact
Business areas tailor guidance to their environment
Proportionate guardrails create confidence to experiment.
Put standards and guidance in place early.
Set common expectations; keep responsibility with agent owners.
Apply controls according to use case and data sensitivity.
Evolve the model using real scenarios and evidence.
Technology enablement is paired with readiness, guidance and local ownership.
What colleagues can use, when and how to work within the guardrails.
Initial soft launch, then expand messages around use cases and value. Align to OBT.
Digital Champions, Sponsors and business representatives help locally.
Options from “getting started” through to building agents, focused on real use cases.
Tailored help to embed useful agents into team workflows. May extend to Tiger Teams.
Reporting is not the outcome. It is the feedback loop that helps us decide what to support, change, scale or retire.
Builders Who is creating agents?
Ownership Are agents visible and actively owned?
Usage Which agents are being used?
Enablement Are learning and communications changing behaviour?
Value What outcomes and benefits are emerging?
Business takes ownership of their use cases and determines the best tool for the job
Local resources to take action on the measures
People and adoption · community
How a cross-government community matures: the engagement arc, the flywheel that turns experience into shared assets, and the honest exchange of what members give and get. Capability without community repeats mistakes; community without operationalisation produces discussion.
How we engage
Greater clarity & confidence Less ambiguity about products, controls and next steps.
Reduced risk & duplication Shared lessons prevent repeated mistakes.
Reusable knowledge Patterns, templates and guidance that can be adapted.
Faster & safer implementation Less rework, clearer pathways, better preparation.
Stronger supplier engagement Consolidated questions and influence based on community insight.
Continuous learning loop Experience Insight Guidance Action Impact Feedback Improved practice
The more we share, the more we all improve.
Surface real needs and uncertainties
Contribute what you have seen and tried
Learn together and test ideas
Turn learning into shared assets
Put learning into practice locally
Share results and what is next
What members contribute, what the community turns it into, and what everyone gains.
Our questions & challenges Share real problems, uncertainties and barriers
Use cases & demos Show what we are building, testing and learning
Successes & lessons Share what works, what did not and why
Policies & approaches Contribute guidance, guardrails and governance ideas
Tools & templates Share artefacts, patterns, checklists and assets
Commercial & cost insight Raise licensing, FinOps and procurement experiences
Security, privacy & assurance Share risk findings, controls and assurance practice
Our time & expertise Offer help, review, connect others and support each other
Open discussions & peer challenge Honest conversations that surface real issues
Shared evidence & patterns Collective intelligence from across government
Expert input & supplier access Access to Microsoft, experts and the wider network
Connections & collaboration Find partners, avoid duplication and work together
Influence & impact Consolidate our voice to shape products, policy and standards
Capability uplift & confidence Improve skills, maturity and organisational readiness
Early awareness & reducing risk Spot changes, risks and opportunities earlier
Reusable outputs & guidance Turn learning into practical guidance and tools
Practical answers Faster solutions to real problems
Time saved Learn from others' experience, avoid repeating mistakes
Safer adoption Better decisions with shared assurance and guardrails
Better quality outcomes Access to proven patterns and worked examples
Cost confidence Understand licensing, usage and FinOps more clearly
Stronger relationships Collaborate across departments, sectors and professions
Measurable progress Track maturity, impact and value together
Greater public value Better services, better use of public money, more trust
Shared principle: be open, be respectful, challenge constructively and focus on learning together to deliver safe, inclusive and effective AI in the public sector.
Operating model & programme
One AI capability operating model — not separate Copilot, agent, data, assurance and FinOps programmes. Cognitive load down, recognition up, DEIA by default rather than afterthought.
Every product, every build. DPIA in, not post.
Existing
Open, auditable, no vendor lock-in.
Existing
Inclusion, equity, access — in every spec.
Adding now
Named people, visible work, earned respect.
Adding now
Decision logs replace approval chains
AI drafts briefings; you review
Single source of truth per project
Open by default, restrict by exception
Weekly cross-team show-and-tells
Shared prompt and tool library
SME pairing across directorates
Secondments and shadow programmes
Monthly programme board updates
Outcome dashboards, not status reports
Senior sponsors own a use-case
Ministers briefed on citizen impact
Every product must actively serve those experiencing the greatest adversity: economic hardship, disabled and neurodivergent users, underrepresented communities, vulnerable households and carers, digital exclusion and low literacy, and everyone algorithmic unfairness would otherwise reach first.
Who is most harmed if this fails or excludes?
Does the design assume internet access or literacy?
Has it been tested with people from affected groups?
Are legal obligations (Equality Act, PSED) documented?
Is there a non-digital fallback route?
Does the data we collect reflect the full population?
Are there language and accessibility alternatives?
How will we measure equitable outcomes, not just usage?
A model trained on historical outcomes learns the discrimination in the record, not the potential of a well-supported person. Four pre-specified bias domains to audit — and abstention ("insufficient data — human decision only") is always acceptable, and must be possible.
Systematically fewer or lesser recommendations for protected groups with equivalent profiles.
Predictions accurate against a biased historical record, wrong about supported potential.
Too few cases for a group; the model fills gaps with majority priors.
Biased outputs produce worse outcomes that retrain the model — break the loop by design.
Training-data standards: exclude invalidated-treatment-era and sanctioned-employer records and proxy features; require minimum cases per group, representation audits, labelled synthetic augmentation for regularisation only, and temporal weighting toward modern practice. Lived-experience panels review real outputs; community bodies sign off the test criteria; explanations reference functional needs, never the diagnosis.
Phase 0 · FoundationWeeks 1–4
Private, data-safe assistant; no lock-in; GDPR by design
DPIA template, handling principles, red lines
Crowdsource the top jobs-to-be-done; impact × feasibility
Phase 1 · PilotMonths 2–4
Prompting fundamentals, hands-on sprints, peer cohorts
Open-weight models on gov cloud; avoid lock-in from day one
Hours saved, quality uplift, citizen outcomes — not vanity metrics
Phase 2 · ScaleMonths 5–9
Modular, containerised infrastructure across directorates
One champion per team; shared prompt library; show-and-tells
AI-assisted modernisation, documentation, automation
Phase 3 · EmbedMonths 10–18
Transparent registers, explainability standards, citizen comms
Awards, case-study library, cross-government showcases
Evergreen curriculum, AI-native onboarding, horizon scans
100%
team with baseline AI literacy
measurable use-cases live
0
proprietary lock-in
embedded in every build
civil servants celebrated
3 months
6 months
12 months
24 months
Sequence matters more than speed: permissions before scale, register before governance claims, cost instrumentation before volume. Recognition runs throughout — named people, visible work, credited contributions, minister-level visibility on citizen wins.
Responsible AI · people and adoption
Eight segments that turn responsible AI into lasting value, with accessibility, guest access and ways-of-working impact. Teams, roles, the academy and the adoption journey in this part go deeper on each.
Build capability
Build capability
Engage and empower
Engage and empower
Support and sustain
Support and sustain
Learn and improve
Learn and improve
Two questions, kept separate: is the AI tool itself accessible, and can AI support someone as a workplace adjustment? Request either through your line manager or the adjustments route. Every tool needs a non-AI alternative.
Do not rely on meeting recaps for guests. Store resources where external participants can reach them, set permissions deliberately and provide an accessible copy.
Maintain a controlled glossary of cross-government terms and acronyms so that transcription and summaries get names right.
Include workload, role, delegation, autonomy, training and workplace-adjustment impact in the adoption assessment, and monitor them after rollout.
Policy and governance Approved tools and data Security and privacy Accessible by design Sustainability by design Cost control and FinOps Supplier and model management Change management Cross-tenant and external working Legal, ethical and regulatory compliance
Worked example
Worked example
Portfolio reporting grounded in the Graph.
Worked example
The same day run on Claude Projects and scheduled tasks.
Worked example
Whole-portfolio reading in one context.
Training film
The Microsoft journey, end to end.
Training film
Both stacks, the same functions.
Comparison
39 capabilities with residency and FinOps layers.
Part 8 of 12 · 9 chapters
Live pilot handbook · prototype pilot pack, 25 August 2026
We are not reviewing a document. We are using the playbook during a live AI deployment to determine whether it improves decisions, delivery, assurance and public-service outcomes, and whether there is sufficient evidence and appetite to establish a permanent, interactive toolkit.
Evidence status. The pilot model, thresholds and templates are proposed practice. They must be agreed for the selected live deployment. No pilot outcome is claimed in advance.
Is the pilot worth initiating?
Mandate, sponsor, scope and evaluation plan. Proceed or stop.
Is the deployment suitable and safe for a pilot?
Candidate assessment and baseline. Select, condition or reject.
Are people, controls, evidence and fallbacks ready?
Readiness and assurance plan. Authorise controlled deployment.
Can the playbook be used during real delivery?
Operational evidence and decision trail. Continue, restrict, pause or stop.
Did governance and assurance improve decisions?
Independent findings and participant evidence. Accept, remediate or retest.
Was the playbook useful, proportionate and valuable?
Evaluation and options appraisal. Stop, retain, simplify or evolve.
Is there appetite and evidence for a permanent capability?
Investment recommendation and roadmap. Toolkit decision.
Decision points, risk boundaries, evidence and investment gate.
Eligibility, readiness criteria, expected contribution and safeguards.
Step-by-step lifecycle, artefacts, ceremonies and feedback routes.
Review lenses, evidence standards and independence rules.
Ways to observe, challenge and contribute without disrupting delivery.
Evidence needed to decide whether to stop, retain a static guide or build a living toolkit.
Live pilot · hypothesis and test dimensions
Primary hypothesis: using the playbook during a live deployment will improve the clarity, consistency, traceability and proportionality of governance and assurance without imposing unjustified delivery burden.
People can find and apply the right guidance without the author present.
Task observation, completion, help requests and interviews.
Controls reflect risk, autonomy, data, decision influence and public effect.
Exceptions, duplicated work, effort diary and reviewer judgement.
Material decisions have clear owners, evidence, uncertainty and conditions.
Decision records and assurance sampling.
The playbook prevents rework or reveals issues early enough to act.
Before and after comparison, avoided rework and issue chronology.
Assurance quality Claims are tested against meaningful evidence. Findings, closure evidence and independent opinion.
Learning value Participants gain usable capability and confidence. Applied assessments, job-aid use and follow-up interviews.
Teams and leaders prefer a maintained toolkit over no framework or a static document.
Investment survey, sponsor decision and committed use cases.
Proving that every AI deployment should use identical controls.
Treating participant satisfaction as proof of safety or value.
Using a successful pilot to claim cross-government impact.
Replacing formal legal, security, privacy, commercial or service-assessment obligations.
Live pilot · candidate selection
Seven criteria, the evidence each needs, and the condition that turns a candidate into a rejection or an escalation.
Named service, users, owner and intended live or limited-live use.
Reject or escalate when: There is only a hypothetical demo.
Defined tasks, prohibited uses and measurable outcomes. Reject or escalate when: Purpose is open-ended or unowned.
Controlled environment, data route, fallback and stop mechanism.
Reject or escalate when: Critical safeguards are unavailable.
The candidate exercises meaningful parts of the playbook.
Reject or escalate when: The use case is too trivial to generate evidence.
Product, operational, technical and governance participation.
Reject or escalate when: Delivery relies on one person or unavailable specialists.
Permission to examine decisions, tests, incidents and effort.
Reject or escalate when: Evaluation cannot access material evidence.
Team accepts challenge and can alter or pause deployment.
Reject or escalate when: Go-live is predetermined regardless of evidence.
Use three contrasting deployments where feasible: a bounded low-risk assistant, a reusable internal knowledge or workflow agent, and a higher-impact agent influencing operational judgement. This is a suggested design, not a mandated sample.
Live pilot · participants and responsibilities
The deployment SRO stays accountable for the live service throughout. The pilot adds evaluation and independent challenge; it does not take accountability away from delivery.
Protects the purpose, secures capacity and receives the investment recommendation.
Pilot SRO Accountable for pilot outcomes, scope and residual risk.
Deployment SRO Remains accountable for the live AI-enabled service.
Runs plan, ceremonies, dependencies, evidence and communications.
Applies the playbook to needs, outcomes and backlog decisions.
Owns operational performance, support, fallback and continuity.
Tests usability and captures staff and affected-user evidence.
Maps whole service, human controls, channels and failure demand.
Builds bounded capability and supplies evaluation and operational evidence.
Assures sources, provenance, metadata, retention and records.
Assess threats, access, personal data, incidents and mitigations.
Challenge fairness, human control, inclusion and public impact.
Delivers induction, job aids, competence assessment and adoption support.
Plans and conducts evidence-based challenge independent of delivery.
Owns baseline, methods, synthesis, limitations and options appraisal.
Provide targeted architecture, service, audit or policy challenge.
Tests appetite and prepares the permanent-toolkit option if warranted.
Live pilot · formal assurance model
Each stage states what it is for, what evidence it reviews, and the opinions it can legitimately reach.
Determine whether controlled deployment can begin.
Evidence: Scope, owners, risk, authority, data, tests, fallback, training.
Test whether controls operate and the playbook supports decisions.
Evidence: Logs, samples, incidents, exceptions, user evidence, decisions.
Assess material changes to model, tools, data, purpose or service.
Evidence: Impact assessment, regression tests, permissions and updated record.
Judge whether claims and conclusions are supported.
Evidence: Outcome evidence, costs, findings, limitations and participant testimony.
Assess whether permanence is justified.
Evidence: Need, repeat demand, value, maintainability, ownership and funding.
Delivery teams own controls and evidence. Independent assurance challenges claims and reports limitations. The SRO makes the accountable decision; assurance does not transfer accountability.
Live pilot · journey and rhythm
Ten steps from mobilise to share, each with its activities and the artefact it must produce.
Approve charter; appoint roles; agree scope, ethics, data and communications; define hypotheses and baselines.
Screen candidates; interview owners; classify risk and learning value; confirm evidence access.
Brief all roles; assess learning needs; issue playbook routes and job aids; rehearse escalation.
Capture current governance journey, effort, uncertainty, rework, readiness and expected outcomes.
Use playbook lifecycle, authority model, internal ATRS record, evidence pack and decision gates.
Begin bounded live or limited-live operation; monitor; sample outputs; record exceptions and effort.
Run architecture, service, security, privacy, responsible-AI, accessibility and audit reviews.
Compare baseline; triangulate metrics, observation and interviews; document limitations.
Options appraisal and sponsor decision: stop, static, simplify or toolkit.
Return findings to participants and communities; publish only what is approved and evidenced.
Pilot, deployment and assurance leads.
Test eligibility, constraints and evidence access. Selection record.
All core roles.
Align purpose, boundaries, responsibilities and stop routes.
Pilot manager, product, operations, evaluation.
Review decisions, friction, incidents, metrics and learning.
SRO, service owner, assurance specialists.
Test controls and conditions; decide remedial action. Assurance note and decision.
Users and delivery team.
Capture usability and burden close to the work. Short structured feedback.
Independent reviewers and community representatives.
Identify blind spots before conclusions harden. Challenge findings.
All accountable and assurance roles.
Agree evidence, limitations and recommendations. Final assurance opinion.
Sponsor, SRO, toolkit owner, finance and governance.
Choose future option. Investment decision.
Live pilot · communications plan
Each moment names its audience, message, channel and owner. The framework is being assessed — not the teams taking part.
Why this is a live deployment test; decision sought; safeguards.
Executive brief and sponsor conversation. Pilot sponsor.
The framework is being assessed, not the team; eligibility and contribution.
Teams post, briefing session and candidate pack. Pilot manager.
Transparent criteria, result, conditions and alternative engagement.
Decision note and direct conversation. Pilot SRO.
Process, roles, boundaries, evidence, learning and escalation.
Workshop, handbook and role cards. Learning lead.
Progress, decisions, risks, changes and support.
Weekly digest and monthly dashboard. Pilot manager.
Focused review question, evidence boundaries and response route.
Review pack and facilitated session. Assurance lead.
What evidence shows, limitations and what will change.
Playback and findings report. Evaluation lead.
Options, value, demand, costs, ownership and recommendation.
Decision paper and deck. Pilot SRO.
Verified lessons, reusable patterns and next decision.
Community session and controlled publication. Communications owner.
This is a controlled test of the framework in real delivery.
Participation is not an audit of team performance.
Safety, service obligations and accountable decisions override pilot learning.
Evidence includes what fails, adds burden or duplicates existing controls.
No decision to establish a permanent toolkit has been made in advance.
Live pilot · success framework and feedback
Each measure names its evidence, owner and cadence. Thresholds are agreed after the baseline is captured: a pre-set percentage without context would create false precision.
Participants navigate lifecycle and required artefacts without author dependency.
Observed task completion; help requests; navigation errors. Evaluation lead. Baseline and end.
Teams and reviewers agree controls match material risk and consequence.
Effort diary; duplicated controls; exceptions; reviewer judgement.
Weekly and close-out.
Material decisions identify owner, evidence, uncertainty, conditions and review.
Decision-record sample. Pilot manager. Fortnightly.
Material gaps are discovered while action remains possible.
Issue chronology and avoided or reduced rework evidence. Service owner. Monthly.
Findings distinguish intent, design, implementation, operation and outcome.
Independent assurance quality review. Assurance lead. Mid-point and close.
Support, monitoring, fallback and incident routes work in exercises and operation.
Readiness checks, exercises and operational records. Operations lead. Pre-live and live.
Participants demonstrate relevant competence in realistic tasks.
Applied assessment and coaching records. Learning lead. Before and after.
Observed service outcomes improve or credible learning justifies continuation.
Service-specific baseline and evaluation. Product manager. Agreed cadence.
Implementation effort and recurring cost are visible and acceptable relative to value.
Time, support and consumption data. Evaluation and finance. Monthly.
Leaders and teams prefer a maintained option and commit use cases, ownership or resources.
Structured option choice and commitments. Toolkit product owner. Close-out.
Claims separate verified results, internal evidence, assumptions and expected outcomes.
Evidence audit. Independent assurance. Close-out.
Research and controls address accessibility and affected-user differences.
Accessibility and inclusion evidence. Accessibility lead. Design and close-out.
Observed use Whether participants can apply the guide in context. At key lifecycle tasks. Observe the framework, not individual performance.
Decision sampling Whether governance claims are traceable. Throughout deployment. Risk-based sample agreed in advance.
Effort diary Where burden, duplication and rework occur. Weekly. Record estimates honestly and label them.
Structured interviews What helped, hindered or remained missing. Baseline, mid-point, close. Separate experience from factual claims.
Tabletop scenarios Whether controls work under failure and uncertainty. Pre-live and mid-point. Use synthetic scenarios where appropriate.
Where the model could fail, be gamed or become bureaucratic.
Mid-point.
Facilitated, blameless challenge.
Independent evidence review Whether conclusions follow from evidence. Close-out. Declare conflicts and evidence limitations.
Option-preference survey Whether appetite exists for future forms. Close-out. Require rationale and commitment, not preference alone.
Community consultation Whether patterns transfer beyond the pilot. After initial evidence. Do not generalise from advocacy alone.
KEEP
Works and should remain.
IMPROVE
Useful but unclear or incomplete.
REMOVE
Adds unjustified burden or duplicates controls.
MISSING
A necessary need or control is absent.
CONFLICT
Contradicts policy, service or assurance practice.
LOCALISE
Valid but needs context-specific adaptation.
Live pilot · decision and future state
Four options, the conditions that select each, and what each commits the organisation to. No option is pre-selected.
Value is weak, burden is excessive, risks are unresolved or demand is absent.
Archive evidence and retain lessons; do not operationalise.
The guide is useful as reference but repeat operational demand is limited.
Assign light-touch ownership and scheduled review.
Evidence shows repeat need, practical value, appetite and sustainable ownership.
Fund discovery and product development for an ongoing service.
Multiple committed user groups or deployments, not general interest alone.
A named product or service owner and multidisciplinary maintenance capacity.
A sustainable source of policy, standards and pattern updates.
Clear content governance, versioning, supersession and retirement rules.
Evidence that dynamic interaction adds value beyond a downloadable guide.
Security, privacy, accessibility, information-management and operating-cost assessment.
A measurable service outcome and roadmap, not only a feature list.
Understand the end-to-end route.
Adaptive journey through lifecycle, roles, gates and evidence.
Evidence: Navigation, completion and repeat-use data.
Know which controls and forums apply.
Context-led route based on purpose, risk, autonomy and public effect.
Evidence: Routing errors, exceptions and decision consistency.
Create and maintain a transparency record.
Guided record, evidence links, review and publication decision.
Evidence: Field usability, maintenance burden and scope decisions.
Build and review an evidence-based assurance case. Claims, evidence, findings, conditions and closure. Evidence: Evidence quality and review efficiency.
Build competence by responsibility. Role pathways, scenarios, assessment and job aids. Evidence: Competence outcomes and support demand.
Reuse proven controls and artefacts. Versioned patterns with evidence, owner and applicability. Evidence: Verified reuse and reduced duplication.
Oversee value, risk, cost and lifecycle. Inventory, status, evidence freshness and decision queues. Evidence: Decision usefulness and data quality.
Keep the toolkit relevant.
Structured proposals, challenge, supersession and release notes.
Evidence: Contribution quality and maintenance capacity.
Product principle: the implementation guide remains the flagship resource. Interactive functions should reduce cognitive load and help people act safely; they must not hide accountability, evidence or applicable authoritative guidance.
Critical security, privacy, legal or accessibility condition breached.
Required human oversight or operational fallback unavailable.
Evidence integrity compromised in a way that prevents accountable decision-making.
Deployment moves outside approved purpose or environment.
Human control, lifecycle management, governance, skills, quality assurance and risk.
User needs, multidisciplinary delivery, security, measures, reliability and improvement.
Criteria for government technology, including user needs, accessibility, security, privacy, data and Service Standard alignment.
ATRS guidance and scope policy Transparency purpose, scope tests and publication route. Official process; internal registry extension is proposed.
Responsible deployment, assurance, governance and practical resources.
Policy, lifecycle, roles, internal registry, learning and roadmap prototype.
Final discipline. Pilot conclusions must distinguish observed and verified results, internally evidenced results, participant views, assumptions, expected outcomes and unresolved unknowns.
Part 9 of 12 · 8 chapters
Delivery documentation · PRD
What this guide is for, who it serves, what it must do and what it deliberately does not do. Derived from published guidance and community discussion; a proposal, not an adopted product mandate.
Each problem is visible in the source material rather than assumed.
The AI Playbook, Service Standard, TCoP, ATRS, NCSC secure-AI guidelines, ethics frameworks and FinOps practice exist separately. Teams reassemble them from scratch, inconsistently, each time.
Personal assistants become shared agents and then business-critical dependencies without registration, named ownership, assurance or funded support.
Token and licence spend is visible; data remediation, workflow redesign, assurance, human review, monitoring, support and exit are not costed or owned.
The same product yields different results under different evaluation methods, so headline productivity figures travel further than the method that produced them.
Sessions, transcripts, repositories and personal networks hold the learning; there is no shared discovery layer, taxonomy or evidence pack, so lessons are relearned.
A reader can go from no policy position to a registered, assured, monitored agent using only this guide.
Every cost class in the investment line is named, owned and modelled.
Criticality drives the gate; low-risk work stays fast.
Claims carry method, publisher and date; illustrative material is labelled.
Register schema, ATRS record, instruction block and evidence card can be copied out and used.
Leaders, engineers, assurance, commercial, finance and frontline staff each have a path.
Guidance synthesis and chapter navigation
Assessment tools (readiness, criticality, tiers, maturity, quality gate)
| Accessibility WCAG 2.2 AA: contrast, focus visibility, keyboard operation, reduced-motion, semantic headings and labels. | Information handling OFFICIAL working content only. No personal data, no case data, no credentials. Assessment answers stay on the reader's device. | Portability Runs as a single file in a browser with no build step, no backend and no network dependency beyond fonts and cited links. |
| Performance First paint before the whole document streams; interactions respond within one frame budget on a standard civil-service laptop. | Provenance Crown-copyright content reused under the Open Government Licence v3.0, with publisher and date on every figure. | Print and export Chapters print legibly; the register schema and instruction block copy to the clipboard as plain text. |
Now
This guide, reviewed by the community, with illustrative figures clearly labelled.
Next
Illustrative figures replaced with measured departmental data; international benchmarking executed.
Later
Referenced in departmental AI policy, with the register schema and assurance gates in operational use.
Label every illustrative frame; keep publisher and date on evidence; publication checkpoint before circulation.
Proposal status on the landing chapter and in the hierarchy; explicit statement that it does not override guidance.
Chapter-to-source map; review date per chapter; ticker tracks guidance in force.
Device-only storage, no scoring authority claimed, evidence-based maturity rather than self-scored levels.
Delivery documentation · FTIR specification
Numbered requirements with priority, traceable to the chapters that satisfy them. Must, should and could follow the usual prioritisation convention.
Navigate 69 chapters in eleven numbered parts with contents, previous, next and sibling links
Must
Search every chapter, principle, gate, register field, role and source from one field
Must
FR-03 Record which chapters have been opened and show progress Should
Rate readiness across eight dimensions and return a banded barometer
Must
FR-05 Determine an assurance tier from a two-question intake Must
Score agent criticality across ten dimensions and derive the gate
Must
FR-07 Score barriers by impact and likelihood and band the result Should
Model cost per run and per month across models, languages and volumes
Must
Model unit economics: value per interaction, net value and return
Must
FR-10 Build a seven-part instruction block and copy it out Must
Assess evidence strength per capability domain across seven maturity levels
Should
FR-12 Step a failure and recovery loop, marking stages complete Should
Present the eighteen-field register schema and the eight-section ATRS record
Must
Show an illustrative portfolio dashboard with lifecycle, assurance, risk and domain views
Could
Single-file browser document; no build step, bundler or server runtime
Must
All state in memory; only the readiness barometer persists, in localStorage on the reader's device
Must
TR-03 Inline styling so content paints as it streams Must
Keyboard access for every control; visible focus; reduced-motion honoured
Must
Contrast at or above 4.5:1 for text and 3:1 for meaningful non-text
Must
TR-06 No third-party analytics, cookies or tracking Must
External links open in a new tab with noopener and carry publisher and date
Should
Deterministic calculations — same inputs, same outputs, no model calls
Must
Register schema aligns field-for-field with the ATRS record sections
Must
Register, information asset register and data catalogue treated as three integrated registers, not duplicates
Must
Criticality score maps to the five assurance gates and their evidence requirements
Must
Cost model outputs feed the investment line and FinOps allocation layers
Should
IR-05 Instruction block output is platform-neutral plain text Must
IR-06 Chapters cite the guidance in force and carry a review date Must
Links out to the AI Engineering Lab, AI Knowledge Hub and departmental policy rather than copying them
Should
Sits alongside the govukOS twin front door and portals as one app in a shared shell
Should
Delivery documentation · SAD
Principles, context, layers and the decisions that shaped them. The architecture is deliberately small: the guide holds interpretation, never the record of a live deployment.
Link to the governed source; hold interpretation and sequence here.
The guide never becomes the register of record for a live deployment.
No backend means no new data controller, no new attack surface, no new service to run.
Contrast, focus and keyboard behaviour are acceptance criteria, not polish.
Every figure states publisher, date and method, or is labelled illustrative.
Crown-copyright content under OGL v3.0; the guide itself is reusable and adaptable.
Reader Civil servant in one of the twelve priority roles Reads chapters, runs assessments, copies artefacts
Owning organisation Department, agency, council or health body Holds the register of record, policy and approvals
Central guidance GDS, DSIT, NCSC, CDDO successor functions Publishes the guidance this guide interprets
Cross-government Copilot forum, AI Engineering Lab, ethics and data communities
Supplies patterns, failures and supplier intelligence
Suppliers Platform, model and integration providers Subject of dependency, exit and commercial chapters
govukOS shell Twin front door and sibling portals Hosts the guide as one app among several
Masthead, ticker, drawer, chapter rail, contents, chapter body, footer navigation
Chapter registry Parts, chapters, numbering, accents, icons, read state
Source-derived data: principles, gates, domains, roles, tiers, schemas, dashboards
Barometer, criticality, tiers, maturity evidence, quality gate, barrier scoring
Token maths, language multipliers, unit economics, allocation layers
In-memory interaction state plus one persisted barometer record on the device
ADR-01
No hosting, no data processing agreement, no service to operate; trades server-side features for portability.
ADR-02
Removes the need for a DPIA on the guide itself; assessments stay on the device.
ADR-03
Keeps navigation, search and cross-references coherent; avoids a fragmented document set.
ADR-04
Protects the evidence boundary this playbook sets out.
ADR-05
Gives readers a positional cue; all accents are contrast-tested as text.
Delivery documentation · HLD
Components, the entities they work on, and the four flows that carry most of the value.
Masthead, guidance ticker, launchpad, search field
Drawer with parts and chapters, chapter rail, contents overlay, prev/next
One section per chapter, streamed in order
Barometer, criticality radar, tier finder, maturity evidence, quality gate
Token cost, language multiplier, unit economics, allocation layers
Instruction block, register schema, ATRS record, evidence card
Flat index of chapters, principles, gates, fields, roles, sources
Illustrative portfolio control tower
Chapter id, part, number, title, icon, accent, read Navigation and progress
Blueprint title, source, ingest date, chapters it feeds Provenance and coverage
Assessment dimension, score, band, timestamp Barometer and maturity evidence
Register record 18 fields from identity to retirement The record an organisation holds
ATRS record 8 sections, each with its own fields Transparency record per agent
Cost scenario tokens in/out, prices, tools, infra, review effort, volume Cost and unit economics
Instruction block goal, context, sources, task, format, guardrails, output Skill injection
Read a chapter Drawer or contents → chapter renders → read state recorded → next chapter offered
Assess readiness Rate eight dimensions → band computed → stored on device → shown in the drawer workspace
Tier an agent Two intake questions → tier → gate, controls, evidence and review frequency shown
Price a job Enter words in and out, model, language, volume → per-run and monthly cost → unit economics
Delivery documentation · LLD
Module responsibilities, the arithmetic behind every score, the validation rules and the accessibility implementation. Enough for someone else to rebuild the behaviour exactly.
groupDefs → flat chapter list → numbering, accents, prev/next, contents
section, groups, read, toc
Barometer module eight dimension scores → mean → band with colour and label scores (persisted)
Criticality module ten dimension scores → total → tier and gate mapping crit, gateTier
Barrier module impact × likelihood → 1–25 → four bands bar
tokens = words ÷ 0.75; cost = in×priceIn + out×priceOut + tools + infra + review
lab, ue
Instruction module seven fields → composed block → clipboard blk, copied
mean of rated dimensions; 0–0.9 Nascent, 1.0–1.9 Emerging, 2.0–2.4 Establishing, 2.5–3.0 Leading; unrated dimensions excluded from the mean and reported as unrated.
sum of ten dimensions scored 1–5; 10–20 personal, 21–30 team, 31–38 departmental, 39–44 operational service, 45–50 citizen impact.
impact × likelihood on 1–5 scales; 1–3 low, 4–8 medium, 9–15 high, 16–25 critical.
value per interaction = (minutes saved ÷ 60 × hourly rate) × quality factor; net = value − cost; return = net ÷ cost expressed as a percentage.
Numeric inputs Clamp to declared minimum and maximum; reject non-numeric silently rather than throwing.
Unrated dimensions Show an em dash, never a zero, and exclude from means.
Currency Two decimal places, pound sign, thousands separated; never rounded away in totals.
Clipboard Plain text only; confirm with a transient state, no alert.
Persistence One key, versioned; parse defensively and fall back to blank on any error.
Colour Every text and meaningful non-text pair contrast-tested; accents have separate text-safe variants where a fill carries text.
Keyboard All controls are buttons or inputs; aria-pressed and aria-current reflect state; focus ring never suppressed.
Motion The guidance ticker pauses on hover and respects prefers-reduced-motion.
Structure One h1 per chapter, sequential headings, aria-labels on sections, skip link to content.
Language British English, plain terms, expanded acronyms on first use in a chapter.
Delivery documentation · theory of change
From inputs to impact, with the assumptions each step rests on. Written so the chain can be falsified: if an assumption fails, the link it supports fails with it.
Central guidance in force Community evidence and patterns Evaluation studies Departmental policy and delegated authority Funded time from named roles
Synthesise guidance into sequenced chapters Build assessment and costing tools Publish register and ATRS schemas Run role-based capability pathways Return findings to the community
One navigable implementation guide Completed readiness and criticality assessments Registered agents with named owners Costed business cases including hidden costs Reusable evidence cards and patterns
Teams find the applicable guidance quickly Proportionate assurance replaces blanket caution Agents have owners, tiers and review dates Cost conversations include people and exit
Fewer orphaned and unregistered agents Duplicated procurement and assurance effort falls Evidence is reused across organisations Oversight is continuous, not one-off
Guidance stays reasonably stable Chapters carry review dates so drift is visible rather than silent.
Teams have protected time Without funded time, the guide is read but not applied.
Leaders accept proportionality Blanket caution defeats the tiering model as surely as no control at all.
Community keeps sharing Reuse depends on organisations publishing failures as well as successes.
Illustrative figures get replaced Measured departmental data must supersede the labelled placeholders.
Chapters opened per role pathway
Agents registered with a named owner and review date
Proportion of agents with an ATRS record
Assurance decisions made at the tier they belong to
Business cases including the full investment line
Patterns and evidence cards reused across organisations
Delivery documentation · user research bundle
Drawn from the roles, frictions and lifecycle material in published guidance and community discussion. These are evidence-derived working personas, not research participants — they need validating with real users before they carry design decisions.
Decide where AI is worth the risk and the money, and be able to answer for it.
Needs
Frustrations
Adopt agents without destabilising a service people depend on.
Needs
Frustrations
Build well-grounded agents and prove they work.
Needs
Frustrations
Apply proportionate control and see the evidence behind it.
Needs
Frustrations
Fund adoption honestly and keep exit possible.
Needs
Frustrations
Do the work faster without carrying unowned risk.
Needs
Frustrations
As a Accountable leader, I need to see the method behind any productivity claim, so that I do not repeat an unevidenced figure in public.
Accepted when: Every figure shows publisher, date and method, or is labelled illustrative
As a Accountable leader, I need to know which agents are high risk and who owns them, so that I can direct attention where consequence is highest.
Accepted when: Control tower lists risk level, owner and next review per agent
As a Service owner, I need to know the gate an agent must pass before it touches my service, so that I can plan assurance rather than discover it.
Accepted when: Criticality score returns a tier with named evidence and review frequency
As a Service owner, I need to have an exit path for a platform or model, so that a supplier change does not stop my service.
Accepted when: Seven-step exit path with the dependencies it addresses
As a AI engineer, I need to decide where grounding information should live, so that I do not create ungoverned copies.
Accepted when: Decision tree returns one of four options in priority order
As a AI engineer, I need to compose an instruction block once and reuse it, so that behaviour is consistent across platforms.
Accepted when: Seven-part block copies out as plain text
As a Assurance specialist, I need to score criticality across ten dimensions, so that control is proportionate to consequence.
Accepted when: Score maps to one of five gates with its controls
As a Assurance specialist, I need to record transparency information for every internal agent, so that oversight and audit have something to inspect.
Accepted when: Eight-section ATRS record with all fields listed
As a FinOps lead, I need to model cost per run and per month, so that a business case includes consumption, not guesses.
Accepted when: Cost lab returns per-run and monthly figures with a language multiplier
As a FinOps lead, I need to see value against cost per interaction, so that I can show whether the agent earns its keep.
Accepted when: Unit economics returns value, net and return
As a Frontline practitioner, I need to know what I may use and when to escalate, so that I can work confidently within the guardrails.
Accepted when: Support chapter states allowed use, escalation and where help comes from
As a Community manager, I need to turn a session into a reusable evidence pack, so that learning survives the meeting.
Accepted when: Evidence pack structure with owner, taxonomy and review date
Arrives with a question Lands on the overview, sees proposal status “Is this authoritative?” Proposal flag and hierarchy answer it in one screen
Finds the applicable chapter Uses search or contents “Where does my case sit?” Chapter numbering and part accents orient quickly
Runs an assessment Tier finder or criticality radar “Am I over- or under-controlling?” Tier arrives with evidence and review expectations
Takes an artefact away Copies the register schema or instruction block “What do I do on Monday?” Copyable artefacts, not prose
Returns for oversight Re-opens the control tower and maturity evidence “Has anything drifted?” Read state and review cadence bring them back
Idea Someone builds a helpful agent “This saves me an hour a day” Register at creation, not at scale
Share Colleagues start relying on it “It works, so why formalise it?” Promotion gate between personal and team scope
Assure Criticality scored, tier applied “How much control is proportionate?” Gate names the evidence needed
Operate Monitoring, cost and quality watched “Who notices if it degrades?” Named owner and continuous duties
Review Re-authorise, improve, restrict or retire “Is it still earning its authority?” Review date forces the decision
Retire Access revoked, records archived “What happens to the evidence?” Archive and audit stage in the lifecycle
Session happens Weekly call, recorded from 17 August 2026 “Will I find this again?” Automatic recording closes part of the capture gap
Capture Transcript, chat, files, commitments “Who owns the follow-up?” Session evidence pack with owner and actions
Classify Taxonomy: capability, use case, sensitivity, owner, review date “Which of these applies to me?” Shared taxonomy across communities
Connect Federated index links to governed sources “Is this the current version?” Index links rather than copies
Reuse Another organisation adopts the pattern “Has anyone tried this already?” Discovery by role, sector, maturity and supplier
Learn Outcome and corrections returned “Did it actually work?” Value reported as reuse, not attendance
Five lanes across the same five moments: what the user does, what they see, what happens behind it, what supports it, and what is recorded.
Requests help Reviews agent output Accepts or challenges Escalates an error Feeds back on value
Agent responds with citation Confidence and limits shown Human review point Escalation route offered Feedback captured in service
Retrieval over authoritative sources Guardrails and permission checks Logging and audit trail Incident triage and rollback Evaluation against baseline
Register and ATRS record maintained Assurance gate evidence held Cost allocated to the use case Supplier and model change managed Review or retirement decision
Source provenance Approval and sign-off Monitoring and drift reports Incident records Benefit evidence
Delivery documentation · external resources
This guide interprets and sequences; these resources supply engineering practice, worked examples and published evaluation. Link to them rather than copying them — they move faster than any document.
Programme site, an open GitHub resource repository under OGL v3.0, a Knowledge Hub community, role-specific material and cross-government hackathons. Enquiries to the GDS engineering-excellence team.
A living resource on building and adopting AI in the public sector, with tool pages carrying what was built, what it cost and what the evaluation found.
Caddy, Consult, Extract, Redbox, Minute, Parlex, Scout, Connect, Lex and rAPId — several open-sourced, with published evaluations.
Proposed cross-economy regulatory sandbox with time-limited exemptions under supervision; call for evidence closed January 2026.
Up to £40m for long-term research into limitations such as hallucination, unreliable memory and unpredictable reasoning.
Ten principles and the life cycle this guide sequences; supersedes the Generative AI Framework.
Need an engineering or evaluation pattern AI Engineering Lab repository and community 3.2 Architecture & quality
Need a worked example with published evidence AI Knowledge Hub tool pages 2.2 Use cases & evidence
Need to see how a product was evaluated i.AI evaluations and blogs 2.1 Evidence before assertion
Need to know if regulation is the blocker AI Growth Lab call for evidence Resource library (home page)
Need the authoritative principles AI Playbook for the UK Government 1.3 The playbook
Need transparency-record structure ATRS and the internal registry chapter 4.5 Internal ATRS registry
Part 11 of 12 · 4 chapters
Deploy for real · where the work actually is
The highest-volume services across UK central government, local government and the NHS, placed on two axes that decide almost everything else: how much of it there is, and what happens when it is wrong. Select a quadrant or a factor to filter. Open any service to load it into the generator.
54 services 4 quadrants 10 factors 27 land in Tier 4 34 sit in the dangerous quadrant
One agent, end to end, in each quadrant
Filter by factor
54 of 54 shown
Volume bands come from published transaction data; criticality and the factor tags are an assessment of consequence, not a classification you can cite. The value is not the placement — it is that placing a service forces a conversation about consequence that a business case usually skips. Move anything you disagree with, and record why you moved it.
Deploy for real · working tool
Two bands and ten factors in. Tier, autonomy cap, oversight regime, currency window, evaluation set, threat priorities, DPIA route, contract clauses, workforce actions, register fields and kill criteria out — with the rules that produced them, so the answer can be argued with rather than accepted.
No factors selected
Q1
The strongest agent case in government. Volume pays for the governance; a mistake is recoverable. The risk here is not harm, it is quietly losing the ability to do the work.
Effective criticality
Autonomy cap
Act with post-hoc sampling — exceptions escalate
Challenge-rate floor
Below this, suspect oversight, not excellence.
Sign-off
2% random sample, weekly
A screening that concludes no full DPIA is needed is still evidence. Keep it, and re-run it when scope changes.
A frozen set of real historical cases with agreed correct outputs, re-run on every model or prompt change.
Cases the agent must decline, escalate or flag rather than attempt.
Per-transaction cost and response time, with the ceiling that triggers review.
Prompt injection via ingested content High
Any document, email or record the agent reads is untrusted input. Test with adversarial content in every ingestion path.
Tool over-permission High
The agent holds broader credentials than its task needs. Scope per action, not per agent.
Written notice before any model, weight or default change that affects behaviour, with a defined notice period.
All prompts, outputs, tool calls and decisions exportable in a documented open format, during and after the contract.
Explicit prohibition, with the right to evidence of compliance.
Prompts, instruction blocks, eval sets and grounding indexes are ours, exportable, and documented well enough to rebuild elsewhere.
Volume automation releases hours. Publish where they were redeployed. Unnamed savings become quiet headcount reduction with no decision behind it.
Staff and unions see the job-redesign proposal while it can still change. After the build is a briefing, not a consultation.
A named person, not a team or a role inbox.
How an effect is undone, tested, with the last test date.
Who currently holds valid approval authority, and when each window expires.
Reviewers overturning less than 0.5% of outputs. Either the agent is perfect or oversight has stopped. Assume the second.
Any drop on the frozen set that cannot be explained and fixed within the rollback SLA.
A single decision the accountable officer cannot account for in plain language.
Very high volume with low criticality places this in Q1 · Automate with sampling.
The weights are a defensible starting point, not a measured value. Change them with your assurance function and record the change — a tool whose reasoning is hidden cannot be challenged, and a governance answer nobody can challenge is not governance.
Deploy for real · four agents, one per quadrant
Every artefact filled in, including the ones that are uncomfortable. Very high volume, moderate criticality — the case most departments would call easy. Named honestly it lands in Tier 3 · Enhanced, and that is the most useful thing in this chapter.
These four agents are composites, not case studies. The figures are internally consistent and illustrative, and the decisions are the ones the sources support. Your organisation will ratify and adapt every one of them.
Signed by the service owner before any build. One page, and it fits.
Inbound correspondence arrives across six channels into one queue. Median time to first correct routing is four days; 11% is routed wrong at least once. Statutory deadlines are met by escalation, not by process.
Classify inbound correspondence to one of 34 work queues, extract sender, reference, subject and any stated deadline, and draft a suggested reply from the approved template set.
Deciding any case. Sending anything. Replying to anyone. Handling correspondence flagged as a complaint, a legal threat, a safeguarding concern or a data-subject request — those route to a human unopened by the agent.
Time to correct routing, misroute rate, deadline breaches, and reviewer challenge rate. Not volume processed, and not staff satisfaction alone.
The agent proposes. A caseworker commits. Nothing leaves the organisation without a named person attached to it.
Twelve months, one delivery team, with a stop point at month four and a re-authorisation decision at month twelve.
Any statutory deadline missed because of the agent, or a challenge rate below 5% sustained for two months.
Same agent, same quadrant, one factor different. Each variant below is scored by the generator on this page, not asserted — so what moves and what does not is the tool's own arithmetic.
The agent sends the reply itself Tier moves
+ Irreversible Tier 4 · Critical Q1 · Automate with sampling L1 · Draft only — a person edits and commits every output
One word in the mandate — send rather than draft — and an effect exists that no later correction undoes. Autonomy is forced to draft-only regardless of tier, and shadow mode becomes mandatory. This is the most common scope change requested after go-live, and the cheapest one to refuse.
Correspondence routinely contains health detail Tier moves
+ Special-category data Tier 4 · Critical Q1 · Automate with sampling L1 · Draft only — a person edits and commits every output
The screening guardrail stops being a nicety and becomes the control the DPIA rests on. A leakage probe joins the evaluation set, and the grounding index has to be provably free of the material.
Internal-only queue, rule-based routing Tier moves
Remove the discretion and the vulnerable recipient and the same agent, at the same volume, drops a tier. Most of the governance in the worked example is paid for by those two factors, not by the volume.
Deploy for real · scope boundary
A guide that does not state its boundary strands the team at the point where the interesting problems start. This chapter exists so nobody reaches gate 3 believing the architecture, the security testing or the legal advice was somewhere in the earlier parts.
Principles, criticality and tiering, autonomy boundaries, registers and transparency records, oversight design including currency, cost and value modelling, roles and adoption, the pilot method.
Instruction blocks, evaluation plans, threat profiles, DPIA routes, contract clause sets, job redesign. The shape is here; the substance is yours.
Reference architecture for your estate, engineering pipeline, executed security testing, model selection, sector regulation, organisational design, safety cases.
Eight things this guide names but cannot supply. Each one has a normal home in an organisation. Find the home before the build starts, not after.
A target architecture for your estate Enterprise architecture
This guide names what an architecture must satisfy. It does not know your identity provider, your network, your data platform or your existing contracts. Someone with authority over those has to draw it.
Executed security testing Security function or an appointed third party
The threat model here tells you what to test. Running the test — penetration testing, red-teaming, adversarial evaluation — is specialist work with a report at the end of it.
Legal advice on your specific case Legal and your DPO
Everything here about lawful basis, automated decision-making and liability is structure for a conversation with a lawyer. It is not the advice.
Model selection and benchmarking Technical owner
Which model, at what cost, with what evaluation on your data. This guide is deliberately model-agnostic, which means it cannot answer the question you will be asked first.
Sector regulation Your regulator and legal
Health, policing, immigration, education and financial services each carry duties this guide does not enumerate. Where it conflicts with your regulator, your regulator wins.
Organisational design Accountable leadership
Who reports to whom, which team holds the capability, and how it is funded past the pilot. The roles here are role names; you have people, budgets and politics.
Accessibility conformance Accessibility specialist
The comprehension and inclusion tests here are necessary and not sufficient. A conformance audit against WCAG is a separate, evidenced exercise.
Safety cases Clinical or safety governance
Where an agent touches physical safety or clinical outcomes, a structured safety case is required and this guide does not contain one.
Read these before citing anything in this guide to a board, an auditor or a regulator.
Nothing in this guide has statutory force. Until your organisation ratifies it, an artefact is a well-argued suggestion. Ratify it, or expect it to be ignored at the first disagreement.
Where this guide shows a threshold, a percentage or a band, it is a defensible starting point drawn from the sources, not a measured value from your service. Replace it with yours.
The fifty-four services in the atlas are placed on published volume bands and an assessment of consequence. They are a starting position for an argument, not a classification you can cite.
It derives a consistent governance set from the inputs you give it. Give it flattering inputs and it will produce a flattering answer. That is a property of every assessment tool, including the ones with logos on.
They are built from the sources, from published service data and from patterns that recur across deployments. They are not case studies of four real agents, and no organisation in them is identifiable. Their numbers are internally consistent and illustrative — treat them as a model answer to compare your own against, not as evidence of what happened somewhere.
The life cycle here is ordered. It carries no durations, because the honest answer depends on your assurance capacity, and that is the constraint nobody plans for.
Part 12 of 12 · 3 chapters
Pan-UK
Counting strategies overstates maturity: a benchmark that measures published strategy ranks all four nations comparable. Only one has a published body of controlled evaluation of deployment. What each nation can borrow from the others is the point of comparing.
England and reserved functions
Strategy
Who leads
DSIT · Government Digital Service (incl. i.AI and the AI Engineering Lab) · AI Security Institute
Signals
Four published Copilot evaluations plus NAO and PAC scrutiny — the UK's distinctive asset is evaluation, not strategy
ATRS transparency records mandatory for departments and ALBs
Cabinet Secretary's AI Innovation Award (quarterly, from Jun 2026); AI & Data Challenge at 252 ideas
Devolved strategy, refreshed 2026
Strategy
Who leads
Scottish Government · Scottish AI Alliance · The Data Lab
Signals
National AI plan, 2025
Strategy
Who leads
Welsh Government · Centre for Digital Public Services (CDPS)
Signals
Council adoption early-stage (AI Readiness of UK Local Authorities, 2025)
Draft strategy, consultation closed 7 October 2026
Strategy
Who leads
The Executive Office · Department of Finance · Office of AI and Digital
Signals
Proposes human-oversight teams, redress mechanisms and AI risk-management standards
Procurement expectations on suppliers: oversight, fairness, transparency, accountability
Stated aim: close the adoption gap with Scotland and Wales
Local-government comparison: AI Readiness of UK Local Authorities 2025 (Heriot-Watt University & GoLLM), 208 councils across all four nations — Scotland leads, with roughly one council in three piloting or using AI.
International
Five reference jurisdictions, five different bets: from the EU's binding risk tiers to Singapore's testing-led pragmatism. Genuinely transferable: Canada's public assessment and inventory approach, Australia's named AI ownership and workforce model, Singapore's central adoption flywheel, US inventory discipline, and the EU's joint procurement direction.
Where Key instrument Model For public bodies
AI Playbook + sector regulators; ATRS; AI Security Institute evaluations
Principles-led, pro-innovation; non-statutory for now
Playbook principles and ATRS records; four published evaluations; frontier legislation under consideration
AI Act (in force 1 Aug 2024; Digital Omnibus 2026 defers high-risk duties to Dec 2027, public authorities to Aug 2030); EuropAI joint procurement
Binding, risk-based regulation
High-risk public deployments: conformity assessment, logging, human oversight, fundamental-rights impact assessment
NIST AI Risk Management Framework; OMB memoranda; GAO oversight
Voluntary framework + agency rules; state laws vary
Federal AI use-case inventories and Chief AI Officers; rapid growth with persistent policy and budget challenges
Directive on Automated Decision-Making + Algorithmic Impact Assessment; agentic-AI guidance (May 2026)
Binding on federal automated decisions
Mandatory AIA before deployment; names skill atrophy as an oversight risk; bounded autonomy and recoverability
Model AI Governance Framework · AI Verify toolkit
Voluntary, testing-led
Central funding, shared platforms, problem intake, applied learning — an adoption flywheel
In force 1 August 2024. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force 27 July 2026) moved stand-alone high-risk obligations to 2 December 2027, embedded high-risk systems to 2 August 2028 and high-risk systems for public authorities to 2 August 2030. Article 50 transparency duties have applied since 2 August 2026. Obligations follow the use case, not the technology. The benchmark other regimes are measured against.
Unacceptable
Banned outright: social scoring, subliminal manipulation, real-time remote biometric ID in public spaces (narrow exceptions).
High risk
Essential public services, law enforcement, employment, critical infrastructure: conformity assessment, documentation, human oversight, EU database registration.
Limited risk
Transparency duties: chatbots must say they are AI; synthetic media must be labelled.
Minimal risk
Most systems: spam filters, recommenders. No mandatory obligations; voluntary codes encouraged.
The shared definition of AI used by the UK Playbook and the EU AI Act. The OECD.AI Policy Observatory tracks national AI policies worldwide; it is the model for this app's observatory.
The first binding international AI treaty: human rights, democracy and the rule of law. The UK is a signatory.
International guiding principles and a code of conduct for organisations developing advanced AI systems.
Frontier-safety declarations and the network of national AI safety and security institutes. Started at Bletchley Park, November 2023.
First UN consensus resolution on safe, secure and trustworthy AI for sustainable development.
Reference · pace and parity
Skill atrophy appears in this guide as a single sentence in the evidence chapter, credited to Canada. Nothing sits behind it: no control, no metric, no register field, no gate criterion, no owner. That is the honest answer to the question, and it is now backlog item PACE-01. This chapter closes it, then applies the same test to everything else the world has built that we have not.
Named
One sentence, evidence chapter, attributed to Canada
Controlled
No control, no owner, no frequency
Measured
No metric in the barometer or control tower
Gated
No criterion in the quality gate
Six moves, and the last is the operable one. The instrument is the Guide on the Use of Agentic Artificial Intelligence, published by the Treasury Board of Canada Secretariat on 22 May 2026, sitting on top of a directive that already made assessment mandatory.
Continued offloading of decisions to AI agents can lead to knowledge and skill atrophy and weakened human judgment. Employees become less practiced at core tasks and slower to notice when something is wrong.
Guide on the Use of Agentic AI, TBS, 22 May 2026
Maintenance of relevant skill sets among agentic AI users and deployers is treated as a prerequisite for meaningful human oversight. Not a wellbeing concern — a validity condition on the control itself.
Same instrument, mitigations
Bounded autonomy and recoverability, introduced alongside the existing generative-AI principles rather than replacing them.
Same instrument
Service quality, privacy and security, legal defensibility and public trust. The risk is framed as legal and operational exposure, not a training preference.
Same instrument
Algorithmic Impact Assessment before deployment, four impact levels, external peer review at levels III and IV, questionnaire published openly including on GitHub.
Directive on Automated Decision-Making
Periodic human-only execution of agent-handled tasks for comparison; staff trained to interrogate agent outputs rather than accept them.
Commentary on the guide, May 2026
PACE-01 · proposed instrument
Canada names skill maintenance as a prerequisite for meaningful human oversight. A prerequisite that nobody measures is a sentiment. Aviation, nuclear operations and clinical practice all solved this the same way: recency requirements, evidenced, with authority attached. If you supervise the work, you stay current in the work. Six components, all of them auditable.
Component What it does Who holds it Evidence
A reviewer executes an agent-handled task unaided, then compares. Tier 1 monthly, tier 2 quarterly, tier 3 twice yearly.
Reviewer
Drill record holding both outputs
A reviewer whose last unaided execution predates the tier window loses approval authority on that queue until re-current. Enforced, not advised.
Service owner
Authority state in the register
Every drill records agreement or disagreement and why. Sustained perfect agreement is itself a signal — it is what rubber-stamping looks like from outside.
Reviewer
Divergence entries with reasons
Proportion of agent outputs materially amended before acceptance. A falling challenge rate against a flat error rate is drift, not improvement.
Control tower
Monthly trend per queue
No reviewer supervises the same agent-handled queue continuously. Rotation preserves unaided practice somewhere in the role.
Line manager
Rota showing unaided work
The register records which skills the agent displaces, who must retain them and to what standard. If nobody is named, the deployment is not approvable.
Deployment owner
Register field, tested at each gate
The uncomfortable component is the divergence log. A reviewer who agrees with the agent every time is either supervising a perfect agent or not supervising at all, and the record cannot tell the difference. Convergence has to be read as a warning rather than a result.
Three are outright gaps: something binding or published elsewhere with no UK counterpart. Four are partial. One is context we should track rather than copy.
EU AI Act Article 4 — the AI literacy duty Gap
Binding since 2 February 2025 and unaffected by the proposed high-risk delay. Applies to deployers of every AI system, not only high-risk ones, and covers staff and contractors regardless of seniority. The compliance test is a dated, role-mapped programme that plausibly matches actual use.
Where we stand
Training is encouraged throughout UK guidance but is nowhere a duty, and course completion is not evidence of competence.
Singapore Model AI Governance Framework for Agentic AI Gap
World-first agentic framework, Davos 22 January 2026; version 1.5 on 20 May 2026 after feedback from over sixty organisations, with more than ten deployment case studies. Four pillars, unique agent identities tied to a supervising agent or user, threat modelling for memory poisoning, tool misuse and privilege compromise.
Where we stand
We hold secure-AI guidance and our own patterns, but no single agentic instrument and no agent identity requirement.
Canada
Gap
Assessment is mandatory before deployment, scored across four impact levels, with external peer review at the top two. The questionnaire is published openly, including on GitHub, and has become the instrument others copy.
Where we stand
ATRS is transparency after the fact, not assessment before it, and it is not a gate.
Australia AI Safety Institute inaugural report, 9 August 2026 Open to all
Found that NIST's risk framework, OWASP's Agentic Top 10, Singapore's agentic framework and ASD's agentic guidance all rest on the same hidden assumption: that one organisation owns every agent in a system. Cross-organisational agent risk is a structural gap none of them covers.
Where we stand
Our register assumes single-owner agents too. The gap is identified and unimplemented everywhere, including here — which makes it available.
Global Singapore Consensus companion report on agentic risk Partial
Ten foundational principles for managing agentic risk: least privilege, traceable identity, auditability, validated deployment, adversarial resilience, multi-agent stability, runtime assurance, interruptibility, legibility and human oversight.
Where we stand
We hold most of these as scattered controls. None is consolidated into a named principle set an approver can test a deployment against.
OECD Digital Government Outlook 2026 Partial
Thirty-two of thirty-six countries run AI training programmes; only thirteen of thirty-six train on using AI in public services or in policymaking. Broad capability is improving faster than role-specific readiness. Thirty of thirty-six have an institution governing public-sector AI; twenty-one of thirty-six offer central procurement support.
Where we stand
Our academy is role-based on paper. Depth of oversight-specific training has never been tested against this benchmark.
G7 GAIN network and the GovAI Grand Challenge Partial
A G7 AI network convened and chaired by Canada, tasked with a roadmap to scale successful public-sector AI projects and a catalogue of open-source, shareable AI solutions for members.
Where we stand
We participate. We have contributed no reusable instrument to the catalogue.
Asia-Pacific Korea AI Basic Act (2026), Taiwan AI Basic Act (2025) Tracking
Regional convergence on human oversight, data safeguards and prohibitions on harmful or deceptive use, arriving through basic law rather than sectoral guidance.
Where we stand
Context rather than gap: the UK route is guidance-led by choice. Worth tracking for interoperability, not imitation.
Thirteen items and three deliberate refusals. Every item has an owner and a first deliverable, because an unowned backlog item is a wish.
This quarter. Each one closes a gap that leaves an existing control unsound.
Manual-mode drills, currency windows, divergence log, rotation and the displaced-skill register field. Turns the Canada finding into something an auditor can test.
Owner: Learning lead with service owners First: Control spec and tier windows
A dated, role-mapped competence requirement for anyone who operates or approves an agent, contractors in scope. The Article 4 test, applied voluntarily and ahead of any obligation.
Owner: Learning lead First: Role-to-competence map
Every agent carries a unique identity bound to a named human owner and, where orchestrated, to a supervising agent. Prerequisite for audit, for handover and for retirement.
Owner: Architecture First: Identity schema in the register
Adopt both as named approval principles. Every agent has a tested interrupt and a tested rollback before it goes live, with the test evidenced.
Owner: Assurance lead First: Two gate criteria and a test script
Challenge rate and divergence surfaced in the control tower beside assurance status, per queue, monthly.
Owner: Control tower First: Two new tiles
Next two quarters. These extend reach rather than repair a fault.
Register fields for agents we interact with but do not own: supplier, other department, arm's-length body. The first implementation of the gap Australia named.
Owner: Architecture with commercial First: Field set and one worked case
An assessment step for orchestrated patterns, covering interaction effects that no single-agent assessment can see.
Owner: Assurance lead First: Assessment addendum
Not general AI awareness. Training in interrogating outputs, recognising plausible-but-wrong, and running a manual-mode drill.
Owner: Academy First: Curriculum for three roles
A scored pre-deployment assessment distinct from the transparency record, with independent peer review at the top tiers.
Owner: Governance office First: Scoring instrument
Memory poisoning, tool misuse and privilege compromise added to the agent threat model and to the security review.
Owner: Security lead First: Threat model revision
Worth doing, sequenced behind the above. This is where the leadership case is actually made.
Register schema, gate criteria and the currency control released openly under the Open Government Licence, versioned, as Canada did with its assessment.
Owner: Governance office First: Public repository
One partner nation, one deployment class, mutual recognition of assurance evidence. Tests whether our evidence travels before we offer it widely.
Owner: Governance office First: Pilot terms with one partner
Longitudinal measurement of unaided performance in agent-supported roles, with an academic partner. Nobody holds this data, which is why everyone is asserting rather than measuring.
Owner: Evaluation lead First: Study protocol
Named, so the omission reads as a decision rather than an oversight.
Not ours to set. We build the evidence base a policy decision would need.
Owner: — First: Revisit when policy opens
The guidance route moves faster, and we do not yet hold the evidence to legislate well.
Owner: — First: Revisit after PACE-13
Capacity we do not have. Contract terms and evidence requirements instead.
Owner: — First: Revisit at scale
Parity is a floor, and reaching it puts us level with instruments already a year old. Each of these is a control we could be first to publish, and three of the four came out of our own case work rather than the international scan.
Canada names skill maintenance as a prerequisite for oversight. Nobody has made it measurable and revocable — an approval right you can lose by going stale. Publishing the control with tier windows and drill records would be first.
Machinery-of-government change moves an owner; active agent work continues unattended and is later inherited by a team that never approved it. No framework surveyed handles ownership transfer on reorganisation.
Approval capacity treated as a finite, costed resource and rationed by risk tier, so review is real where it matters instead of uniform and hollow everywhere.
Australia's institute named cross-organisational agent risk as a gap no framework covers. A register field plus a handshake at departmental and supplier boundaries would be the first implementation of it.
Four of the eight instruments above were published or revised in the last eight months. A guide refreshed annually is a guide that is wrong for eleven of them.
Weekly
Horizon scan into the guidance ticker: new instrument, new version, new enforcement action.
Fortnightly
Instrument diff across the eight tracked jurisdictions. What changed, what it obliges, whether it touches a control we hold.
Monthly
Re-MoSCoW the backlog. Anything sitting in Should for two cycles is promoted or killed, never carried.
Quarterly
Publish the delta pack and contribute one reusable instrument to a shared catalogue.
Annually
Reciprocal review with a partner nation: they read our instruments, we read theirs, both publish the findings.
Five routes already open to us. None of them takes a position paper; all of them take a working instrument.
The network's remit includes a catalogue of open-source, shareable AI solutions for members. A register schema and a currency control are exactly that shape.
The comparative baseline every country is measured against. Contributing method, rather than only returning survey answers, changes what the baseline measures.
The first binding international AI treaty, on human rights, democracy and the rule of law. The UK is a signatory, and the validity of human oversight sits squarely inside its scope.
The UK allocated roughly A$460m at the 2025 spending review, against about A$50m over five years in Canada and about A$11m a year in Singapore. Capacity to convene is not the constraint.
Open Government Licence v3.0, public repository, versioned. Canada published its assessment questionnaire openly and it became the reference others adopted.
Sources: Guide on the Use of Agentic Artificial Intelligence and Directive on Automated Decision-Making, Treasury Board of Canada Secretariat; EU AI Act Article 4 and the AI Office Q&A on AI literacy; IMDA Model AI Governance Framework for Agentic AI v1.5 (20 May 2026); Australian AI Safety Institute inaugural report (9 August 2026); Singapore Consensus companion report on agentic risk; OECD Digital Government Outlook 2026 and Building an AI-ready public workforce (2026); G7 GovAI Grand Challenge and GAIN, Treasury Board of Canada Secretariat and Impact Canada. Figures are as published by each source and are not converted or combined.
Working tool
Five questions give a first route. Governance follows what the AI can see, create and do, not which company made it. Nothing you enter leaves your browser.
The first rule that matches wins. A first sort, not an approval: every route other than an approved tool leads to a registered use case.
| When | Route | What it means |
|---|---|---|
| Q1 is Yes | Seek advice before going further | AI must not make or directly change this kind of decision on its own. Register a use case; it takes the specialist route with a named human decision-maker. |
| Q5 is Yes | Try conventional automation first | A rule, form or workflow is usually safer, cheaper and more consistent. Use AI only where judgement over unstructured material is the task. |
| Q3 is Yes or Not sure | Take the agent route | Anything that acts, runs on a schedule or works through steps without you needs an agent design record, approval points and a stop control. |
| Q1 is Not sure, or Q2 is Yes or Not sure | Register a use case | Personal or sensitive information, or an unclear impact on people, needs data readiness and privacy screening before use. |
| Q4 is Yes | Go ahead with an approved tool | Use it within its approved uses and check its catalogue entry for conditions. Review every output before you rely on it. |
| Otherwise | Check for reuse, then register | No approved tool is known to fit. Search your organisation's catalogue first; if nothing fits, request a tool through a use case. |
Knowledge map
Parts on the inner ring, chapters on the outer ring; select any node to go to it. The taxonomy is published as SKOS in JSON-LD and the content as JSON, so other portals can harvest both.
Transparency
Ported from an internal AI agent governance toolkit on 18 November 2026. Each chapter was rendered, classified as public or withheld, redacted where needed and rebuilt as plain semantic content. Figures the source labels illustrative stay labelled illustrative; nothing here is an official policy position. The source has 87 chapters: 76 are published here and 11 are withheld or merged into the rest of the portal. The decision for every chapter is in the content classification (opens GitHub).
| Chapter | Part | Decision and reason |
|---|---|---|
| AI tool catalogue | Decide and prioritise | Organisation-specific approved-tool statuses and links to internal pages; publishing them would read as endorsement. |
| Analytics, estimators & dashboards | Decide and prioritise | Its charts count entries in the withheld tool catalogue; the estimators it links to are published in "Cost, value and investment". |
| Blueprint coverage | Product documentation | Coverage map of internal source documents and blueprints, which are not published. |
| Portal specification trace | Product documentation | Traces requirements from an internal specification document. |
| Knowledge graph | Knowledge and navigation | Interactive graph of the internal edition; replaced by the generated knowledge map in this playbook. |
| Portal & site map | Knowledge and navigation | Map of the internal edition, including internal pages. The public playbook has its own contents and drawer. |
| Guided tours | Knowledge and navigation | Guided tours route through internal pages; the home page personas and journeys replace them. |
| Media centre | Knowledge and navigation | Films and decks: never kept in a repository (owner rule). They stay in the media library outside GitHub. |
| Self-assurance console | Knowledge and navigation | Self-test console of the internal edition; the public site runs its own checks in CI (see Transparency). |
| Sources & blueprints | Reference | Lists internal source documents. Public sources are in the resource library on the home page. |
| Glossary | Knowledge and navigation | Merged into the portal glossary, so the whole portal has one glossary: 39 new terms added, existing definitions kept. |
tools/build_playbook.py; check S21 fails the build if they drift.Six tools that turn the guidance into an artefact you can paste into a ticket, a change record or a board paper. Everything runs in your browser — nothing you type is sent anywhere, stored or transmitted.
Privacy by design: these tools are client-side only. Do not enter personal data, case details or anything above OFFICIAL. Use references and descriptions instead.
Your result will appear here.
Produces a copy-ready record. Attach it to the artefact, the ticket or the change record.
Your result will appear here.
A screening tool, not a substitute for a full equality impact assessment. It tells you whether you need one.
Your result will appear here.
Drafts a public-facing record of how an algorithmic or AI tool is used. Publish it, and link it from the service.
Your result will appear here.
Your result will appear here.
Paste AI-drafted content to check reading difficulty, sentence length, jargon and non-inclusive phrasing before publication.
Your result will appear here.
Every asset is designed to be catalogued, versioned, reviewed and reused. Filter by purpose, or search across names, descriptions and linked standards.
Showing 18 of 18 toolkit assets.
Machine-readable schemas for skills, agents, prompts, tools, model-use events, evidence packs and source provenance.
Output: JSON Schema + YAML template
Anchored to: PROV-O (opens an external website), CycloneDX ML-BOM (opens an external website), DCAT 3 (opens an external website)
Approved, draft and deprecated skills with owners, sources, risk ratings, compatibility and evidence rules.
Output: Register + lifecycle policy
Anchored to: Model cards (opens an external website), alphagov (opens an external website)
Bounded agents by autonomy level, permissions, allowed skills, prohibited actions, model constraints and kill-switch owner.
Output: Register + autonomy policy
Anchored to: MCP (opens an external website), OWASP LLM Top 10 (opens an external website)
Source consultation records, AI-use declarations, validation results, human acknowledgements and audit trails.
Output: Generator on this page
Anchored to: ATRS (opens an external website)
Equality, accessibility, assisted digital, digital inclusion, protected characteristic and vulnerability screening.
Output: Screening tool on this page
Anchored to: PSED (opens an external website), WCAG 2.2 (opens an external website)
Inventory cryptography, identify long-life data, assess harvest-now-decrypt-later exposure and plan crypto-agility.
Output: Assessment tool on this page
Anchored to: NCSC PQC (opens an external website), NIST PQC (opens an external website)
Risk-based routing that decides how many eyes, which specialists and what evidence an AI-assisted output needs.
Output: Triage tool on this page
Anchored to: NIST AI RMF (opens an external website)
Reviewed prompt patterns with source-citation requirements, refusal conditions, data-handling rules and test cases.
Output: Pattern library + tests
Anchored to: AI Playbook for the UK Government (opens an external website)
Executable checks turning WCAG, Service Standard, Secure by Design and TCoP clauses into automated tests.
Output: CI pack
Anchored to: Pa11y (opens an external website), Secure by Design (opens an external website)
Produce an ATRS-shaped public record for any algorithmic or AI-assisted tool in use.
Output: Generator on this page
Anchored to: ATRS hub (opens an external website)
Questions, evidence asks and contract clauses for AI suppliers: model provenance, evaluations, accessibility, exit.
Output: Question bank + clauses
Anchored to: ISO/IEC 42001 (opens an external website), EN 301 549 (opens an external website)
Report AI failures, hallucinations, bias incidents, prompt injection and source drift with learning loops.
Output: Process + register
Anchored to: MITRE ATLAS (opens an external website), NCSC secure AI (opens an external website)
Signed, tamper-evident record of which model, version, prompt, source and human touched each artefact.
Output: Ledger spec
Anchored to: C2PA (opens an external website), SLSA (opens an external website)
Estimate energy, carbon and unit cost of model calls, and test whether a non-AI option is proportionate.
Output: Calculator + reporting
Anchored to: Greening Government ICT (opens an external website), W3C WSG (opens an external website)
Recruitment, consent, reimbursement and session plans that reach disabled, digitally excluded and seldom-heard users.
Output: Research kit
Anchored to: Research with disabled people (opens an external website), Assisted digital (opens an external website)
Test AI-drafted content against GOV.UK style, reading age 9, and inclusive terminology before publication.
Output: Checker on this page
Anchored to: GOV.UK style guide (opens an external website)
Classify AI interaction logs as records; set retention, disclosure and redaction rules ahead of FOI/SAR requests.
Output: Retention schedule
Anchored to: DPA 2018 (opens an external website)
Role-based learning paths, licence-to-operate checks and refreshers for anyone using AI on public work.
Output: Curriculum map
Anchored to: DDaT framework (opens an external website)
No assets match. Try a broader term or reset the filter.
Risk-based human-in-the-loop controls make sure AI never grants authority a user does not already possess, and that every material output can be defended after the fact.
| Level | Name | What it means | Typical examples | Evidence |
|---|---|---|---|---|
| H0 | No review needed | Personal, throwaway, non-published exploration with no personal or sensitive data. | Explaining a concept to yourself; drafting private notes. | None required. |
| H1 | Self-check | Author verifies facts, sources and accessibility before use. Low impact, easily reversible. | Internal meeting notes, first-draft ideas. | AI-use note in the document. |
| H2 | Peer check | A second competent person checks output before it leaves the team. | Internal reports, non-public analysis, low-risk code in a sandbox. | Reviewer name, date, issues found. |
| H3 | Specialist review (4-eyes) | A named specialist in the relevant discipline reviews. No self-approval. | Accessibility, security, legal, statistical or clinical content; production code. | Specialist sign-off + test results. |
| H4 | Multi-specialist review (6-eyes) | Two or more specialisms plus the accountable owner. Independent challenge required. | Public-facing decisions, casework support, anything affecting entitlements. | Panel record, dissent captured, residual risk stated. |
| H5 | Governance board and publication | Formal board approval, published transparency record and monitored deployment. | Automated or semi-automated decisions about people; national services. | ATRS record, DPIA, EQIA, monitoring plan, review date. |
| Activity | Responsible | Accountable | Consulted | Informed |
|---|---|---|---|---|
| Classifying risk and review level | Author | Service owner | Specialist reviewers | Delivery team |
| Verifying sources and outputs | Author | Author | Peer reviewer | Service owner |
| Specialist sign-off (H3+) | Named specialist | Service owner | Author | Assurance board |
| Equality and accessibility screening | Author | Service owner | Equality & accessibility leads | Governance board |
| Data protection and lawful basis | Product owner | Data controller / SIRO | DPO | ICO where required |
| Publishing a transparency record | Service owner | SRO | Comms, legal | Public |
| Incident and near-miss reporting | Anyone who spots it | Service owner | Security, DPO | Assurance board |
Every reusable AI capability should be discoverable, accountable, interoperable and governed like any other public asset — with an owner, a version, a risk rating and a retirement plan.
| Field | Why it matters |
|---|---|
asset_id | Stable identifier so evidence, incidents and dependencies can point at it. |
owner | A named person, not a team inbox. Accountability needs a face. |
status | draft / approved / restricted / deprecated / withdrawn. |
version | Semantic versioning; breaking changes force re-review. |
authority_tier | 0–3 tier of the sources the asset relies on. |
risk_level / human_review | H0–H5 floor that consumers must apply. |
source_provenance | Which guidance, data and code it derives from, with dates. |
permissions | Least-privilege scopes, tools and credentials the asset may use. |
prohibited_actions | Explicit red lines an agent must never cross. |
deia_screen | Date and outcome of equality and accessibility screening. |
evaluation | Test suite, pass thresholds and last run. |
compatibility | Models, runtimes and versions it is validated against. |
lifecycle | Review cadence, next review, deprecation and retirement date. |
contact / redress | How to report a problem or challenge an outcome. |
asset_type: ai_skill
asset_id: review-accessibility-govuk
title: GOV.UK accessibility review assistant
status: approved
version: 2.1.0
owner: AI assurance lead (role, not a named person)
authority_tier: 1
human_review: H3
specialist_review: [accessibility, content_design]
self_approval: false
source_provenance:
- id: wcag22
url: https://www.w3.org/TR/WCAG22/
checked: 2026-07-14
next_check: 2026-10-14
- id: govuk-design-system
url: https://design-system.service.gov.uk/
checked: 2026-07-14
permissions:
tools: [read_page, run_axe]
data_classification_max: OFFICIAL
prohibited_actions:
- claiming_conformance_without_test_evidence
- processing_personal_data
- autonomous_publication
deia_screen:
completed: 2026-06-02
outcome: mitigations_applied
evaluation:
suite: a11y-regression-v4
threshold: 0.95
last_run: 2026-08-01
compatibility:
models: [">=v3.1"]
runtimes: [node20, python3.12]
evidence_rules:
require_citations: true
require_reviewer_signoff: true
retain_for: P7Y
lifecycle:
review_cadence: P3M
next_review: 2026-11-01
deprecation: null
contact: https://github.com/cyberdudeuk/ai-capability-governance/issues/newInteroperability note: map these fields to DCAT 3 (opens an external website) for cataloguing, PROV-O (opens an external website) for evidence chains and CycloneDX ML-BOM (opens an external website) for supply chain, so registries in different organisations can federate rather than fork.
Inclusion is a delivery activity, not a compliance report. This section states what we commit to, how we test it, and how you can hold us to it.
Every asset traces to a named user need and a measurable outcome. If we cannot name the need, we do not build the feature.
Recruit disabled participants, assistive technology users, people with low digital confidence and people in vulnerable circumstances — in every round, not a one-off "accessibility round".
Manual testing with screen readers, magnification, voice control, switch access and keyboard-only. Automated tools catch a minority of real barriers.
Including offline, assisted digital, third-party support, and what happens when the AI is unavailable or wrong.
Plain English, short sentences, no unexplained acronyms, inclusive terminology, and alternative formats on request.
Track drop-out by channel and need, assisted digital volumes, complaint themes and accessibility issue ageing — and publish them.
prefers-reduced-motion and no motion-only meaningAccessibility statement guidance → (opens an external website)
Known equity risks of AI-assisted delivery. Performance that degrades for under-represented groups; language models that default to majority dialects and names; outputs written above the reading age of the audience; automation that removes the human route people in crisis rely on; accessibility regressions in AI-generated code and content; and reviewers who trust fluent text more than accurate text.
What we do about them. Screen every change with the DEIA tool; disaggregate evaluation by affected group where lawful and proportionate; keep a human route that cannot be removed by efficiency targets; test AI-generated interfaces with assistive technology before merge; publish limitations honestly; and give people a real way to challenge an outcome.
This code complements behavioural codes of conduct and the Civil Service values. It defines operational rules for model-backed tools, direct AI use, coding assistants, multi-agent orchestration and AI-generated artefacts.
Anyone may pause an AI-assisted process they believe is unsafe, unfair or unlawful, without needing permission and without detriment. Near-misses should be reported as readily as incidents — they are the cheapest learning available.
Members of the public affected by an AI-assisted outcome must be able to find out that AI was involved, understand the outcome in plain English, request a human review, and escalate to an independent route. If any of those four is missing, the service is not ready.
The corpus is organised as a knowledge graph linking laws, policies, standards, design guidance, security controls, skills, agents and evidence — so that a rule can always be traced to its source and its consequences.
Index GOV.UK, GDS, CDDO, departmental, international standards, GitHub and engineering-lab sources with authority tiering, freshness checks and applicability metadata.
Link services, repositories, datasets, APIs, controls, skills, agents, people, approvals, models, risks and evidence so impact analysis is a query, not a workshop.
Translate policy and standards into schemas, conformance packs, tests, scripts, checklists and prompt patterns that run in delivery pipelines.
When two sources disagree, the lower tier number wins. Record the conflict and the resolution in the evidence pack.
| Tier | Source type | Examples | Treatment |
|---|---|---|---|
| 0 | Law and statutory instruments | Equality Act 2010, DPA 2018, Accessibility Regulations 2018 | Binding. Non-compliance is not a risk to accept — escalate. |
| 1 | Mandated standards and cross-government policy | Service Standard, Technology Code of Practice, Secure by Design, WCAG 2.2, ATRS | Comply or record a formal, time-boxed exception with an owner. |
| 2 | Recognised international and professional standards | ISO/IEC 42001, NIST AI RMF, OWASP LLM Top 10 | Adopt as good practice; justify divergence. |
| 3 | Community, vendor and research material | Blogs, model cards, toolkits, preprints | Useful signal. Never the sole basis for a control or a public claim. |
68 curated sources, each with its authority tier and a note on what it is actually useful for. Search, filter, then follow the link to the source of truth.
Showing 69 of 69 resources.
| Resource | Publisher | Type | Tier | Use it for |
|---|---|---|---|---|
| AI Playbook for the UK Government (opens an external website) | DSIT | Policy | Tier 1 | Ten principles plus per-technology guidance for safe, effective public sector AI adoption. Replaced the Generative AI Framework for HMG, withdrawn on 10 February 2025. |
| Algorithmic Transparency Recording Standard (ATRS) (opens an external website) | DSIT / CDDO | Standard | Tier 1 | Mandatory-style template for publishing how an algorithmic tool is used in decision-making. |
| Data Ethics Framework (opens an external website) | DSIT | Policy | Tier 1 | Structured ethical appraisal for public sector data and AI projects. |
| Service Standard (14 points) (opens an external website) | GDS | Standard | Tier 1 | The assessment backbone for any public service — align AI features to each point. |
| Government Service Manual (opens an external website) | GDS | Guidance | Tier 1 | End-to-end practice guidance for user research, design, technology and operations. |
| Government Design Principles (opens an external website) | GDS | Guidance | Tier 1 | Ten principles — start with user needs, do less, design with data, do the hard work to make it simple. |
| Technology Code of Practice (opens an external website) | CDDO | Standard | Tier 1 | Criteria for technology spend approval — interoperability, open standards, security, sustainability. |
| Digital, Data and Technology Capability Framework (opens an external website) | GDS | Toolkit | Tier 1 | Map AI-related skills to recognised roles and progression levels. |
| Open source guidance (coding in the open) (opens an external website) | GDS | Guidance | Tier 1 | When code should be open or closed, and how to publish safely. |
| GDS API technical and data standards (opens an external website) | GDS | Standard | Tier 1 | How to design, document and version APIs that agents and skills can call. |
| WCAG 2.2 (opens an external website) | W3C | Standard | Tier 1 | The normative success criteria — target AA as a minimum, AAA where feasible. |
| ARIA Authoring Practices Guide (APG) (opens an external website) | W3C WAI | Guidance | Tier 1 | Accessible patterns for tabs, accordions, comboboxes, dialogs and more. |
| EN 301 549 accessibility requirements (opens an external website) | ETSI/CEN/CENELEC | Standard | Tier 1 | The procurement-facing accessibility standard referenced by UK/EU regulations. |
| Accessibility requirements for public sector websites and apps (opens an external website) | GOV.UK | Guidance | Tier 1 | What the 2018 regulations require, including the accessibility statement duty. |
| Public Sector Bodies (Websites and Mobile Applications) Accessibility Regulations 2018 (opens an external website) | legislation.gov.uk | Law | Tier 0 | The underlying legal duty. |
| Equality Act 2010 (opens an external website) | legislation.gov.uk | Law | Tier 0 | Protected characteristics, reasonable adjustments and discrimination duties. |
| Public Sector Equality Duty guidance (opens an external website) | EHRC | Guidance | Tier 1 | Due regard duty — evidence it in every AI-assisted decision or service change. |
| Make your service accessible: an introduction (opens an external website) | GDS | Guidance | Tier 1 | Practical starting point for teams new to accessibility duties. |
| Publish an accessibility statement (opens an external website) | GOV.UK | Toolkit | Tier 1 | Model wording and required disclosures for non-accessible content. |
| Research with people with disabilities and access needs (opens an external website) | GDS | Guidance | Tier 1 | How to recruit and run inclusive sessions, including with assistive technology users. |
| Helping people to use your service (assisted digital) (opens an external website) | GDS | Guidance | Tier 1 | Assisted digital support and digital inclusion planning. |
| Inclusive Design Principles (opens an external website) | Paciello/TPGi et al. | Guidance | Tier 2 | Seven principles: provide comparable experience, be consistent, give control, offer choice. |
| Inclusive Design Toolkit (opens an external website) | Microsoft | Toolkit | Tier 3 | Persona spectrum and exclusion-mapping activities for workshops. |
| Home Office accessibility "dos and don’ts" posters (opens an external website) | UK Home Office | Toolkit | Tier 1 | Print-ready posters for autistic users, screen readers, dyslexia, low vision, D/deaf users. |
| GOV.UK Design System (opens an external website) | GDS | Toolkit | Tier 1 | Tested, accessible components and patterns — reuse before you invent. |
| GOV.UK Prototype Kit (opens an external website) | GDS | Toolkit | Tier 1 | Rapid, realistic prototypes for usability testing without production risk. |
| GOV.UK style guide (content design) (opens an external website) | GDS | Guidance | Tier 1 | Plain English, reading age, formatting and inclusive terminology. |
| WebAIM contrast checker (opens an external website) | WebAIM | Tool | Tier 3 | Verify 4.5:1 text and 3:1 non-text contrast before shipping. |
| Pa11y automated accessibility testing (opens an external website) | Pa11y | Tool | Tier 3 | Add automated WCAG checks to your pipeline — never as a substitute for manual testing. |
| axe DevTools (opens an external website) | Deque | Tool | Tier 3 | Browser-based rule engine used widely for triage. |
| Guidelines for secure AI system development (opens an external website) | NCSC + international partners | Guidance | Tier 1 | Secure design, development, deployment and operation across the AI lifecycle. |
| Secure by Design (UK government) (opens an external website) | Cabinet Office / GSG | Standard | Tier 1 | Mandated activities and artefacts for digital service security in HMG. |
| Cyber Assessment Framework (CAF) (opens an external website) | NCSC | Standard | Tier 1 | Outcome-based objectives A–D for essential services and government systems. |
| Post-quantum cryptography guidance and migration timeline (opens an external website) | NCSC | Guidance | Tier 1 | Discovery by 2028, high-priority migration by 2031, complete by 2035. |
| Post-Quantum Cryptography standards (FIPS 203/204/205) (opens an external website) | NIST | Standard | Tier 1 | ML-KEM, ML-DSA and SLH-DSA — the algorithms your crypto inventory should target. |
| OWASP Top 10 for LLM Applications (opens an external website) | OWASP | Standard | Tier 2 | Prompt injection, insecure output handling, supply chain and agency risks. |
| MITRE ATLAS adversarial ML knowledge base (opens an external website) | MITRE | Toolkit | Tier 2 | Tactics and techniques for threat modelling AI systems. |
| Secure Software Development Framework (SP 800-218) (opens an external website) | NIST | Standard | Tier 2 | Baseline secure SDLC practices to bind AI-generated code to. |
| SLSA supply chain integrity framework (opens an external website) | OpenSSF | Standard | Tier 2 | Provenance levels for build artefacts — extend to model and skill artefacts. |
| CycloneDX ML-BOM / AI-BOM (opens an external website) | OWASP CycloneDX | Standard | Tier 2 | Machine-readable bill of materials for models, datasets and AI components. |
| Guidance on AI and data protection (opens an external website) | ICO | Guidance | Tier 1 | Lawful basis, fairness, transparency and accountability for AI processing. Under review by the ICO after the Data (Use and Access) Act 2025. |
| AI and data protection risk toolkit (opens an external website) | ICO | Toolkit | Tier 1 | Risk statements and controls mapped to the AI lifecycle. Under review by the ICO after the Data (Use and Access) Act 2025. |
| Data protection impact assessments (DPIA) (opens an external website) | ICO | Toolkit | Tier 1 | When a DPIA is mandatory and what a defensible one contains. Under review by the ICO after the Data (Use and Access) Act 2025. |
| Data Protection Act 2018 (opens an external website) | legislation.gov.uk | Law | Tier 0 | UK GDPR implementation. Its automated decision-making rules were rewritten by the Data (Use and Access) Act 2025. |
| Data (Use and Access) Act 2025 (opens an external website) | legislation.gov.uk | Law | Tier 0 | Replaced UK GDPR Article 22 with Articles 22A to 22D from 5 February 2026: a significant decision taken with no meaningful human involvement needs specific safeguards, including the right to contest it. |
| Automated decision-making, including profiling (opens an external website) | ICO | Guidance | Tier 1 | What counts as meaningful human involvement, and the safeguards significant automated decisions need. Draft updated for the 2025 Act; the consultation closed on 29 May 2026. |
| ISO/IEC 42001 AI management systems (opens an external website) | ISO/IEC | Standard | Tier 1 | Certifiable management system for responsible AI — useful for supplier assurance. |
| ISO/IEC 23894 AI risk management (opens an external website) | ISO/IEC | Standard | Tier 1 | AI-specific application of ISO 31000 risk management. |
| ISO/IEC 27001 information security management (opens an external website) | ISO/IEC | Standard | Tier 1 | Control baseline most suppliers will already hold. |
| AI Risk Management Framework (AI RMF 1.0) + GenAI Profile (opens an external website) | NIST | Standard | Tier 1 | Govern, Map, Measure, Manage — pairs cleanly with UK assurance practice. |
| EU AI Act (Regulation 2024/1689) (opens an external website) | EUR-Lex | Law | Tier 0 | Risk tiers and obligations that may bind UK bodies operating in the EU market. |
| OECD AI Principles (opens an external website) | OECD | Policy | Tier 2 | Internationally agreed values-based principles for trustworthy AI. |
| Recommendation on the Ethics of AI (opens an external website) | UNESCO | Policy | Tier 2 | Global normative instrument covering human rights and inclusion. |
| Understanding AI ethics and safety (opens an external website) | The Alan Turing Institute | Guidance | Tier 2 | The SUM values and FAST Track principles used across UK public sector. |
| AI Standards Hub (opens an external website) | BSI / NPL / Turing | Toolkit | Tier 2 | Searchable database of AI standards in development and published. |
| Ada Lovelace Institute research (opens an external website) | Ada Lovelace Institute | Research | Tier 3 | Public attitudes, algorithmic impact assessment and deliberative methods. |
| Open Data Institute (opens an external website) | ODI | Research | Tier 3 | Data ethics canvas and data institution patterns. |
| Model Cards for Model Reporting (opens an external website) | Mitchell et al. | Research | Tier 2 | The reference format for documenting intended use and performance by group. |
| Datasheets for Datasets (opens an external website) | Gebru et al. | Research | Tier 2 | Provenance, composition and collection questions for training/eval data. |
| PROV-O provenance ontology (opens an external website) | W3C | Standard | Tier 1 | Entity–activity–agent model for machine-readable evidence chains. |
| DCAT 3 data catalogue vocabulary (opens an external website) | W3C | Standard | Tier 1 | Describe datasets and services so registries can federate. |
| C2PA content provenance and authenticity (opens an external website) | C2PA | Standard | Tier 2 | Cryptographic manifests for AI-generated or edited media. |
| Model Context Protocol (opens an external website) | MCP | Standard | Tier 3 | Open protocol for connecting assistants to tools and data with explicit scopes. |
| Model cards on Hugging Face Hub (opens an external website) | Hugging Face | Toolkit | Tier 3 | Practical templates and YAML front-matter conventions. |
| alphagov on GitHub (opens an external website) | GDS | Toolkit | Tier 1 | Reference implementations, frontend, and publishing platform source. |
| x-govuk community tools (opens an external website) | Cross-government community | Toolkit | Tier 2 | Community-built plugins, design tooling and prototypes. |
| Open Government Licence v3.0 (opens an external website) | The National Archives | Law | Tier 0 | Default licence for publishing public sector content and guidance. |
| Greening Government ICT and Digital Services Strategy (opens an external website) | Defra / GGC | Policy | Tier 1 | Carbon, circular-economy and reporting expectations for digital estates. The 2020 to 2025 period has ended; the 2024 to 2025 annual report (July 2026) says a new government digital sustainability strategy is coming. |
| Web Sustainability Guidelines (WSG) (opens an external website) | W3C Community Group | Guidance | Tier 3 | Practical efficiency criteria for pages, media, hosting and AI workloads. |
Content reviewed on 2026-11-13: every link checked, automatically or by hand in a browser where a site refuses automated checks, and each page read for withdrawal, replacement or review notices. Next review due 2027-02-13. Guidance changes without notice: always confirm you are reading the current version at the source.
29 signals on governing AI agents and the data they use, gathered from public LinkedIn content and refreshed weekly. Each names its author, links to where it was found and says how far it has been checked. Inclusion is not endorsement.
Read with care. LinkedIn does not allow automated reading of posts, so every summary comes from search-indexed text and is paraphrased. 15 of 29 come from vendors, consultancies or law firms and are marked. Statistics without a primary source are flagged as unverified. Check the source before you quote anything. The register is open data: ai-data-governance-signals.json (scanned 2026-11-09, version 1.2.0).
Lee Stockwell, Databricks · 2026
Activity highlights George Bettsworth, Head of Data Science at the Foreign, Commonwealth and Development Office, putting AI progress down to good governance, with the argument that governance should sit at the data layer rather than in each application. Also trails an Innovation 2026 panel with Sue Bateman (Defra CDO), Jacqui Leggetter (DWP) and Aimee Smith (Government Chief Data Officer).
Why it matters: Direct UK government voices on governing AI through the data estate; supports BOT's position that agent governance starts with data governance.
Caution: Vendor-hosted content; the FCDO view is relayed by the vendor, not quoted from FCDO.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Elizabeth Seger, Tony Blair Institute for Global Change · 2026 (about one month before scan)
Argues AI governance is shifting from principles to implementation and that governance capacity is what lets organisations adopt AI agents safely. Points to the UK Government's Trusted Third-Party AI Assurance Roadmap (September 2025), tied to commitment 29 of the AI Opportunities Action Plan.
Why it matters: Links BOT's assurance pages to a named UK policy instrument and commitment number.
Caution: Roadmap and commitment number should be checked against GOV.UK before citing in BOT's evidence base.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Usman Ikhlaq, techUK (author); now University of Bath Institute for Policy Research · 21 July 2026
Profile lists a techUK report he authored on moving agentic AI beyond pilots to responsible deployment at scale across the UK economy, including how oversight and accountability keep pace.
Why it matters: UK industry baseline for agent oversight; a comparator for BOT's agent governance playbook.
Caution: Report itself not read; summary is from the profile listing.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
AI Security Institute (AISI), UK Government · 2026
AISI's page links to NCSC guidance on managing the cyber risk of agentic AI and to its Alignment Red Team work. The page describes AISI as part of the Cabinet Office.
Why it matters: Authoritative UK security sources BOT should cite for agent permissions and runtime controls.
Caution: Departmental home is as stated on the LinkedIn page; confirm on aisi.gov.uk given recent machinery of government changes.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Keith Vincent (speaker); hosted by Ashurst Perkins Coie, Government Legal Department, Treasury Legal Advisers · 7 October 2026
Event activity frames AI agents around autonomy, accountability and where control boundaries sit, and notes that customers' own agents bring new liquidity and fraud risks.
Why it matters: Government legal perspective on the same question BOT answers: what an agent may do without a human.
Caution: Event summary written by the host firm, not by the speaker.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Sophie Sheldon (quoted in Law360 UK); liked by Devina Shah, Simmons & Simmons · 2026
Activity reports the FCA's plan to become an agentic regulator, with a warning about over-reliance on a single model, platform or provider, and separately references new FCA guidance on AI in financial services.
Why it matters: A UK regulator itself adopting agents: a live case for BOT's concentration-risk and vendor-neutrality guidance.
Caution: Secondary report via Law360; check the FCA's own statement.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Yasmin Liverpool, Chief Data Office, National Highways · Undated (likely 2025)
LinkedIn article proposing shared accountability for AI governance (for example the CDO with the CSO) or a cross-executive AI steering group, citing the NAO finding that poor data quality is a barrier to AI in government and the Generative AI Framework for HM Government as a starting point.
Why it matters: A practising government CDO's operating-model view; maps onto BOT's roles and accountability pages.
Caution: Predates later UK guidance that replaced the Generative AI Framework; treat as historical context.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Melvyn White, Agentic AI strategy, governance and management (independent, London) · Current profile
Profile thesis: the gap between agent adoption and the ability to govern agents is the core risk.
Why it matters: A crisp framing of the governance-debt problem BOT exists to close.
Caution: Opinion; no evidence attached in the indexed text.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Jake Boudsocq, Department of Government Services (Victoria, Australia) · 2026 (about five months before scan)
Takeaways from the 2026 Melbourne Data Protection and Security Summit: data governance is the operating model that makes responsible AI possible at scale; uncatalogued legacy data compounds risk; Digital Transformation Agency model clauses are becoming a public sector baseline; AI governance is not set-and-forget.
Why it matters: Peer-government evidence for BOT's 'know your data before AI' gate and for catalogue-first sequencing.
Caution: Australian context; DTA clauses are not UK instruments.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Partnership on AI (reposted by Stephanie Ifayemi), Partnership on AI · 2026 (about eight months before scan)
Sets out priorities led by governing AI agents with accountability, security and privacy, plus stronger documentation and public reporting across the AI value chain and shared international baselines. Also notes work on how public sector AI governance differs from private sector approaches.
Why it matters: Multi-stakeholder agenda that BOT's agent registry and transparency pages already address.
Caution: Agenda-setting piece; not guidance.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Patrick Upmann, AIGN Global · 2026
As agents move from generating information to taking actions across data, tools and systems, the question becomes how governance decisions are enforced at runtime. Situates this against ISO/IEC 42001, the NIST AI Risk Management Framework and the EU AI Act.
Why it matters: Matches BOT's 'the app is the leash' principle: permissions enforced by scope, not policy text.
Caution: Author markets a proprietary governance framework.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Roberto Mandiola, Amazon Web Services (post about MuleSoft and Google Cloud) · 2026
Describes scanners that find AI agents and capture metadata on what each can do, its underlying model and its data access rights, mapped to the agent-to-agent protocol's agent card so organisations hold one registry of agents, including home-built agents and MCP servers.
Why it matters: Strongest metadata link in this scan: an agent register is a metadata problem, which ties BOT to the Metadata Exchange Model and DCAT work.
Caution: Vendor announcement relayed in another person's activity feed (indexed under Andrea Radojicic's profile).
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Bigeye (and others in the feed), Bigeye · 2026
Argues that governing agents that read enterprise data differs from human-era data governance because agents fail at different points in the pipeline; recommends knowing data quality signals and what data agents access before rollout. The feed also cites a Sinch survey of 2,527 senior decision-makers in which 74% had rolled back a live agent after a governance failure.
Why it matters: Useful framing for BOT's data-access controls; the rollback figure is a candidate indicator.
Caution: Survey figure is vendor-reported and unverified; do not use in BOT's evidence base until the Sinch source is read.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Subeer Sehgal, Grazitti Interactive · 2026
A playbook for governing agents: where risk concentrates, which controls hold and how to prove compliance as agent numbers grow. Warns that unvalidated feedback loops, where agents treat their own inferences as fact, are the real blind spot. The company feed claims 54% of organisations have between 1 and 100 unsanctioned agents.
Why it matters: Supports BOT's agent inventory and lineage controls.
Caution: The 54% figure has no stated source; treat as unverified.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Robert Noyce, Shell · 2026
Agents need business context, semantic definitions, governance, memory and trusted data; a knowledge layer linking business definitions to data gives agents that context.
Why it matters: Endorses the glossary, ontology and knowledge-graph layer the govukOS build standard already requires.
Caution: Practitioner reflection after an event; no evidence attached.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Joyce Goh, Salesforce (Informatica) · 2026
Data readiness affects AI economics, and governance must evolve as AI moves from generating insight to agents taking action, needing more visibility, accountability and control.
Why it matters: Corroborates S01 and S09 from the platform side.
Caution: Vendor promotional content.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Herman Yu, Teeup Data Technology · 2026
Data governance and data quality issues are long-standing; the non-deterministic behaviour of LLMs and agents stretches them further, so enterprise architecture must join governance dimensions into one framework.
Why it matters: Supports treating agent governance as an extension of data governance, not a new silo.
Caution: Opinion.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Robert Gasparovic, HAVEL & PARTNERS, Slovak Republic · 2026
Notes the EU AI Act has applied since 2 August 2026 but Member States differ in enforcement readiness; Europe is also revisiting text-and-data-mining rules.
Why it matters: Cross-border context for BOT's regulatory crosswalk.
Caution: Applicability dates differ by obligation under the Act; check the specific article before citing.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Pascal Jiang (post by Rimon PC), Dacheng Law Offices / Rimon PC · July 2026
Three measures effective July 2026 move China from principles to operational governance, including AI ethics guidance, a framework for anthropomorphic AI services and a policy framework for governing AI agents.
Why it matters: One of the first national-level agent governance frameworks; a comparator for BOT.
Caution: Law-firm summary; the primary texts are not read.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Anthony Akpan, Tizee Consult; Institutional AI Governance newsletter · April and July 2026
An SSRN preprint (April 2026) on institutional readiness as a condition for responsible AI in the public sector, and a newsletter launched in July 2026 on institutional responsibility in AI governance.
Why it matters: Academic framing of readiness that maps onto BOT's maturity and adoption model.
Caution: Preprint, not peer reviewed; focus is developing countries.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Brad Martineau, LinkedIn article · Undated
Compares Credo AI, IBM watsonx.governance, Holistic AI, Fiddler AI and Atlan, noting that a data governance and metadata platform such as Atlan supplies the lineage regulators expect.
Why it matters: Market scan only; useful for BOT's tooling landscape, not for recommendations.
Caution: Listicle; possible commercial interest. BOT stays vendor-neutral.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Tess Buckley, Post about the AI Security Institute (DSIT) · 18 December 2025 (derived from the post's activity ID)
Post announcing that the UK AI Security Institute, part of DSIT, published its first Frontier AI Trends Report.
Why it matters: A UK government evidence source on frontier AI capability and risk that BOT's assurance material can cite once read at source.
Caution: Only the headline and URL were indexed; LinkedIn refuses page fetches, so the body was not read. Read the report on GOV.UK before citing any finding. Older than the last scan date; indexed only now.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Khyati Thakkar, LinkedIn article · Undated
Article framing AI governance as an extension of existing data governance practice.
Why it matters: Supports BOT's premise that AI and agent governance should be built on data governance foundations.
Caution: Only the headline and URL were indexed; LinkedIn refuses page fetches, so the body was not read. Author name taken from the article URL; organisation and date unknown.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Heather Dawe, LinkedIn article · Undated
Article on how organisations can build AI assurance frameworks.
Why it matters: Possible practitioner input to BOT's assurance pillar.
Caution: Only the headline and URL were indexed; LinkedIn refuses page fetches, so the body was not read. Author name taken from the article URL; organisation, region and date not confirmed.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Mahmoud Abufadda, LinkedIn article · Undated
Article on the rise of AI control planes as a way to govern models, agents and their decisions centrally.
Why it matters: Relates to BOT's control-tower and agent-registry thinking: one governed layer over many agents.
Caution: Only the headline and URL were indexed; LinkedIn refuses page fetches, so the body was not read. Author name taken from the article URL; organisation and date unknown. Check for vendor interest before citing.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
Katharina Koerner, LinkedIn post · 15 February 2024 (derived from the post's activity ID)
Post sharing the UK government's Introduction to AI assurance guidance.
Why it matters: Points to the baseline UK assurance guidance BOT should align with.
Caution: Only the headline and URL were indexed; LinkedIn refuses page fetches, so the body was not read. Dated 2024, so background rather than new signal; check GOV.UK for the current edition.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
The Alan Turing Institute (CETaS), The Alan Turing Institute · 17 January 2024 (derived from the post's activity ID)
Turing Institute post on a CETaS paper about AI risk management and assuring third-party AI systems in a UK national security setting.
Why it matters: Complements S02 on third-party assurance with a public-sector research source.
Caution: Only the headline and URL were indexed; LinkedIn refuses page fetches, so the body was not read. Full paper title is truncated in the URL; dated 2024, so background rather than new signal.
Source on LinkedIn (opens an external website) · evidence: search-indexed text, post not opened
NVIDIA, Auto Ontology (open source, Apache 2.0) · November 2026 (repository read on 13 November 2026)
An open-source service that drafts an ontology from database metadata: business terms mapped to columns, reviewed SQL expressions, and the joins between tables. People then test it with real questions and certify the names, descriptions and SQL. An agent answers questions by writing SQL from those definitions, checks the query against the database, and returns the answer with its SQL and the steps it took. It serves this context to other agents over MCP, and says when a question cannot be answered from the data.
Why it matters: It makes certification a separate human step from drafting, which is where this toolkit's H0 to H5 review levels apply: see the advanced guide's worked examples on certifying a definition and reviewing an agent's answer, and the evidence pack's field for the definitions an answer relied on.
Caution: Vendor project, and it does not accept outside contributions. Accuracy is measured with a separate text-to-SQL benchmark repository; this review did not run it or the software.
Source repository on GitHub (opens an external website) · evidence: primary source, the repository's README (summary and how it works)
Luís Rodrigues, Not stated on the post (the author's headline describes him as a CPTO; he writes the newsletter Build What Matters) · About 7 November 2026 (the post showed '1w' when read on 17 November; the date is decoded from its activity ID)
Frames AI governance as four layers: regulation (the EU AI Act, plus GDPR, sector rules, NIS2 and DORA), standards and frameworks (ISO/IEC 42001 and the NIST AI RMF), roles (a named owner for every AI system) and controls. Argues that controls are where most organisations fall short, and lists them: risk-rate every system, give agents only the tools they need, train people in AI literacy, keep a human able to stop the system, keep an inventory, log everything, and test continuously. Suggests starting with an inventory of every AI tool in use.
Why it matters: The four questions map onto this toolkit: the law (Resources, tier 0), proof (evidence pack builder, TK-04), accountability (the accountable owner and H4 to H5 review) and production (agent registry, TK-03, and incident reporting, TK-12).
Caution: A practitioner's opinion post that also promotes the author's newsletter. The legal points (for example AI literacy as an EU duty) and the standards named should be checked against the primary texts before citing.
Source on LinkedIn (opens an external website) · evidence: post read in full in the owner's signed-in browser on 17 November 2026
Credit belongs to the named authors. To correct or remove an item, use the contact routes below.
No. Declare material assistance — where AI shaped the substance, structure, analysis or code of an output. Trivial uses such as spell-checking do not need a declaration, but your local policy may set a lower threshold.
No. AI can suggest, draft, check and explain. Approval is a human act by someone with the authority to make that decision. AI never grants authority a user does not already hold.
Treat it as an incident. Do not publish. Verify every citation against the original, log the failure in the drift and incident register, and check whether the same prompt pattern has been used elsewhere.
No. This toolkit helps you prepare and evidence those assessments; it does not replace them or the professionals accountable for them.
Not unless the tool is explicitly approved for that classification and the data-handling terms have been assured. When in doubt, do not paste — ask your information asset owner.
Treat it as untrusted third-party code: review it, test it, scan dependencies, record provenance and never merge without a human reviewer who understands it.
The same ones that apply to everything else. WCAG 2.2 AA, plain language, alternative formats and the accessibility statement duty all apply regardless of who or what drafted the content.
Your service owner and the named AI assurance contact for your organisation. Report near-misses too — they are the cheapest form of learning.
A guidance utility that cannot be questioned is not trustworthy. Here is how this one is governed, and how you hold it to account.
Every substantive claim links to a tiered source. Where this portal expresses an opinion or a house rule rather than a legal requirement, it says so.
Owner: AI Capability Governance working group
Version: 3.3.0
Last reviewed: 2026-11-18
Next review due: 2027-02-13
Open Government Licence v3.0 (content) / MIT (code). Fork it, adapt it for your organisation, and tell us what you changed so others benefit.
This site is static and client-side. It sets no tracking cookies, runs no analytics that identify you, and sends nothing you type in the tools to any server. Preferences are stored locally in your browser and you can clear them at any time.
This is a concept portal, not a statutory instrument. It does not replace your organisation's policies, your DPO, your security team or your legal advice. Some linked standards sit behind paywalls.
Tell us about a broken link, an out-of-date claim, an accessibility barrier or a gap. Accessibility issues get a response within five working days.
| Version | Date | Change |
|---|---|---|
| 3.3.0 | 2026-11-18 | One portal. The playbook moved into this page: one masthead with grouped menus, one contents drawer for the whole portal with read marks, one glossary (39 playbook terms merged in, now 62) and one footer. Redesigned after the original AI Capability Governance portal: Figtree typeface (self-hosted), navy front page with the orbit, the guidance-in-force ticker, tinted part cards and the chapter view with part chips and progress bars. playbook.html now redirects here. New check S25. |
| 3.2.0 | 2026-11-18 | Added the agent governance playbook: 77 chapters in 12 parts ported from an internal AI agent governance toolkit, with a working “Can I use AI for this?” route tool, a knowledge map and a SKOS taxonomy. Ten chapters are withheld, each with its reason. New checks S21 to S24. Fixed the masthead line that repeated the old name. |
| 3.1.0 | 2026-11-13 | Content review: replaced the withdrawn Generative AI Framework for HMG with the AI Playbook, added the Data (Use and Access) Act 2025 and the ICO's automated decision-making guidance, flagged ICO guidance now under review, and fixed four moved links. Added a field to the evidence pack for the data definitions an agent's answer relied on, and signals S22 to S28. |
| 3.0.1 | 2026-11-08 | Replaced one broken and four moved resource links after an automated check of all 68; 11 sites refuse automated checks and are still to be checked by hand. The Monday health run now checks every link. |
| 3.0.0 | 2026-11-08 | Renamed to AI Capability Governance and Implementation and moved to its own address. Added the Signals section (a weekly horizon scan of AI and data governance with attribution, evidence labels and an open JSON register). Added Beginner, Intermediate and Advanced levels to every persona, journey and common task, and a new data and AI governance lead persona. Contact is now by the GitHub contact form only: the email-client route and reply-address field are removed, and the two email addresses in the worked metadata example are replaced with a role and the form. Added the issue forms the contact links already pointed to, and automated checks, including browser checks, that run before anything is published. Fixed the page scrolling sideways on small phones. |
| 2.0.0 | 2026-08-08 | Added task-first entry points, six interactive tools, a tiered resource library, H0–H5 review table, RACI, transparency and redress commitments, MoSCoW roadmap, glossary and FAQ. Added dark, high-contrast and large-text modes, site search, on-page navigation and a print stylesheet. |
| 1.0.0 | Initial | Concept portal: mission, knowledge corpus, toolkit, assurance workflow, catalogue, UCD and code of practice. |
An honest backlog is a trust signal. These are the gaps identified in this portal and in most public sector AI guidance, ordered by whether they are essential, valuable, desirable or deliberately out of scope.
Nothing on this page is credible without a person who owns it and answers questions.
Legally required for UK public sector sites; also the fastest trust signal.
Freshness is the main failure mode of guidance portals.
Without a schema, registries cannot federate and evidence cannot be automated.
The core control that keeps accountability with people.
Prevents hallucination becoming policy.
Equality duty is a due-regard duty — it must happen before decisions, not after.
Findability is the top usability complaint for guidance sites.
Users arrive with a task, not an interest in your information architecture.
Supports assisted digital, low bandwidth and audit contexts.
Abstract principles do not change behaviour; examples do.
Lets teams re-check work when guidance changes.
Statutory in some contexts and essential for inclusion in all.
Broken authority links silently destroy trust.
Enables reuse by other departments and tooling.
Helps organisations see where they are without a consultant.
Scales the corpus beyond a single team.
Reduces the "which tool may I use?" support burden.
Makes proportionality measurable rather than rhetorical.
Out of scope by design — approval stays human.
The portal is public and stateless; sensitive records belong in assured systems.
Conflicts with interoperability and exit obligations.
The toolkit supports specialists; it does not substitute for them.
Every route below goes through GitHub, so it is public and auditable. There is no email route. Choose the one that fits — if you are not sure, raise an issue and we will redirect it.
Broken links, tools that do not work, layout failures, incorrect calculations or console errors.
Anything that stops you using this site with a screen reader, keyboard, magnification, voice control or any other assistive technology. These are treated as priority defects.
Target first response: 5 working days.
A tool you need, a section that is missing, or an authoritative source that should be in the library.
Superseded guidance, a withdrawn standard, a mis-stated legal duty or a dead authority link. Include the source that proves it.
Do not open a public issue. Use private disclosure so we can fix it before it is known. See our security policy for scope and timescales.
Not sure it is a defect? Want to talk about approach, reuse this in your organisation, or contribute? Start there.
This form helps you write a clear report, then opens the GitHub contact form with it filled in. Issues are public: do not include personal data, case details or anything above OFFICIAL. For a security problem, use private reporting instead.
Start with the corpus, encode the metadata, publish the skills, bound the agents, enforce the review — and keep improving it in the open.