Board AI policy and disclosure
- Chair and board
- A · Accountable
- Company secretary
- R · Responsible
- Solutions architect
- C · Consulted
- Technical PM
- I · Informed
- Engineers
- I · Informed
- General counsel & DPO
- C · Consulted
- Security
- C · Consulted
Board packs that are checked, consistent and easier to read, with every summary traced to its source. Authors and the company secretary sign off, and directors still do the judging.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
A board pack is assembled from papers written by a dozen or more people: the CFO's finance report, divisional updates, risk and assurance reports, committee chairs' reports, and papers asking the board to approve something. Each author writes in their own style and to their own length. The company secretariat has a few days between the paper deadline and issue to chase late papers, check that each one says what the board is being asked to do, and spot where figures don't match.
In a typical listed group, that checking is done by hand by two or three people. A revenue figure in the CEO's report differs from the finance appendix because one was updated and the other wasn't. An action from three meetings ago quietly drops off the log. A paper runs to forty pages without saying, until page thirty, what decision it needs. Directors receive the pack late, read it at weekends, and good challenge suffers.
The stakes are not just efficiency. Directors' duties to exercise independent judgement and reasonable care depend on the information they're given. The FRC's Code Guidance lists poor quality papers and insufficient notice among the risk factors for poor decisions. Papers also carry some of the most sensitive information an organisation holds: inside information, legal advice, remuneration and personnel cases. Any tool that touches them has to be at least as well controlled as the board portal itself.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Days between the meeting and the pack reaching directors. Most governance guidance asks for papers well ahead of the meeting.
ExampleAt least 7 days before every main board meeting
Figure mismatches and broken cross-references found by the checks and fixed before the pack goes out, against those found after.
ExampleNearly all caught before issue
Share of papers whose first page states the decision or input needed from the board.
ExampleEvery paper
How much authors change the draft summaries. Very high means it isn't helping; near zero can mean drafts are being signed unread.
ExampleStable, and reviewed each cycle
Restricted papers or extracts found in the shared index by the monthly access test.
ExampleZero
Measured through the annual board evaluation, not assumed.
ExampleImproving year on year
Cross-sector rules are on the UK regulation map.
Directors must exercise independent judgement (s.173) and reasonable care, skill and diligence (s.174), and have regard to the matters in s.172 when promoting the company's success. A tool can help them read the pack; it can't do the judging for them. Summaries point back to the paper and never replace it. Nothing the AI drafts is written as if it were management's recommendation or the board's view.
Minutes of all board meetings must be recorded and kept for at least ten years, and a minute signed by the chair is evidence of the proceedings. The minute has to reflect the real deliberation. The AI may turn the company secretary's own notes into a draft outline, but the company secretary writes the record of decisions, conflicts and challenge, and the chair approves it.
The guidance asks for board papers that are accurate, clear, comprehensive and up to date, with a summary and a statement of what's expected of directors. Under the chair's direction, the company secretary is responsible for good information flows. Provision 29 applies to financial years beginning on or after 1 January 2026: the board declares the effectiveness of its material controls at the balance sheet date. If the board relies on the pack to monitor those controls, document and test the checks like any other control.
Large private companies, charities, NHS trusts and public bodies report against different codes, but all of them expect boards to receive good information and to own their decisions. The design doesn't change. The disclosure to directors and the board's AI policy are worded for the code you follow.
For listed companies, results, trading updates and transaction papers often hold inside information. Those papers go only to the restricted workspace, open to people already on the insider list, or stay out. The provider contract rules out model training and limits retention, and every access is logged so it can be checked against the insider list. In November 2026 the FCA's review of sponsors' use of AI noted approved enterprise tools, human review of every output, and close attention to confidentiality.
Privilege depends on confidentiality, and a summary of legal advice can disclose its substance. Legal advice papers stay out of the shared index, and the tool doesn't summarise them unless general counsel agrees and the workspace is confined. A 2026 Upper Tribunal decision said that putting confidential documents into a public AI tool amounts to placing them in the public domain. Enterprise tools need checking too: what the provider keeps, and who can see it.
Remuneration papers, HR cases, whistleblowing reports and safety incidents hold personal data, sometimes special category. That means a DPIA before the pilot, collecting only what's needed, processing in UK regions, retention that matches the board record, and a contract that stops providers training on your data. Personnel cases stay out of the shared workspace.
The IoD's business paper sets out governance questions for directors on AI. It's a useful starting checklist for the board's own AI policy, including how the board itself uses AI to prepare for meetings, alongside the company's own legal advice.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake: papers, appendices, source reportsPapers come from your board portal's export or API, after the author marks them final. Authors don't upload anything new. | API Gateway, Amazon S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| Reading papers and tablesTest on real packs: scanned signed papers, landscape financial tables and charts, not clean Word files. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Drafting summaries and promptsBoth clouds offer a choice of models, including Claude. Every sentence carries a citation, or it isn't shown. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| RetrievalOne index per meeting and committee, filtered by the reader's portal permissions. Previous minutes and the actions log are included; restricted papers are not. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| Figure and cross-reference checksPlain code compares each quoted figure with the cell it should come from. A model can find the figures; it never decides whether they match. | AWS Lambda | Azure Functions | Containers on Kubernetes |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| WorkflowTracks each paper from final draft to author sign-off to company secretary approval. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Restricted workspaceFor inside information, M&A, legal advice and personnel papers. Named users only, matched to the insider list or the deal team. | A separate AWS account with its own keys and index | A separate Azure subscription with its own keys and index | A separate namespace on your own GPUs, or no AI at all |
| GuardrailsBlocks recommendation language ("the board should approve") and uncited claims in drafts. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionPermissions come from the board portal and are enforced by the platform. Check each provider's retention and abuse-monitoring settings for prompts. | IAM Identity Center, KMS with customer-managed keys, Macie | Entra ID, Key Vault, Purview sensitivity labels | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every draft, edit, sign-off and access, with the paper version used. The access log is reconciled with insider lists. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
Your board portal, document management system, minute book and insider lists stay where they are and remain the record. So do the finance and risk systems the source figures come from. Restricted papers can stay entirely on your own servers.
Uses cloud AI
Reading, drafting, retrieval and the checks run in a UK cloud region, for papers the company secretary has released to the shared workspace. Organisations that won't send any board material to a cloud provider can run the same design on their own GPUs with open-weight models.
How they connect
Through the board portal's export or API and an API gateway, over a private link such as AWS Direct Connect or Azure ExpressRoute. Papers flow one way into the workspace; drafts and check results flow back to the secretariat for review, never straight into the portal.
If the link drops
The pack is prepared and issued exactly as it was before the AI, by the secretariat, on the same timetable. Checks and summaries run when the link returns, if there's still time before issue.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Paper assistant
Where a person signs offThe paper author edits and signs their summary, and the company secretary approves the pack before issue. AI-assisted parts are labelled for directors. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Minutes and actions assistant
Where a person signs offThe company secretary writes and owns the record of decisions, and the chair approves the minutes. Action owners close their own actions.
Your board portal stays the single place directors read papers. The AI works on an export of final papers, through the portal's API or a controlled file drop, and sends drafts and check results back to the secretariat. It never writes into the portal itself. That keeps the board portal's access controls, version history and audit trail exactly as they are.
Permissions follow the portal. Each paper keeps its audience, such as the full board, a committee, or a named group, and the index filters every query by it. A director on the audit committee sees prompts for audit papers; someone outside it doesn't. Restricted papers never reach the shared index at all, so a mistake in filtering can't expose them.
Figure checks use the source, not another summary. Finance and risk reports are pulled from the systems that produced them, or attached as appendices, and the check compares like with like. If a source isn't available, the figure is marked unchecked rather than assumed correct.
Map today's pack timetable, measure the baseline, classify papers by sensitivity, and agree what your board portal's export or API allows.
Gate · must be true to passBaseline agreed, classification rules signed off by the company secretary and general counsel, DPIA started, board's AI policy drafted.
One committee whose papers carry no inside information, such as a sustainability or health and safety committee, over two cycles in shadow mode: drafts and checks are produced, the pack is prepared as normal, and the two are compared.
Gate · must be true to passSummary accuracy and edit rate within agreed limits, no uncited claims shown, no restricted material in the shared index, committee chair content.
Summaries, checks and question prompts go live for the pilot committee and the main board's open papers, with monitoring and a rollback plan.
Gate · must be true to passBoard approves the AI policy and the disclosure wording; monitoring, support and rollback in place.
Add committees one at a time. The audit and remuneration committees come last, with the restricted workspace and insider-list controls tested first.
Gate · must be true to passEach committee passes the same pilot gate, and its chair agrees.
| Activity | Chair and board | Company secretary | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | General counsel & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|
| Board AI policy and disclosure | Accountable | Responsible | Consulted | Informed | Informed | Consulted | Consulted |
| Paper classification and restricted handling | Informed | Accountable | Consulted | Informed | Informed | Responsible | Consulted |
| Target architecture and portal integration | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted |
| DPIA, privilege and MAR review | Informed | Consulted | Consulted | Informed | Informed | Accountable and Responsible | Consulted |
| Build and testing on real packs | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Consulted |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Responsible | Consulted | Responsible |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the workspaces and how papers are classified into them, the integration with your board portal, the checks and the gates, working closely with your company secretary and general counsel. I'd report to whoever sponsors the programme, usually the company secretary or the CFO.
Option B
If your partner or your own team already has an architect, I'd take on the retrieval, the figure checks and the restricted workspace under their lead, and give them a second senior opinion on access control, provider settings and the audit trail.
Goes wrong
The cover summary reads like a recommendation, and directors can't tell management's view from the tool's.
How the design avoids it
Summaries only restate the paper, with a citation for every sentence. Guardrails block recommendation language, and the author signs the summary as their own.
Goes wrong
A figure in the summary doesn't match the appendix, because the model rounded or restated it.
How the design avoids it
Numbers are copied by code from the source and checked by code. The model never generates a figure, and an unchecked figure is marked as such.
Goes wrong
A remuneration or M&A paper ends up in the shared index, and someone outside the committee can find it.
How the design avoids it
Restricted papers go to a separate workspace or stay out. Permissions come from the portal, default to deny, and are tested every month.
Goes wrong
Directors read the summaries and stop reading the papers.
How the design avoids it
Summaries are short and point to pages. The chair's question prompts invite challenge, and the board evaluation asks how directors use them.
Goes wrong
Meetings are recorded and transcribed "to help the minutes", creating a verbatim record nobody planned for.
How the design avoids it
No recording by default. The company secretary writes the minute from their own notes, and any change is the board's decision, with legal advice.
Changes already announced or under way that affect this design. More in the UK horizon scan.
If the board relies on the pack to monitor material controls, write the figure checks and sign-offs down as part of that control framework, with an owner and test evidence.
Keep the checks configurable, so new or removed reporting requirements change a rule set, not the design.
Your board portal provider and AI provider may come into scope. Update supplier assurance and contracts as the duties are set.
Board minutes are kept for at least ten years and papers are highly sensitive. Record which encryption the workspaces use and keep it easy to change.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.