Ten Questions Every UK Board Should Ask About AI

How a board can tell whether it is governing its AI or only hoping

James A Lang · 26 October 2026 · Board briefing

Every UK board is now accountable for AI decisions it did not build. A pricing model written by a supplier, a chatbot configured by marketing, an agent a developer switched on last quarter: each one acts in the company's name, and each one lands on the directors' duty of care.

Most boards respond in one of two ways. Some delegate the subject to the technology function and wait for a report. Others commission a policy and treat the signature as the control. Neither tells the board whether its AI is governed. Both tell the board only that someone has been asked.

The gap is not knowledge of the technology. Directors do not need to understand how a model is trained. They need to know which questions to ask, what a good answer looks like, and which law sits behind the answer. That is what the ten questions below provide.

Watch, control and stop

One sentence sits behind all ten questions: we will not delegate a decision to AI that we cannot watch, control and stop.

I learned the shape of that sentence in senior commanders' briefing rooms long before AI arrived. A commander delegates authority, never accountability. Delegation is safe only when the commander can see what is happening, can change it, and can halt it. Remove any one of the three and delegation becomes abdication.

The same test applies to a model in production. Watch means the board receives evidence, not assurance. Control means a named person can change the system's limits. Stop means the system can be switched off quickly, and someone has proved it. A board that holds all three can afford to experiment. A board that holds none is relying on luck.

Purpose and accountability

The questions follow the categories of the AI Blindspot Framework, starting with purpose and ending with the people affected. The order is deliberate. Strategy and accountability come before technology, because a well-engineered system serving no clear purpose is still a governance failure.

1. What is each material AI use for, what value does it deliver, and how will we know it is working? AI adopted without a declared purpose cannot be judged, so it cannot be governed. Ask for a register of material uses, each with its value stated in money, time or learning, and a measure of success the board reviews. The anchor is Principle A of the UK Corporate Governance Code: the board's role is long-term sustainable success.

2. Who is accountable, by name, for each AI system, and has the board approved a policy and risk appetite that lists the uses we will not allow? A team is not an owner. Ask for a named owner per system and a board-approved risk appetite with prohibited uses written down. Directors' duty of care under section 174 of the Companies Act 2006, Provision 28 of the Code and ISO/IEC 42001 all assume someone answers for each decision.

3. Do directors and staff understand AI well enough to question and override it, and is human oversight real rather than nominal? A human who never overrides the machine is a signature, not a safeguard. Ask for board training records and a log of overrides that are made and reviewed. UK GDPR Article 22 governs solely automated decisions; for EU markets, Articles 4 and 14 of the EU AI Act require literacy and effective oversight.

The system itself

4. Do we know where our data and models come from: their quality, lineage, lawful basis and provenance? A model inherits every flaw in its data and every term in its supplier's contract. Ask for lineage records, a lawful basis for each data source and written assurances from third-party model suppliers, measured against UK GDPR and the Data (Use and Access) Act 2025.

5. Are our models tested before release and monitored for accuracy and drift once they are live? Models degrade quietly as the world they were trained on moves. Ask for pre-release test results and live monitoring against agreed thresholds, structured on the NIST AI Risk Management Framework and ISO/IEC 42001.

6. Are our AI systems protected against poisoning, prompt injection and data leakage, and are they securely retired at end of life? AI creates attack surfaces that conventional security testing misses. Ask for AI-specific threat models, red-team results and a decommissioning record for every retired system. The anchors are the DSIT Code of Practice for the Cyber Security of AI and ETSI EN 304 223.

Operations and exposure

7. Do we hold a complete AI inventory, including shadow AI that staff use without approval, with an owner and an incident process for each system? A board cannot govern what it cannot see, and staff adopt AI tools far faster than procurement approves them. Ask for an inventory reconciled against discovery tools and an AI incident log the board sees, in line with ISO/IEC 42001.

8. How dependent are we on outside AI providers, on rules that may change and on energy costs, and do we have an exit plan? Concentration on a single provider is a strategic risk, not a technical one. Ask for a supplier dependency map, tested exit plans, a regulatory horizon scan and energy cost data for AI workloads. The Procurement Act 2023 applies, and the EU AI Act for EU markets.

9. Which of our systems can act on their own, who set their limits, and can we watch, control and stop them immediately? This is the controlling idea in its sharpest form. An agent that can send, buy or change things acts at machine speed. Ask, for each agent, for its permissions, its oversight model, tamper-proof logs and a shutdown that has been tested, not merely designed. The anchor is the NCSC's August 2026 guidance on managing the cyber risk of agentic AI.

The people affected

10. Can we show that our AI produces fair, accurate and explainable outcomes, including what our chatbots tell customers, and can the people affected challenge them? A chatbot's promise is the company's promise. Ask for fairness test results, explanation records and a working route to contest an AI decision. The Equality Act 2010, UK GDPR Article 22 and the Digital Markets, Competition and Consumers Act 2024 all apply.

Score with evidence, not confidence

A board that feels sure about its AI and a board that can prove it should not receive the same score. So the method starts from what the board believes, then lets each answer move that belief.

Prior. Before answering, each director gives a starting score from 0 to 1 and a confidence: low (weight 2), medium (5) or high (10).

Evidence. Each question is answered from what has been seen. Yes, evidenced in board papers, data or audit, counts as a success. Partly, or asserted but not seen, counts half to each side. No counts as a failure.

Don't know. This changes nothing in the score. It is logged as an information gap with an owner, a deadline and a default action. Uncertainty is a task to close, never a low score.

Posterior. The updated score is the mean of a Beta distribution: (a + Y + P/2) divided by (a + b + Y + P + N), where a is the weight multiplied by the prior, b is the weight multiplied by one minus the prior, and Y, P and N are the counts of Yes, Partly and No. It is reported with a credible range holding 0.8 of the probability.

A worked example. A board starts at 0.7 with medium confidence, giving a = 3.5 and b = 1.5. It answers four Yes, three Partly, two No and one Don't know. The score is 9 divided by 14: 0.64, with a credible range of 0.48 to 0.80, and one gap to close.

The board felt like a 0.7. The evidence says 0.64, and it could be as low as 0.48. That difference is the conversation the board needs to have. A wide range shows where the board lacks information. Scores far apart across directors show where they disagree. Both are more useful than a single confident number.

What to do at the next board meeting

Start small and start with evidence. Put the ten questions on the agenda as a single item, not a programme. The printable one-page version is built for the board pack.

Before the meeting, ask each director for a prior score and a confidence level. In the meeting, answer each question with what the board has actually seen, and treat every assertion as Partly until the evidence is tabled. After the meeting, give each Don't know an owner and a deadline, and score again next quarter.

The first score matters less than the second. A board that moves from 0.64 to 0.75 in a quarter, and narrows its range, is governing its AI. A board that cannot run the exercise at all has learned something more important.

None of this asks a board to be certain. It asks a board to know where it stands, to close the gaps it can, and to keep the ability to stop what it cannot yet trust. Resilience, in the end, is what buys you the right to be wrong.

Legal references reflect UK law as at October 2026 and are not legal advice.