Presenting an AI decision to the board: what boards ask

The board has asked for an AI strategy, or a budget line needs approval, or a vendor contract is waiting for a signature. Sometimes a tool a team built for itself is about to become something the whole business relies on. The decision has your name on it and the meeting has a date. Most of the published guidance on what boards ask about AI is written for the directors asking. This article is written for the person answering.

Directors cannot check the model themselves, so they will check whether you can show where its output will be relied on, who checks that output, and what happens when it is wrong. A paper that answers those three points before the meeting gets a different discussion from one that leaves the board to dig for them.

Directors know they cannot check the model, so they will check the presenter

Directors are candid about what they can and cannot oversee. In PwC's 2026 Annual Corporate Directors Survey of 599 directors, reported by ESG Dive, 71% said their boards need to improve their AI skills to oversee the technology. They are also holding each other to a higher standard. The share of directors who see a colleague's insufficient expertise as grounds for replacement nearly doubled, to 39% from 21%.

EY's analysis of Fortune 100 disclosures, published by the Harvard Law School Forum on Corporate Governance, shows both sides of this. Only 12% of those companies disclosed that board members had received AI training, while 22% flag AI hallucinations, inaccuracies or bias as a material risk. Boards are naming AI output errors as a risk faster than they are building the skills to judge them.

The person presenting has to plan around that. A director who knows they cannot assess a model will not try to. They will assess what they can: whether the presenter understands where the system's output goes, whether someone named is accountable for it, and whether the paper reads as a record that will hold up later. Some boards now include a director with AI expertise, but most of the room will still judge the decision by the quality of the paper and of the answers.

This works in the presenter's favour. You do not need to make the board technical. You need a paper that shows the decision has been examined by people who could examine it, with the results written down.

Most of the board's questions are predictable, and few are technical

Directors work from published question lists, and those lists are long. NACD's boardroom tool from March 2025 runs to around fifty questions. Many cover strategy, cost and governance structure, and the ones with teeth are about operations: where the data came from, where humans review output, whether red teams have tested generative AI uses, and whether testing and monitoring are, in NACD's words, "reflected in our logging" rather than held in someone's memory. EY's Forum piece offers eight questions for boards, one of which asks what mechanisms exist to catch AI-generated errors.

Read together, these lists show that a board rarely asks how the model works. It asks how the company would know the model was wrong, and who would find out first. A presenter who has read the lists can answer most of them in the paper before the meeting starts.

UK boards have a further reference point. The FRC's Code Guidance asks under strategy, "Are we aware of emerging technologies (e.g. Responsible Artificial Intelligence) being used by the company, for example, in reporting?" and lists "new technologies (e.g. artificial intelligence)" among technology risks. Provision 29 of the UK Corporate Governance Code, the board's declaration on the effectiveness of material controls, applies to financial years beginning on or after 1 January 2026. Where AI sits inside a material process, the controls around it fall within the scope of that declaration. The Code applies to companies in the commercial companies and closed-ended investment funds listing categories; for a private or PE-backed company it is a reference standard rather than a duty, though many boards hold themselves to it anyway.

Directors who use these lists are preparing to approve something they cannot inspect, and that is a reasonable way to prepare. The presenter's job is to make the answers easy to find.

Can you trace, defend and explain the output you will rely on?

Prof. Joe O'Mahoney put one question to consultancy boards on LinkedIn on 30 April 2026, the eighth of eight: "Can we trace, defend and explain every AI-generated output we put in front of a client?" He wrote it for consultancies, but it applies to any company whose AI output reaches a customer, a lender, a regulator or the accounts. Most of the questions on the longer lists reduce to this one, and each verb tells the presenter what to have ready.

Trace: which system, which inputs, whose sign-off

For any output the business will rely on, the paper should say which system produced it and which version was running, what inputs it used, and who signed off the configuration that went live. If a vendor updates the model underneath, the paper should say how the company will find out. A board does not need the technical detail, but it needs to know that someone can reconstruct any given output without guessing.

Defend: what was checked before anyone relied on it

Defending an output means showing what was checked before it was used, who did the checking, and whether that check is logged. A check that happens but leaves no record cannot be defended to an auditor or a regulator, and NACD's point about logging is aimed squarely at this gap. Where the honest answer is that nobody checks output before it reaches a customer, the paper should say so and explain why that is acceptable for this use.

Explain: an account that works without a data scientist in the room

The board, and later a customer or regulator, will want to know in plain words why the system produced what it did. The paper should contain that account, written so that the presenter can give it alone. If the explanation only holds up with the data science team in the room, the company cannot yet explain the output, and the board will notice.

What the board paper should contain before anyone asks

A paper built around trace, defend and explain covers most of what the published lists ask. Five items do the most work, and for each one there is an answer that will worry the board.

Where AI touches something that matters, with an owner for each. List every point where AI output will affect a number in the accounts, a customer decision or a regulatory filing, and name the person accountable at each point. The answer that worries a board is a list without names, or a team named as owner instead of a person.

How errors will be found, and how long that takes. State how a wrong output would come to light: a review step, a reconciliation, a customer complaint. Give a realistic time to detection. If errors would first surface through a complaint or an audit finding, say so plainly, because the board will work it out in any case. Directors who have read the questions a director asks of a live system's performance will push on this item.

What was asked of the vendor, and the terms accepted. Research by MIT Sloan Management Review and BCG, from a 2023 survey of more than 1,240 executives, found that more than half (55%) of all AI failures come from third-party tools. The paper should record what was asked of the vendor before signing and the data retention, model-change and exit terms the company accepted. The worrying answer is that the standard contract was signed and nobody read the model-change clause.

What it costs to stop or reverse, and when. Set out the cost of unwinding the decision at each stage: after the pilot, after the contract renews, after a process has been redesigned around the system. Boards approve reversible decisions more readily, and they want to know the point after which a decision stops being reversible. A paper that does not mention exit leaves the board assuming there is none.

Who outside the team has checked the answer. The team that built or bought a system is the one least able to judge it neutrally, which is no criticism of the team. Say who outside it has reviewed the case, the risks and the claims made about performance. "Nobody yet" is a common answer, and it is often the one that turns an approval into a deferral.

The paper sits within whatever wider governance the company runs, and if it has adopted a governance framework, the paper should map onto it. The record matters as much as the paper. The minutes should show the question asked, the answer given, the owner named and the date, so that the board can later show it asked before the company relied on the system.

FAQ

What will the board ask when I present an AI investment? Boards ask mostly about where AI output will be relied on, who checks it, how errors are caught and what the vendor terms say, rather than about how the model works. Agathon reads published director question lists, including NACD's and EY's, as a guide to the paper an executive should write, because most of their questions can be answered in advance. The questions that carry the most weight concern logging, human review and the cost of reversing the decision.

What should a board paper on an AI decision include? A board paper on an AI decision should name each point where AI touches the accounts, a customer decision or a regulatory filing, with an owner for each, and set out how errors will be found and how quickly. Agathon also advises recording what was asked of the vendor, the data retention, model-change and exit terms accepted, the cost of stopping at each stage, and who outside the team has checked the case. The minutes should then record the question, the answer, the owner and the date.

How do I explain AI risk to a board that is not technical? AI risk is easiest to explain to a non-technical board as three practical questions: can the company trace an output to its system, inputs and sign-off, defend the checks made before relying on it, and explain in plain words why it was produced. Agathon recommends writing that plain-words account so the presenter can give it without a data scientist in the room. Where an honest answer is weak, the paper should say so and explain why the risk is acceptable for this use.

Does the UK Corporate Governance Code require boards to oversee AI? The UK Corporate Governance Code's provisions do not name AI, and the FRC's Code Guidance, which mentions artificial intelligence under strategy and technology risk, is not mandatory. Provision 29 requires a declaration on the effectiveness of material controls for financial years beginning on or after 1 January 2026, so where AI sits inside a material process, its controls fall within that declaration's scope. The Code applies to companies in the commercial companies and closed-ended investment funds listing categories, and for private or PE-backed companies it is a reference standard rather than a duty.

References

An AI decision going to the board with your name on it
The article's last checklist item asks who outside the team that built or bought the system has checked the answer, and "nobody yet" is the reply that often turns an approval into a deferral.

Agathon's AI leadership advisory gives executives who have to stand behind an AI strategy, a budget line, a vendor contract or a tool the team built a clear, independent read on the decision before it reaches the board; Agathon does not sell implementation or remediation on anything it assesses, and has no vendor affiliation.

Send an email if you are weighing how the checklist applies to a paper you have not yet written, and Agathon will tell you quickly if it is not a fit.
Book a short confidential call if the board date is set and the decision is on the table.

Email us Book an initial consultation