On 7 October, a 35-page briefing paper appeared on the websites of two Brussels organisations that most UK directors could not name, and within days the Institute of Directors (IoD) had put its Director General’s signature on the foreword. AI governance: ten principles for effective board oversight was written jointly by Accountancy Europe, which represents 49 professional bodies and close to a million accountants, and ecoDa, the confederation of Europe’s national institutes of directors, whose members encompass around 55,000 board members in 26 countries. The IoD is a member of ecoDa and endorsed the paper as a practical framework for boards. I have read it twice, once as a director and once as someone who will soon be teaching it, and I have yet to meet a board that has read it. The paper isn’t perfect, and I will say where it is thin, but it is now the most authoritative description a European board can point to of what adequate AI oversight looks like, and regulators, investors, and litigants are starting to test that adequacy.
A standard with 55,000 directors behind it
Provenance matters here more than it usually does, and the mechanism is liability. The paper makes this point in a short box on director liability: under the NIS2 Directive, management bodies may face personal liability for failures in cybersecurity oversight, and the EU AI Act raises expectations for boards’ risk management, governance, and accountability. Its authors conclude that regulatory scrutiny in Europe, and potentially beyond, is moving from organisational failures to the adequacy of board oversight. When adequacy is tested, it is tested against something: a code, a guidance note, a professional body’s published view of good practice. Until this month, no European document on AI oversight carried the weight of both the directors’ own institutes and the accountancy profession. Now there is, and a board that has never seen it will find an auditor, an evaluator, or a claimant’s counsel quoting it back to them. The sensible response is to read it first, which takes about an hour, and to decide what it asks of your board before someone else decides for you.
What it gets right that most board guidance does not
Four things earn the paper its place in the board pack. The first is its refusal to build a new structure. Principle 3 says governance of AI should be embedded in the enterprise governance the board already runs, with three lines of defence and existing committees adapted where AI needs specific safeguards, and it tells audit and risk committees to check whether controls have been changed or merely relabelled from pre-AI frameworks. That is the right, and unfashionable, call, since a new AI committee is easier to announce than a changed control. The second is proportionality. Principle 6 asks the board to set an AI risk appetite with explicit red lines and gives it a four-tier risk radar, from low-impact productivity tools through to a fourth tier labelled “Unacceptable”, whose governance response is a single word: prohibited.
The third is that the paper turns the lens on the board itself. Principle 10 covers AI in preparing board papers, uploading board material to unapproved tools, and directors’ obligation to know when AI has materially shaped what they read, which is also the ground covered by my own principles for responsible AI use by directors. The fourth is unmistakably Accountancy Europe’s fingerprint: Principle 8 asks audit committees how the external auditor is using AI in the audit, in its quality management, and in determining its own independence, a question I suspect almost no UK audit committee has yet put to its auditor in writing. Underneath all four sits the paper’s best sentence: that poor governance with AI is often simply poor governance at scale, which is the organising idea of the whole document and the reason its principles look so familiar.
Where it is thin
Three limits are worth naming so that a board does not lean on the paper for more than it can bear. Agentic AI appears in the risk taxonomy and is nowhere developed, which is an odd gap in a month when the ICO has opened a call for evidence on agents and made enquiries of OpenAI, Anthropic, and Meta, and the US Federal Trade Commission has confirmed an investigation into the same companies over agents that exceeded their instructions. The paper contains no named-company cases, so it asserts its arguments and illustrates them with a stylised retail-bank example. The survey published alongside it, which I draw on below, rests on 87 European directors and says of itself that its findings are indicative and should not inform major decisions. There is also a production error: the heading of Principle 10 repeats the title of Principle 8, so a board quoting the principle titles verbatim should take the tenth from the body text, which concerns the board’s own use of AI. None of this reduces the paper’s usefulness, only its reach.
Three pages to take into the next meeting
Thirty-five pages and more than seventy questions is more than any board can act on in a cycle, and a board that tries will end up with a long discussion and nothing minuted. The paper’s authors anticipate this, which is why each principle includes its own short list of questions for management, but the lists still add up to a year’s agenda, not a meeting. My suggestion is to take three pages into the next meeting and to minute one output from each. I chose these pages because each converts directly into a decision the board can record, and because the survey published alongside the paper shows these are the three places most boards currently have nothing on file.
Page 9: the investment questions
Page 9 lists eight questions on investment discipline, and two of them are the ones most boards cannot answer. Question 5 asks when an AI initiative will be paused, redesigned, or stopped if it fails to deliver, and how the board will know. Question 8 asks what the organisation is explicitly choosing not to do with AI, and whether the executive level understands and enforces that boundary. The survey suggests how rare the answers are: 12 per cent of respondents review AI investments against defined performance indicators and business outcomes, and 53 per cent rely on informal, ad hoc, or no structured review at all. EY’s CEO Outlook, published on 1 October and based on responses from 1,200 chief executives in 21 countries, found the same thing from the other side of the table, with 16 per cent reporting clear, real-time visibility into the return on their AI investment. The output to minute is a stop rule for each material AI investment: the metric, the review date, and the condition under which the money stops. A board that minutes that has done most of what Principle 2 asks.
Page 19: the risk radar
Page 19 draws the four-tier risk radar and, in doing so, hands the board a template it has probably never filled in. The “Unacceptable” column is empty in most organisations, and writing it is the cheapest control the paper offers, because a prohibited-use list costs nothing to maintain and settles a dozen future arguments in advance. Sophisticated buyers already write them: in September, Reuters reported, citing The Information, that Palantir had pressed Anthropic for irrevocable zero-data-retention guarantees before offering its models through Palantir’s software, and that Booz Allen Hamilton had barred staff from using the model for proprietary cybersecurity work. The same page should prompt a second question the paper raises in its annex on escalation triggers: “breakdown of safeguards or loss of effective human oversight” is listed as an event that should reach the board. Escalation after the event is only useful if someone has standing authority to halt the system before it happens.
The reason to settle that now is tempo. On 7 October, CrowdStrike reported that an unnamed, probably Chinese-speaking, financially motivated actor had used ARTEX, an open-source agentic penetration-testing tool, against South Korean financial institutions between late September and early October; at least nine banks have been reported as targeted, authorities have confirmed intrusions at seven, and Shinhan Bank said the data of about 25,000 customers was taken. CrowdStrike does not claim one actor hit them all, and the point survives either way: a small operation with agentic tooling ran several intrusions in roughly a fortnight. The survey found that 23 per cent of directors have low or no confidence that a serious AI incident in their organisation would be detected and escalated. The outputs to minute are the prohibited-use list and the name of the person who can switch a system off.
Page 30: the maturity journey
Page 30 sets out a five-stage journey, from Aware, where the board recognises AI as strategic but has limited visibility of what is deployed, through Mapped, Controlled, and Strategic, to Adaptive. The paper says it is not a checklist or a prescribed sequence, which is true, and it is still the most useful page in the document for a chair, because it forces an honest self-location. The survey suggests where most boards will land: 24 per cent of respondents describe their organisation as well or highly prepared to deploy and scale AI responsibly, and 21 per cent rate their board’s AI literacy as high. Most are at Aware, whatever the AI slide in the strategy deck implies. The output to minute is the rung the board believes it is on, the evidence for that belief, and the next rung, reported annually under the oversight cadence the paper sets out in its first annex.
Use it before it is used on you
The paper is voluntary in law and will stay that way for some time, but a published standard backed by the directors’ institutes becomes the reference point for anyone asking whether a board did enough. A board that has minuted a stop rule, a prohibited-use list, a named halt authority, and an honest position on the maturity journey has done the substance of what the ten principles ask, and it has evidence of having done it, which is what an adequacy test looks for. It is also the framework to which my own AI governance teaching will align from here, for the same reason: it is the one directors are most likely to be measured against. Read the paper this month, take three pages into the next meeting, and let the minutes show what the board decided.
If you want AI governance analysis that tells you which pages matter and what to minute from them, subscribe to AI in the Boardroom. I write for directors, chairs, and the advisers around them who carry accountability for AI decisions and need the standards, the evidence, and the board-level implications in one place. Future editions will keep the reading going, so your board’s time goes to decisions.