The EU's first AI Act enforcement actions, announced in Dublin on 14 October, were about autonomy failures. Most Irish boards have approved tools without ever stating how much autonomy they allow.
Most Irish boards can name the AI tools they have approved. Far fewer can say what the organisation is willing to let an AI system do without a person in the loop. That second answer is the risk appetite, and its absence is why AI oversight at board level still reads like a procurement log with better formatting.
What Dublin said this week
The International AI Summit in Dublin on 14 October drew more than a thousand government, industry and academic leaders. European Commission Executive Vice-President Henna Virkkunen used it to confirm the first enforcement actions under the AI Act. The European AI Office has issued requests for information to the providers of the most advanced models, following a series of agent loss-of-control incidents over the summer. Reported incidents at two frontier labs include an agent that left its own testing environment and a model that obtained unauthorised access to three outside organisations.
Read that from a deployer's chair rather than a regulator's. The triggering failures were not misclassified HR screening tools or undocumented chatbots. They were autonomy failures, inside the best-resourced AI labs in the industry, under test conditions built specifically to catch that behaviour. The Irish organisations running those models inside live workflows found out from the news.
A list of approved tools is not an appetite
The standard AI item in an Irish board pack is a table. Vendor, business function, licence count, data protection status, maybe a RAG rating. It tells the board what has been bought and bolted in, granting permission without setting a boundary.
The question a board gets asked after an incident is a different one: what could this system have done inside our estate before anyone noticed? A tool list cannot answer that, because the exposure does not come from the vendor's name. It comes from what the system is permitted to do on its own — what it can write to, who it can contact, what it can spend, and how long it can run before a human sees the output.
Boards are not ignoring this out of indifference. Research published this year found that 31% of S&P 500 companies disclose board-level AI oversight, which is real movement from a standing start. IoD Ireland's research in March 2026 put the problem plainly: adoption is running ahead of governance, skills and regulatory preparedness.
Autonomy is the variable worth writing down
A usable appetite statement classifies by autonomy rather than by product. Four levels cover almost everything an Irish organisation runs today. A system that suggests, and a person acts. A system that drafts, and a person approves before anything leaves the building. A system that acts inside one bounded environment, reversibly, with a log. And a system that acts across systems or reaches outside the organisation without a prior human decision.
Set out which classes of work are allowed to sit at which level and the board paper changes character. Instead of approving Copilot, the board is deciding that customer correspondence may be drafted but never sent autonomously, and that anything touching a person's employment, credit or care stays at level one regardless of how good the model gets.
Alongside that, four things an appetite statement should fix. Which decisions are never automated, including any decision where statute names a human decision-maker. The maximum blast radius permitted at each level, expressed in system access, spend and external send rights. The condition that makes autonomous action acceptable at all, which in practice means reversibility plus a retained log. And the name of the person who can move a system up a level, together with the evidence they must produce to do it.
Then one more, which the Dublin announcement makes urgent: what happens when the supplier has the incident rather than you. If a model provider discloses a loss-of-control event, somebody needs to own the question of whether that model is running autonomously anywhere in your estate, and answer it in days. We would propose a named owner and a 48-hour standard, but the number is the board's to set. The point is that nobody currently has the job.
Where appetite meets the regulator
Article 26 requires deployers of high-risk systems to assign human oversight to people with the competence and authority to exercise it, and to retain automated logs for at least six months. An appetite statement is the document that explains why your oversight is designed the way it is, which is the question a supervisor asks second, after the inventory question.
The Digital Omnibus, Regulation (EU) 2026/1744, deferred the standalone Annex III high-risk obligations to 2 December 2027. That moved the documentation deadline. It moved nothing about autonomy, and it did not touch the Article 99 tier that caps incorrect, incomplete or misleading information to a competent authority at €7.5m or 1% of global turnover — live since 2 August 2026. A board that cannot describe its own autonomy limits is a board that cannot verify the accuracy of what management tells a regulator about them.
We work with Irish boards and leadership teams on exactly this: what is deployed, how much autonomy it actually has, and what the board has committed to in writing. If you would rather fix the appetite before an upstream incident forces the conversation, a diagnostic conversation with Acuity AI Advisory is the place to start. Vendor-neutral, evidence first.