Most boards now get a regular AI update. Fewer have answered a simpler question. Which AI decisions actually need the board’s approval?
Until recently, the answer could stay vague. AI was a set of tools employees used, and decisions about those tools belonged to management. That is changing quickly, because the tools have started to act on their own.
AI agents don’t just answer questions. They send messages, move data, write and deploy code, and increasingly hand work to other agents. Gartner expects the average Fortune 500 company to have more than 150,000 agents in use by 2028, up from fewer than 15 in 2025. In the same release, only 13% of organizations said they believe they have the right AI agent governance in place (Gartner).
Those numbers describe a management problem first. Inventories, permissions and monitoring are operating work. But when software can act on its own at that scale, some of those decisions become board decisions, and most boards haven’t drawn the line.
The line should be materiality. A board doesn’t approve every agent any more than it approves every hire. It does need to see the AI decisions that could materially affect the organization, and it should agree with management in advance on what that means.
A practical test asks a handful of questions. Could the system affect reported financial results? Does it touch a legal or regulatory obligation? Will it make or shape decisions about customers or employees? Does it act with significant autonomy, without a person approving each step? Does it introduce material cybersecurity or privacy risk? Is it central to the product or the business model, where a failure would damage the organization’s reputation? If the answer to any of these is yes, the decision belongs with the board, or at least with the committee that owns technology risk.
A test is only as good as the people applying it, and the board should know who they are. Materiality calls about AI tend to land between the CFO, the general counsel, the CIO and whoever leads AI. The person most often left out is the one responsible for privacy, even though data is where much of the exposure sits. It is worth asking who is in the room when management decides whether an AI issue is material, and who isn’t.
Incident response is the next gap. Most boards have walked through a cyber breach plan. Far fewer have thought about what happens when an AI agent does something it wasn’t supposed to do. It might send confidential data to the wrong place, take an action outside its permissions or keep running after it should have stopped. The board’s questions are familiar from cyber. Who has the authority to shut the agent off, and how fast can they do it? When does the board hear about it, and from whom? Is there a record of what the agent did and why? For public companies, an agent incident that is also a material cybersecurity incident still runs on the SEC’s four-business-day disclosure clock (Skadden). That is not the moment to be working out the escalation path.
Directors also need a shared vocabulary for oversight, because management presentations often blur it. “Human in the loop” means a person approves before the system acts. “Human on the loop” means a person monitors after the fact and can step in. Both are legitimate designs, but they are very different levels of control. When a presentation says a human is involved, ask which one.
The same goes for accountability. An AI recommendation is not a decision. When an agent recommends and a person clicks approve, the responsibility should still sit with that person, and the record should show it. Boards should be wary of any process where it is hard to say who actually decided.
None of this requires the board to become technical. It requires a few written decisions. The board should set the materiality test for bringing AI matters to it, and know who in management applies the test. It should agree on how agent incidents escalate and when the board is told. And it should decide how much autonomy the organization is willing to grant before a person has to approve. A short policy or a committee charter amendment can hold all of it.
This is the same discipline boards already apply elsewhere. Boards don’t approve every contract, but they set the threshold above which contracts come to them. They don’t review every expense, but they approve the budget and the exceptions. AI deserves the same treatment, and the threshold should be set before the first serious incident, not in response to it.
We think about this from both sides at Aureclar, because we build AI for the board’s own work. The design question is the one boards should be asking of management. What is the AI allowed to see, and who stays accountable for what it produces? In our case the assistant drafts, flags and recommends. Directors decide.
A board doesn’t need to approve every agent. It does need to make sure it never learns about a material one for the first time from the press.
Josh Schiffman is the founder and CEO of Aureclar, AI for the boardroom that helps directors prepare and exercise better judgment.