The accountability test

Boards are being told that they must oversee the organisation's use of artificial intelligence, and most have no idea how. The frameworks they are offered run to dozens of pages. Here is a test that fits on an index card, and that tells a board more about the state of its AI governance than any framework will.

Take the single most consequential agent the organisation runs: the one that touches the most customers, moves the most money, or screens the most people. Imagine it fails at scale overnight. Then ask four questions.

One. Is there a named role, not a department, that would get the call? If the answer is a person's job title, the organisation has at least the beginning of accountability. If the answer is "the technology team" or "the business unit" or, most commonly, a pause followed by "we'd convene the relevant people," there is no accountability, only a plan to find some.

Two. Does that person have the expertise to judge whether the agent's output was wrong? Accountability without expertise is theatre. A senior manager who can be blamed but cannot evaluate the work is not accountable in any meaningful sense; they are a scapegoat held in reserve. The accountable human must know the task well enough to recognise a bad answer, which usually means they learned the task by doing it.

Three. Do they have the authority to switch the agent off without a meeting? Responsibility cannot be separated from authority. If stopping the agent requires the approval of IT, the vendor, or a committee that meets on Thursdays, then the person who answers for the agent cannot act on their own judgment, and the arrangement will fail at exactly the moment it is needed.

Four. Is the version of the agent running tonight the version that was approved? Vendors update models continuously. If the organisation has no process by which an update triggers re-testing and re-approval, the agent that fails will not be the agent anyone signed for, and the first question becomes unanswerable in retrospect.

There is experimental support for taking the first two questions seriously. A 2026 study by Boston Consulting Group and Boston University, published in Harvard Business Review, found that when agents were framed as colleagues rather than as systems with a named human owner, individual accountability fell and review quality dropped. Diffuse ownership is not a neutral default; it degrades the checking.

Four yes answers mean accountability exists for this agent. The board should then ask whether it exists for every agent, which is where most organisations discover that they do not know how many agents they have. Two or three yes answers mean the design is partial, and the missing answer tells you exactly what to fix. One or none means the organisation is in the condition most organisations are in: many participants in the decision to deploy, no owner of the consequence.

The test has a second use. It defines the job. Whoever can make the answer to all four questions yes, for every agent, across the whole organisation, is doing the work I have called the Chief Workforce Officer's, whether or not they hold the title. They are maintaining the register of agents, assigning the named humans, ensuring those humans have expertise and authority, and controlling the update process. If no one in the organisation can do that today, the board has learned what it needed to learn from the test: not that the technology is dangerous, but that the organisation has not yet decided who is responsible for it.

I have found that this test, asked in a boardroom, produces silence and then a conversation that no framework produces. The silence is the finding.