The accountability was in the person being removed, not in the machine.

For as long as a human being stood in the decision path, that person was the accountability. Not the log, not the policy, not the control framework: the person. Everything downstream, from professional liability to the chain of command to the management assertion, is built on there being somebody who decided and can be asked why.

A self-made record was enough, because the person carried the accountability and the record only supported it. Machines now carry more and more of that authority, and nothing has been designed to take the place of the person being removed.

A leading industry analyst firm puts an average global Fortune 500 enterprise at fewer than fifteen agents in 2025 and more than a hundred and fifty thousand by 2028. Whatever one makes of the figure, no oversight function will grow ten-thousandfold in three years. The same firm reports that thirteen percent of organizations think they have the right AI agent governance in place.

That arithmetic disposes of a whole class of answer. Any approach that scales oversight by adding reviewers or approval steps has already lost, and it will lose quietly, as coverage falling rather than a control failing.

The same firm predicts that by 2027, forty percent of enterprises will demote or decommission autonomous AI agents because of governance gaps identified only after production incidents occur. Detection after the fact is not governance. It is archaeology.

Read that as a schedule, not a statistic. Today, the gap is discovered by the incident, and the remedy is retreat: pull the capability back because the accountability cannot be produced. Both are expensive. Neither is governance.

The instruments are already written. A declaration of conformity for a high-risk system. An authorization to operate. An operational resilience attestation. A management assertion. Each names a party who must answer for what a system did.

In our reading, the question on the form arrives before enforcement does. An organization that cannot evidence a decision is likely to be excluded by a procurement questionnaire long before it is penalized by a regulator, and an exclusion of that kind is rarely appealable, because nothing was alleged.

Rules written after an incident are written to prevent the last one. They are drafted fast, under pressure, in a week when no one argues for capability. They are broad because breadth is fast, and hard to unwind because unwinding looks like special pleading.

Rules that permit the mission get written when an architecture already exists to point at, so that what is codified works rather than stops. That is the window. It is open now, and it closes on someone else's schedule.