Accountability

One name

Accountability is satisfied by a person. It is not satisfied by a committee, a function, a policy, or a distributed sense that everyone should be careful.

This is the most frequent failure in the domain and the least likely to be recognized as one, because the arrangements that fail here look orderly. There is a steering group. It meets. It has terms of reference and a membership list. What it does not have is a person who answers when an outcome is examined, which is what accountability means here.

The test is retrospective. If this system produced a materially wrong outcome six months from now, and the matter were examined by a regulator, a court, or a journalist, whose name would appear? Where the answer is a group, accountability has not been assigned. Where the answer is a person who has since moved roles, it was assigned once and not maintained, which ACC-6 addresses.

This is a rule about operational accountability for a system. It does not displace the collective or statutory accountability of a board, its directors, or its officers, which applies as the law of each jurisdiction provides. What it prevents is accountability for a system being diffused until nobody holds it.

Seniority matched to consequence

Accountability placed at the wrong level is a distinct failure from accountability not placed at all, and it is harder to see.

AI systems are frequently owned by the function that built them. That is a natural consequence of how they arrive and it is sound while the system is experimental. It becomes a governance problem once the system makes decisions affecting customers, employees, or the organization's regulatory position, because the accountable person then lacks the authority to stop it, absorb the cost of stopping it, or answer for the consequence outside the organization.

ACC-1 therefore requires seniority to be measured against the system's risk classification rather than against the complexity of the work. The two are unrelated. A technically simple system that declines credit applications carries more consequence than a sophisticated one that drafts internal summaries.

What the board keeps

A board delegates the operation of the organization to management. It cannot delegate its own duty of care, and directors remain individually subject to it regardless of what has been assigned downward.

ACC-2 requires the split to be recorded rather than assumed. The value of the record is not compliance. It is that the act of writing it down forces the board to decide questions it would otherwise leave open: whether deployment of a system above a given risk classification requires board notification, whether the board expects to see realized benefit as well as proposed benefit, and at what point an AI incident must reach it rather than being managed within management.

A board that has never recorded this has usually not decided it. When the question arises, it arises during an incident.

Approval separated from delivery

ACC-5 states a control that is standard elsewhere in an organization and frequently absent for AI: the person accountable for delivering something does not also approve its release.

Where the two coincide, the approval step still occurs and still generates a record. It is not a control, because no independent judgment was applied. The person with the strongest interest in proceeding assessed whether to proceed.

This matters more for AI than for conventional systems, because the question at release is not whether the system functions. It is whether the residual exposure is acceptable, which is a risk decision rather than a technical one, and which the delivery owner is the least suitable person in the organization to take.