What is AI accountability?

AI accountability is the principle that a named person or body answers for the outcomes of each AI system an organisation uses, and can show how those outcomes came about. It has two halves: someone owns the result, and the evidence exists to explain it. Without both, accountability is a policy statement rather than something that can be tested.

Why does AI accountability matter once agents act?

With AI that only suggests, accountability sits with the person who acts on the suggestion. Agents change that. An agent that issues a refund, sends a letter or updates a record has acted, and there may be no person between its decision and the effect. The question "who answers for this?" no longer has an obvious answer unless the organisation sets one in advance.

Regulators in financial services already expect named individuals to answer for areas of the business, and agents do not change that expectation. DORA places responsibility for ICT risk on the management body, a point covered in DORA management body accountability for AI agents. The UK's Senior Managers and Certification Regime takes the same approach to individual accountability. The task is to map agent activity onto those existing lines. The approach is set out in agent governance is a bound, not a model.

How does AI accountability work in practice?

• An owner for every agent: a named person who answers for what the agent does, usually the head of the work it supports.
• An owner for every capability's authority: the person who sets the customer ceiling and limits for that capability.
• A named approver for every held action: the person who decided.
• A record that links them: each verdict shows the agent, the bound that decided and, where relevant, the approver.

Frameworks give this structure. The Govern function of the NIST AI Risk Management Framework centres on roles, responsibilities and culture, as discussed in NIST AI RMF accountability: roles and owners for AI agents.

What does that look like in an example?

A regional lender deploys an agent to handle hardship requests. The head of collections is named as owner. She sets the ceiling: the agent drafts payment plans, a team leader approves them. Six months in, a borrower complains that the plan offered was unaffordable. The record shows the agent's draft, the income and expenditure data it used, the team leader who approved it, and the policy in force at the time. The head of collections can answer the complaint and decide whether the fault lay with the data, the draft, the approval or the policy.

How is accountability different from responsibility?

Responsibility is about who does the work; accountability is about who answers for it. An agent can be responsible for a task, in the sense that it carries it out. It cannot be accountable, because it cannot be held to account: it cannot explain itself to a regulator, accept a sanction or change its own conduct. Accountability always lands on a person. Mixing the two up leads to the phrase "the AI decided", which no supervisor will accept.

Why do staff resist agents, and what does accountability have to do with it?

Often people object less to the agent than to being blamed for its mistakes without the means to prevent them. A team leader asked to approve agent drafts at volume, with no time to review and no power to change the agent's bounds, carries accountability without control. Fixing that mismatch is the subject of staff resistance is usually about accountability, not jobs.