Why can’t accountability live in a policy document?
Because an agent acts faster and more often than any committee reviews. Writing in Forbes, the chief executive of an engineering consultancy argues that accountability “cannot live solely in policies, steering committees or employee training” and must be built into the technology itself (Forbes Technology Council, 26 October 2026, read 2 November 2026). A policy can say an agent must not pay invoices above a limit; only a permission can stop it. A policy states who should answer; a design makes sure someone can. That is the gap between AI governance on paper and governance at runtime.
What does accountability by design look like in an AI system?
Four parts, each set before launch. A named owner per agent who can defend its scope. Explicit decision rights: what the agent may access, decide and change alone. A tool approval gate wherever an action touches money, customers or regulated records. And an agent audit trail that records identity, approvals and effects, not only chat messages. The same Forbes piece puts it as defining “what every agent can access, decide and modify, and when it must seek human approval.” Every one of the four is enforced by the system, not remembered by staff.
What are examples of accountability in AI?
An accounts payable agent that releases invoices below a threshold and routes the rest to a controller, with both paths logged. A customer service agent that drafts refunds a supervisor sends. A lending model whose every decline carries an explanation a person can give the applicant. In each case the question “who approved this?” has an answer in the logs, not in someone’s memory. Autonomy moves the work to software; it never moves the answerability.
Who is accountable when an AI agent gets it wrong?
The company that deployed it, in nearly every case a customer or regulator will accept. Vendor contracts and insurance can shift the cost, not the responsibility, as our piece on who is accountable when an AI agent gets it wrong argues. Design for the post-incident question before the incident: one owner, clear limits, and evidence that survives the investigation. Under agent governance, that record is what turns a failure into a fixable event rather than a dispute.