Most of the time, an AI agent does what it was set up to do. The hard questions come with the exception: the payment outside the normal range, the record changed in an unusual way, the customer contacted when they should not have been.

Afterwards, someone will ask three things. Was the action authorised? Who allowed it, the rules or a person? And what record shows it? In many organisations, the honest answer is a reconstruction from system logs, chat messages and recollection, assembled under pressure.

Why exceptions expose the accountability gap

Routine actions rarely get examined. Exceptions do, by a manager, an auditor, a regulator or a customer. That is when three gaps show up:

  • Logs show activity, not authority. A log entry says the action ran. It does not say whether the agent was entitled to do it, or whether anyone approved it.
  • The exception route is informal. When an agent hits something outside its normal limits, the "approval" is often a message or a call, recorded nowhere that links it to the action.
  • Ownership is unclear. It is not obvious whether the agent's owner, the workflow owner or the person who approved answers for the outcome.

The control question

Ask: for any unusual action, can we show who or what requested it, whether it was within the rules or approved as an exception, and by whom?

In an agent workflow, that means:

  1. Exceptions are defined. Actions outside agreed limits are recognised as exceptions by rule, not discovered afterwards.
  2. Exceptions need an explicit decision. An exceptional action is refused, or held until a named person approves it. It is not quietly allowed.
  3. The decision is evidence, not a reconstruction. Each decision is recorded when it happens, with the requester, the result and the reason, in a form that can be checked later.

What IntelXview's AI Control Plane does here

For actions routed through it, IntelXview's AI Control Plane (ACP) evaluates each request against policy before it runs, and records the decision. A paid proof of concept can test exception handling on one agreed workflow:

  • Human approval: hold actions outside agreed limits until a named person approves them.
  • Failure handling: test expiry, revocation and control unavailability, limited to scenarios ACP supports today, and record what happened in each test.
  • Evidence: inspect the decision record for each allowed, refused and approved action. The published decision evidence example shows records that can be recomputed, and the signed evidence pack shows how signatures are checked.

ACP covers only actions routed through it, and a record shows what ACP decided, not everything that happened elsewhere. A proof of concept usually begins with a controlled demonstration on synthetic data; integration with your systems is scoped and priced separately.

Where this comes from

We posted on this theme on LinkedIn in October 2026, but those posts drew no external replies. It is included because it completes the control path in the other three articles, from authority and approval to evidence. Treat it as our view of the problem, not as something we have heard repeatedly from buyers.

Next step

If you need to answer for unusual actions taken by an AI agent, book a free scoping call to test exception handling on one workflow. If you advise clients on AI risk or assurance, book a partner call instead. See what a PoC includes or how partner work is organised.