Are operations managers accountable for an AI agent’s mistakes?

It depends on what they approved. Operations managers are rarely accountable for a single AI agent action, but they are accountable for the conditions those actions run in: the depth they asked for, the limits they set, who they allowed to approve held work, and how much review time they gave people. The record shows each of those decisions and who made it.

What does an operations manager actually decide?

An operations manager runs the process the agent works inside. They decide which queues an agent handles, ask for its depth to be raised once Attest shows it can clear the gates on the business's own cases, set the business's own limits on value and volume, and staff the review team. They also decide how quickly held items must be cleared.

Each of those is a decision with consequences across thousands of actions. A held queue with a four-hour target and too few reviewers produces rushed approvals. A limit raised without an end date produces drift. Neither shows up as one bad action; both show up as a pattern.

Which decisions belong to the manager, and which do not?

• The manager owns: requests to raise depth, the business's own limits, approval rights for each reviewer, service-level targets for held items, and staffing of review.
• Reviewers own: each held action they approve or reject.
• The agent's builders own: its brief, sources and certification evidence.
• The platform owns: what is withheld at the floor for every agent.

The minimum chain makes this concrete. Every action meets five bounds at once, and the verdict names the one that decided. When the deciding bound is a business limit, the manager who approved it is visible.

What does the record show?

The record files every verdict, and the change history files every approval of a limit, a depth request or a reviewer's rights. It also shows patterns: how many held items were approved per hour, by whom, and how long each was open. A manager can be asked why a reviewer cleared three hundred held items in an afternoon, and the record answers part of the question.

That cuts both ways. A manager whose team rubber-stamped held work under a target the manager set owns the target. A manager who staffed properly and set sound limits can show it.

An example: the backlog cleared in bulk

Take a regional insurer's claims operation. After a storm, the agent holds 600 small property claims for review. The operations manager, under pressure to pay quickly, allows reviewers to approve held claims in batches of fifty for one week. The change is filed with their approval.

A cluster of fraudulent claims passes through the batches. The record shows the agent's fraud flags on several of them, the batch approvals and the manager's temporary rule. Accountability for the pattern sits with the manager. Individual reviewers share it only where they opened a flagged claim and approved it anyway.

Which controls protect the operation?

New managers benefit from supervised practice: a two-week practicum for operations managers, logged in a practicum log template for operations managers. As agent certificates lapse after 90 days, plan for preparing operations managers for 90-day recertification.

Owners of risk appetite face a related question; see risk managers and AI agent mistakes. The wider topic is in AI literacy and certification for people who work with agents.

Where the Academy fits

Operations managers are named as a core audience of Pryme Intelligence Academy's Operator track, whose modules include usage and billing and running the rhythm. An admin view shows the team's progress, results, certificates and expiry dates, which gives the manager evidence for every approval right they grant.