Drill it before it happens. Short tabletop exercises, run every quarter, teach a team the three moves that matter in an agent incident: pause or narrow the agent, escalate to a named owner, and report from the record. Reading a procedure once does not build the reflex. Rehearsing it on a realistic scenario does.
What counts as an AI agent incident?
An incident is any event where an agent acted, or nearly acted, in a way the business did not intend. Common examples: an agent drafted payments against an invoice whose bank details came from a spoofed email; an agent's certificate lapsed and a queue of held work appeared overnight; a reviewer approved a batch of held refunds without reading them; a source the agent relies on went stale.
Some of these are agent faults. Many are people or process faults that the agent exposed. Teams should learn to treat both the same way at first: contain, escalate, record, then work out the cause. The full response sequence is in how to handle an AI agent incident.
What should every team member be able to do?
| Move | Who | What they do |
|---|---|---|
| Spot it | Any reviewer | Recognise an unusual pattern of verdicts, approvals or refusals |
| Contain it | Named owner or deputy | Pause the agent or lower its depth for the affected capability |
| Escalate it | Reviewer to owner | Use the named escalation route, not a group chat |
| Preserve it | Owner | Leave the record untouched; it is append-only and is the evidence |
| Report it | Owner with risk or compliance | Write the incident report from the record, not from memory |
| Restore it | Owner, after review | Return the agent to its earned depth only with fresh evidence |
The table matters because incidents happen out of hours. The night reviewer must know who can pause the agent and how to reach them.
How do you run a tabletop drill?
- Pick one scenario grounded in your own operation. A short script helps, such as the one in the AI agent incident drill script template.
- Give the team the first signal only. For example: "Three supplier payments were approved in ten minutes, all to new bank details."
- Ask what they do first. Wait for an answer before revealing more.
- Reveal the record. Show the filed verdicts. Ask what the deciding bound was and which sources the agent used.
- Ask who they call. The answer should be a name, not a team.
- Ask what they write. Have someone draft the opening lines of the report using the AI agent incident report template.
- Debrief for 15 minutes. What was slow? Who was missing? What was unclear?
Keep it to an hour. The goal is a reflex, not a lecture.
What does a good drill reveal?
Take a regional insurer's claims team. In a drill, the scenario is an agent whose run that morning cleared no depth, dropping it to depth 1. Suddenly every small claim it used to pay sits held.
The drill exposes three things. Reviewers assumed the agent had broken and started paying claims by hand outside the workspace, which leaves no record. Nobody knew who could explain the drop to the claims director. And the team did not know that a failed run drops depth the same day by design. Each finding becomes a fix: a rule that held work is cleared inside the workspace, a named owner for certificate events, and a line in onboarding.
How often should you drill?
Quarterly for teams that run agents at depth 3 or above, and after any change in depth, limits or sources. Rotate scenarios across agent faults, source faults and people faults. A drill for the technical team looks different; see running an incident drill for the IT team. The general shape of a drill is explained in what an incident drill is.
Where the Academy fits
The Operator track of Pryme Intelligence Academy includes a module on when something goes wrong, alongside human review and reading the workspace. Labs run on the real product in a free Build workspace, so a team can practise pausing an agent and reading the resulting record without touching production. Wider training choices are covered in training teams for AI agents.
Questions people ask
Should the whole team attend every drill?
Rotate attendance so everyone takes part at least twice a year, and always include the named owner and their deputy. Night and weekend staff should be included, since incidents do not keep office hours.
What if a drill shows the procedure is wrong?
Change the procedure that week and rerun the scenario at the next drill. A drill that finds nothing is usually too easy.
Who signs off that an agent can return to its earlier depth?
The agent's named owner, on fresh evidence from certification. Depth is earned again, not restored by request.