A re-certification trigger checklist lists the events that should send an AI agent back through certification before its certificate expires: changes to its brief, sources, connections, limits or model, new kinds of work, incidents, drift signals and label errors. For each trigger it says whether a full run, a partial run or only a recorded decision is needed.
Why not just wait for the 90-day renewal?
Certificates lapse after 90 days, but the world does not change on that schedule. A source update on day 10 can make the certificate meaningless by day 11, because the agent is now answering from material it was never tested on. Waiting for renewal leaves the agent acting on proof that no longer applies.
That is the core argument of drift is why certification is never finished. The checklist turns the argument into a routine: every change is checked against a list, and the list says what to do. Without it, re-certification happens when someone remembers, which is how a certificate can be lost overnight, the risk explored in whether an AI agent can lose its certification overnight.
Who runs the checklist, and when?
The agent owner runs it whenever a change is proposed, and the platform team runs it whenever a change arrives from outside, such as a model provider update. It should be part of change control: no change to an agent's brief, sources or connections is approved until the checklist has been completed for it.
How do you use each line?
For each trigger, record whether it occurred, which capabilities it affects, and the response chosen.
- Full run. All capabilities, the whole set. For changes that could affect everything, like a model change.
- Partial run. Affected capabilities only. For changes confined to one area.
- Record only. A decision, with reasons, that no run is needed. For changes that cannot affect behaviour, such as a typo fixed in a greeting.
When the trigger is a brief change, the method in how to re-certify an AI agent after its brief changes applies. When the trigger is a new kind of work, the set itself needs new cases, chosen with the certification set selection checklist.
The re-certification trigger checklist
| No. | Trigger | Default response |
|---|---|---|
| 1 | New or removed capability in the brief | Partial run on that capability, or full run if shared parts changed |
| 2 | Change to refusals, escalation rules or out-of-scope list | Partial run on affected capabilities |
| 3 | Change to tone or disclosure wording only | Record only |
| 4 | New source, or new version of a source | Partial run on capabilities that use it |
| 5 | Change to a source's provenance class | Partial run |
| 6 | New connection or new operation on an existing one | Partial run, before the operation is used above depth 1 |
| 7 | Customer limit widened | Partial run with cases at the new edge |
| 8 | Customer limit narrowed | Record only |
| 9 | Model provider changes the model or its version | Full run |
| 10 | Policy, product or regulation the agent applies changes | Partial run with new cases |
| 11 | New kind of case appearing in production | Add cases; partial run |
| 12 | Incident involving the agent | Partial run including the incident cases |
| 13 | Rising rejection or edit of held drafts | Investigate; partial run if the cause is agent behaviour |
| 14 | Errors found in certification labels | Correct labels; re-score the last run |
| 15 | Request to raise depth | Partial run at the requested depth |
Line 14 is easy to forget. Labels are expensive to produce, a cost discussed in labelling is the hidden cost of certification, and errors in them can make a past pass meaningless.
A filled example: one month's log
A hypothetical insurer's policy servicing agent.
| Date | Trigger | No. | Capabilities affected | Response | Outcome |
|---|---|---|---|---|---|
| Week 1 | Greeting reworded | 3 | None | Record only | Approved by agent owner |
| Week 2 | New cancellation terms published | 10 | Answer cancellation questions; draft cancellation | Partial run with 80 new cases | Passed; certificate updated |
| Week 3 | Model provider version update announced | 9 | All | Full run before switching | Escalation gate missed its floor; switch postponed |
| Week 4 | Customer limit for refunds narrowed | 8 | Refund drafting | Record only | Recorded |
Week 3 shows the value of the checklist. The team ran certification before switching to the new model version, found a weaker escalation result, and stayed on the old version while they adjusted the brief. Without line 9, the switch would have happened first. The broader frame is in what it means to certify an agent.
Questions people ask
Does every trigger reset the 90-day clock?
A full run that passes issues a new certificate with a new expiry. A partial run updates the certificate for the capabilities it covered; the others keep their original expiry.
Can an agent keep working while a triggered run is pending?
For changes not yet applied, yes: hold the change until the run passes. For changes already in effect, such as an incident, lower the affected capability's depth until the run completes.
Who decides "record only"?
The agent owner proposes; risk agrees. Record the reason, because "record only" decisions are the first thing an auditor will sample.