Check the system, not the conversation. Decide before the work starts what must be true when it’s done, then read the actual record afterwards and compare.
If it says “done,” isn’t that enough?
No. An agent can update the wrong record and still report success. Two questions matter: did something change, and was it the change requested?
1. Define what done means first.
Example. Assign service request SR-104 to the agreed team.
Done means the same request ID, the approved team and the required status.
Where that comes from. The authorized request, not a target the agent works out later.
Practical check: Decide what must be true before the work starts.
2. Check the record itself.
Read the actual service record, or another result you can rely on.
Compare the target, the values and whether any required step is missing.
Keep a task identifier, the result and the check you made.
Practical check: A log of what it tried and a check of what’s there answer different questions.
3. Three possible outcomes.
- Verified complete
- The required change is confirmed.
- Verified not complete
- The record shows the action didn’t happen.
- Unknown
- A timeout or partial result means you can’t confirm. Check what changed before retrying.
Practical check: No response doesn’t prove no change. Uncertain actions need a way to recover.
A completion rule.
Write down three things: the right target, the state that proves the work is done, and who checks uncertain cases and prevents duplicates.
Practical check: “Done” should follow the check, not replace it.
Read more: Letting AI Update HR or Payroll Records: Approve Every Change First · Why an AI Agent Works in the Demo but Fails on Real Work
The example is fictional. This is general guidance, not advice for your situation.