Updated 17 October 2026. This article has been revised to separate the policy timeline from the product argument. See the current AI Act timeline for the category-specific dates and primary reference.
A delayed deadline is a useful moment to inspect the monitoring plan. A system can change its behaviour through new inputs, dependencies and configuration long before the policy calendar changes again.
Define what deserves attention
A monitoring plan should name the conditions that matter and the response they trigger. Counting every event is not the same as noticing a dangerous one. Start with the failure modes that would change the release decision if they appeared during evaluation.
For an illustrative support agent, repeated unsupported answers and attempted out-of-scope actions deserve different treatment. One may require reviewing retrieval and answer quality. The other may require examining permissions and stopping an action path. Combining them into one health score hides the response.
Assign the alert to a person or team with enough context to investigate. Record what they should inspect, when they should escalate and which actions they can take. An alert without an owner is a notification looking for a hobby.
Test the monitoring process with a controlled example. Did the signal appear? Did it reach the right person? Could that person tell what happened? This checks the process rather than merely the existence of a dashboard.
| Decision | Evidence to request |
|---|---|
| Detect | Relevant signal appears |
| Investigate | Reviewer can reconstruct the event |
| Respond | An authorised action follows |
What to do with this
Choose signals that help someone act. A smaller set of interpretable alerts is more useful than a wall of numbers that nobody can connect to a decision.
Record the decision, the responsible owner and the evidence still missing. The decision-record example shows a lightweight way to do that. Where a claim concerns system behaviour, connect it to evaluation evidence. These are proposed product practices, not a finding that a system complies with the law.
Primary timeline reference: European Commission, AI Act. The operational recommendations above are my analysis.
