Where does the answer come from?
An answer should be grounded in your plant records, and you should be able to see which ones: the telemetry, the SPC chart, the maintenance history. An answer without sources cannot be checked, and operators stop trusting a system after its first confident mistake.
Who approves the action?
Recommending a feed-rate change is useful. Making it silently is a risk. A named person should approve anything that changes how the line runs, and the approval should be recorded with a reason.
This is not caution for its own sake. It keeps accountability where your quality system expects it to be.
Where is it logged?
If AI recommendations and the actions taken on them live in a separate log, your batch record is no longer the full story of the batch. They belong in the same audit trail as signatures and deviations.
Ask to see that trail in a demo. If it takes more than a click to find, it will not survive an audit.
In short
- Demand sources for every answer.
- Keep a person in the approval loop.
- One audit trail, AI included.