Where should an AI workflow stop and ask a person?

Start with what the action changes

Imagine an assistant preparing a supplier update. Reading an approved record, drafting a change, and sending that change to the supplier have different consequences. Treating them as one permission hides the most important decision.

Our starting point is to map each action: what it can read, what it can change, who is affected, and how a mistake would be corrected. A draft can be easy to discard. A message sent outside the organisation may require a different level of review.

Make approval specific

A useful approval request shows the proposed action before it happens. It identifies the destination, the exact content or values being changed, the source evidence, and anything the system could not resolve.

“Continue?” gives a reviewer little to work with. “Send this revised delivery date to these two contacts?” makes the decision visible. The interface should also offer a clear way to correct or decline the proposal.

• Show what will change and who will receive it.
• Place the supporting source beside the proposed result.
• Explain uncertainty in plain language.
• Keep correction and cancellation within reach.

Put the boundary in the application

Approval needs to be enforced where the action runs. A sentence in a prompt asking the model to wait is not the same as an application that prevents an unapproved write.

Amazon Bedrock, for example, provides user confirmation for selected action-group functions. The application still needs to present the proposal clearly and handle the user’s response. The platform feature and the review experience need to be designed together.

Test the uncomfortable cases

Try a missing approver, an edited proposal, an expired approval, and the same request arriving twice. Decide what should happen if a source changes while someone is reviewing it.

For a first release, we recommend one bounded workflow with a visible decision boundary. Observe where reviewers need more context, then improve the system around those questions. The aim is an approval someone can understand and stand behind.