Where an AI workflow needs a human decision

An AI workflow can prepare useful work without being allowed to complete every action. Before building an automation, list what it will read, what it will generate and what it could change in another system. The risk often sits in the action, not the text on the screen.

Consider a practice assistant that drafts a customer reply. It can retrieve an approved policy, extract relevant details and produce a proposed response. The reviewer should see the sources, the draft and any uncertainty before deciding whether to send it.

Do not turn missing information into a confident answer. Give the workflow a clear way to say that a source is absent, records conflict or a request is outside its scope. Test those cases explicitly. A system that stops at the right time can be more useful than one that always produces a response.

Treat external documents and messages as data, not authority to change the workflow. A document may contain instructions that conflict with your task. Keep permissions narrow, keep credentials out of prompts and require approval before consequential external actions.

Make approval specific. Show what will be sent or changed, the destination and the relevant evidence. A generic approval at the start of a long workflow may not give a person enough information to assess the final action.

Test normal inputs, ambiguous requests, unsupported claims and malicious instructions using synthetic data. Measure useful behaviour rather than judging a single polished answer. Keep a record of failures and repeat the same tests when prompts or tools change.

In a student project, demonstrate the approval boundary and a recovery path. Explain what happens if the system retries, the data source is unavailable or the reviewer rejects the draft. That makes the automation easier to assess and maintain.

More business perspective →