What human review should actually look like

Most of the AI work we are asked about falls into two groups: routine communications and approved-information workflows. Both can save real time. Both can cause harm if a draft goes out unread.

Review is a place, not a promise

A review point should be a specific step in the process with a named person and a clear decision. Before this message is sent. Before this record is updated. Before this answer is given to a client. If nobody can say where the check happens, there is no check.

What the reviewer should be looking for

• Source: does the content come from approved material, and can the reviewer see which part?
• Scope: has the draft answered the question asked, without adding advice nobody authorised?
• Facts: are names, dates, amounts and references correct? These are the details that quietly go wrong.
• Tone: does it sound like your organisation rather than a generic template?
• Sensitivity: is anything in there that should not leave the organisation?

Design so review is easy

Review fails when it is tedious. Short drafts, visible sources, an obvious edit box and a deliberate send action all make checking realistic. If the reviewer has to read three screens to approve a two line reply, the step is more likely to be skipped.

Where responsibility stays

These systems help people work faster on repetitive tasks. They do not replace professional judgement, and we do not describe them as if they do. Boundaries, data handling and who remains responsible are agreed before work begins, as described on how buying works, and there is a short responsible use checklist you can work through yourself.