Human in the loop: designing approvals that don’t slow you down
Approvals keep AI safe, but badly designed ones kill the time savings. Here is how to put people in the loop without creating a bottleneck.

Keeping a person in the loop is the right call for many AI workflows. The trouble starts when every small step waits for someone to click “approve”, and the automation ends up slower than the manual process.
Approve by risk, not by habit
Sort actions into three groups:
- Low risk: sorting, tagging, summarising, drafting internal notes. Let the AI do these alone and review a sample each week.
- Medium risk: customer replies, record updates. Approve in batches, or approve only when the AI is unsure.
- High risk: refunds, pricing, contracts and promises. Always a named person.
Make approving fast
Bring the approval to where people already work, such as email, chat or their task list. Show the draft, the source information and a one-line reason, with approve, edit and reject in one click.
Learn from the edits
Every edit a person makes is feedback. Review the most common edits monthly and update the instructions or rules. Over time, the share of items needing approval should fall.
Set a service level
Agree how quickly approvals must happen, and escalate when they wait too long. A human in the loop is a safeguard, not a parking spot.


