· 4 min read
Where to put the human in an AI workflow
Approving every AI draft sounds safe. In practice, reviewers stop reading. Here is how to design review points people can actually keep up with.
Somewhere in your organization, a person is being asked to approve their two-hundredth AI draft of the day. By the fortieth, they read it. By the hundredth, they scan it. By the two-hundredth, they click.
The control is still there. The log says every draft was reviewed. But nobody is really in the loop.
Approving everything feels like the safe design. It is often the least safe one, because it teaches people to stop looking.
Review is a limited resource
Attention works like a budget. Every request for approval draws from the same account. Spend it on low-risk items and there is nothing left when the risky one arrives.
Hospitals learned this with alarms. When every beep is urgent, none of them are. The fix was not more alarms. It was fewer, better ones.
The question to ask is not “should a person review this?” It is “which decisions deserve a person’s attention, and how do we protect it?”
Sort every action into one of three modes
For each thing a workflow might do, choose one mode.
| Mode | What happens | Use it when |
|---|---|---|
| Act | It happens on its own and is logged | It is easy to undo, low stakes, and uses only approved sources |
| Draft | It is prepared, then waits for a person to approve | It leaves the building or is hard to undo |
| Ask | It stops and asks a named person | Information is missing, confidence is low, or the case is outside the rules |
Two questions place any action: how bad is it if this is wrong, and how easy is it to undo? Tagging a support ticket is Act. Replying to a customer is Draft. Issuing a refund above an agreed limit is Ask.
Most teams start with everything in Draft, which is a sensible place to begin. The real work is moving actions to Act as the evidence earns it, and moving them up to Ask when it does not.
Make the approval screen do the work
A reviewer’s attention is shaped by what the screen puts in front of them. Three details change more than any policy document.
- Show the evidence next to the claim. Not “customer is eligible for a refund,” but the sentence of policy that says so and the record it came from. A reviewer can check that in seconds.
- Show what is unusual. Highlight what differs from the normal pattern and what the system was unsure about. The reviewer’s eye should land there first.
- Make rejecting cheap and specific. One click plus a reason from a short list. Those reasons are the best data you will collect. They tell you exactly what to fix.
If approving takes one click and rejecting takes a paragraph, you have designed a rubber stamp.
Sample instead of approving everything
For actions in Act mode, do not review everything. Review a slice, say one in ten, plus every item the system flagged.
Then watch the error rate in that sample. If it climbs above the level you agreed, move the whole category back to Draft until the cause is fixed. Factories have controlled quality this way for a century. It works because it puts human attention where the evidence says it is needed.
Test whether your reviewers are actually reviewing
Here is the uncomfortable part. Once approval becomes routine, you cannot tell from the outside whether anyone is reading.
So test it. Mix a few deliberately wrong items into the queue, and record which ones they are in a private log. Count how many get caught.
If reviewers catch four out of ten, you do not have a control. You have a click. Treat that as a design problem, not a discipline problem. Tell people that checks like this happen, without saying when, and use the misses to improve the screen, the pace, and the rules.
Write it down where a stranger could read it
For every workflow, keep four lines:
- What it may do on its own.
- What needs approval, and from whom.
- What triggers an escalation, and to whom.
- What it must never do.
Put a name next to each line. Then log what the system did and who approved it. When an auditor, a customer, or your future self asks “why did this happen?”, the answer should take a minute to find, not a week.
Human judgment is not a checkbox you add at the end. It is a resource you place on purpose, where it counts most.
If you want a second pair of eyes on where your workflows put people, an AI Governance Audit tags every AI action with its control and owner. And if you are still choosing your first workflow, start here.