A human in the loop is not automatically a control
Model risk management in banking taught me a lesson many AI programs are now relearning: a review step is only a control if it can catch the error that matters. Examiners did not accept “someone looks at it” as evidence. They wanted to know what the reviewer checked, what authority they held and what happened when they disagreed. I apply the same test to every human review step in an enterprise AI workflow.
The reviewer needs a specific decision: accept a summary, approve a communication, authorize an action or request more evidence. A vague instruction to “check the AI” leaves both the person and the implementation team without a usable standard.
Match the review to the consequence. An internal draft may need a light accuracy check. A credit decision, a client communication in private banking or a result release in a diagnostics lab needs stronger evidence and explicit authority. Set those requirements for each workflow rather than putting the same approval screen on every task.
Design for real, limited attention
Reviewers are busy. If the system asks for approval constantly or repeats low-value warnings, people start approving by reflex. Treat that as a design problem, not a character flaw. I have watched well-intentioned control designs train people to click through, which is worse than no control because it creates false assurance.
Give the reviewer enough context to judge the proposal without reconstructing the conversation: scope, reporting period, source evidence and material uncertainty. Put what changed, what is uncertain and what the action would affect at the top. A long explanation is not a useful one if the decisive fact is buried. Avoid decorative confidence indicators that carry no clear operational meaning.
Test the review experience with the actual reviewers. Something a technical specialist reads easily may confuse a relationship manager or a clinical operations lead. Observe whether they can spot a planted error, explain the evidence and choose a next step without coaching.
Measure the review, not just the output
A high approval rate can mean excellent output or superficial review. Track review effort, rejection reasons and whether important errors are caught, so you can tell the difference.
Include cases where the right decision is to defer to another source, a definition owner or a higher authority. If deferral is cumbersome and approval is one click, the interface is quietly encouraging premature decisions.
Do not assign review to someone without business authority. Familiarity with a system is not authority to approve a transaction or interpret financial policy. The operating model must name both subject matter competence and decision responsibility.
Scale review with the institution and the risk
In wealth management, an advisor reviewing a client summary brings judgment they already use in their work. The review design should help them apply it: show the relevant evidence, make omissions visible and leave time to challenge the draft. Record the review standard and test whether the person can detect the failures the workflow is likely to produce. Adding an approval button does not answer that question.
Reassess whenever the workflow changes. More users, new sources or broader actions alter the consequence of an error. A review process sized for a small drafting pilot can be inadequate for a widely used operational service.
Use review to make the system better
Repeated corrections to the same definition or source mapping should flow back into the product and data process. Human review is not a permanent excuse for known defects. The review record tells you which improvements would remove unnecessary effort while preserving meaningful control.
Before rollout, run a short reviewer exercise with three proposals: one correct, one plausible but unsupported and one incomplete. Ask the reviewer to explain each decision. If all three look equally credible in the interface, fix how evidence and limitations are presented.
The goal is an assistant that removes preparation work while leaving the person a clear, realistic opportunity to exercise judgment. Test that opportunity with real reviewers before expanding the workflow.