Human in the loop: designing approval gates that actually work

Saying a person is in the loop is easy. It appears in almost every AI proposal, risk assessment and vendor brochure. Making that oversight real is much harder, and the difference matters. A reviewer who sees a hundred items an hour and approves them all is not a control. They are a signature on a process that is, in practice, fully automated.

Good design starts with a simple question: where does human judgement actually change the outcome? Answering it honestly leads to fewer, better-designed review points rather than a person nominally checking everything.

Decide where humans add value

Human review costs time and money, and attention is limited. Spread it across every step and it becomes thin everywhere. Concentrate it where errors are expensive and it becomes a real safeguard.

Place people at decisions that have one or more of these characteristics:

• Costly: a wrong answer leads to a significant financial loss, such as a large refund, a pricing error or an incorrect payment.
• Hard to reverse: once the action is taken, undoing it is slow or impossible, such as a message sent to thousands of customers or a contract signed.
• Legally sensitive: the decision is regulated or could be challenged, such as credit, insurance or employment decisions.
• About people: the outcome affects an individual’s opportunities, wellbeing or rights.

Let automation handle low-risk, high-volume steps on its own, and review a sample rather than every item. Sampling, done properly, gives you evidence about quality without consuming reviewers’ entire day. Increase the sample when quality falls or when something upstream changes.

Consider a typical accounts payable process supported by AI. Invoices under a modest value that match an order exactly can flow through with a weekly sample check. Invoices that do not match, come from new suppliers or exceed a threshold go to a person before payment. Changes to supplier bank details always require two people, because that is where fraud concentrates. Three levels of oversight in one process, each matched to the risk.

Give reviewers what they need

A reviewer can only exercise judgement if they have the information, time and authority to do so. Many approval gates fail on all three.

Context

Show the source material the AI used, its reasoning or the key factors behind its recommendation, and an indication of confidence. A reviewer asked to approve a proposed credit limit should see the application, the relevant history and the factors that drove the recommendation, not just a number and an approve button.

Time

Give reviewers enough time to think. If the throughput target implies thirty seconds per case and a proper check takes three minutes, the target wins and the review becomes a formality. Set targets from the time a careful review actually takes, and adjust staffing accordingly.

Authority

Make it easy to correct and to escalate, and give reviewers clear authority to reject without pressure. If rejecting a recommendation means filling in a long form, explaining yourself to a manager and missing your daily target, people will approve. The cost of saying no must be low.

Avoid rubber-stamping

Rubber-stamping is not a character flaw. It is the predictable result of a well-performing system and a repetitive task. If the AI is right most of the time, reviewers learn to expect it to be right, and their attention drifts. The rare error then slips through precisely because it is rare.

Watch for two warning signs. If approval rates are near 100% and review times are very short, the gate may be decorative. Neither figure alone proves a problem, but together they deserve investigation.

1. Rotate reviewers so that nobody spends a whole day on the same repetitive check.
2. Audit a sample of approved cases with a second, independent reviewer.
3. Include known test cases where the right answer is to reject, and track whether reviewers catch them.
4. Share the results with reviewers, framed as learning rather than blame.

Test cases are particularly effective. Seeding a small number of deliberately flawed recommendations into the queue keeps reviewers alert and gives managers a direct measure of whether the gate works.

Explain the purpose openly. Reviewers who understand that seeded cases measure the design of the gate, not their personal performance, tend to welcome them. Used as a disciplinary tool, they breed resentment and gaming.

Close the loop

Every correction and override is information. Capture each one with a short reason: the AI misread the document, used outdated policy, missed a relevant fact or made a judgement the reviewer disagreed with. Over a few weeks, these reasons reveal patterns.

Use them to improve prompts, rules and training data, and report override rates to the owner of the process. A falling override rate, alongside stable audit results, is evidence that the system is improving and may justify a lighter review level. A rising rate signals that something has changed and needs attention.

Be wary of reading too much into a very low override rate on its own. It may mean the system is excellent, or it may mean reviewers have stopped looking. Pair it with independent audit results before relaxing any control.

The override log turned out to be the most useful document in the project. It told us exactly where the model was weak.

Claims operations manager, insurance client

A quick design test

For each AI step in a workflow, ask four questions. What is the worst realistic error? Who would notice it? How fast? And how would it be undone? If nobody would notice, or would notice only after the damage was done, add a gate. If the error would be caught quickly and reversed easily, sampling may be enough.

Run this test with the people who do the work, not just the designers. They know where errors hide and what a realistic review looks like on a busy day. Then measure the gates you build, and adjust them as evidence accumulates. Oversight that is designed, resourced and measured is a real control. Oversight that is only written into a policy is not.

Source: Intelliversal analysis.

Frequently asked questions

Does every AI decision need human approval?

No. Reserve approval before action for costly, irreversible, legally sensitive or people-related decisions, and use sampling and audit for low-risk, high-volume work.

How can we tell if reviewers are just rubber-stamping?

Look for near-total approval rates with very short review times, audit samples independently and seed known errors to see whether reviewers catch them.

What should we do with override data?

Record a reason for each override, look for patterns and use them to improve prompts, rules and data. Report override rates to the process owner.

Explore Generative AI, copilots and agentic automation