"Human in the loop" has become a phrase people use to reassure regulators and boards. Too often it means a person clicks "approve" at the end of a process they don't understand, on an output they have no time to check.
That isn't oversight. It's a rubber stamp with a name attached.
Why the checkbox version fails
When the human step is added at the end, three things go wrong.
First, reviewers lack context. They see a recommendation but not the evidence behind it, so they either trust it blindly or redo the work from scratch.
Second, volume overwhelms judgement. If an agent produces hundreds of outputs a day and every one needs a click, people stop reading.
Third, nobody learns. Reviewer decisions disappear into an approval log, so the agent never gets better and the organisation never sees where it disagrees.
Design the review, not just the agent
Good human-in-the-loop design starts with a question: which parts of this decision need human judgement, and what does the person need to exercise it well?
Answering that usually leads to four design choices.
1. Show the evidence, not just the answer. Every finding should open next to the document, clause or record it came from. A reviewer should be able to check a finding in seconds.
2. Route by risk. Not every output needs the same scrutiny. Low-risk, high-confidence items can be sampled; consequential or uncertain ones go to an experienced reviewer. Your authority limits already encode this. Use them.
3. Make disagreement easy. Reviewers should be able to change or reject a finding with a reason in one step. Those reasons are the most valuable data you will collect.
4. Close the loop. Feed reviewer decisions back into the agent's checks and into your procedures. Over time, the agent should get closer to how your best people think.
What this looks like in practice
Take a credit review. A checkbox design gives the credit officer a draft paper and an approve button.
A designed review gives the officer a pack where each finding (falling revenue, covenant pressure, a missed payment) links to the page in the accounts or the line in the account history. The officer adds their judgement on what the numbers don't show: the client's new contract, the director's track record. The approver sees both the agent's evidence and the officer's reasoning. Every change the officer makes is recorded and reviewed monthly to improve the agent.
The officer is faster, the decision is better documented, and the bank learns where its people and its agent see things differently.
The regulator's view
Regulators increasingly expect firms to show that people can meaningfully oversee AI systems, not just that a human signed off. Meaningful oversight means the person understood what they approved, had the evidence to challenge it, and had the authority to change it.
A designed review produces exactly that record. A checkbox produces a signature.
Where to start
If you already have an AI pilot, look at its review step. Ask your reviewers three questions: Can you see why the agent said what it said? How long does checking take? What happens when you disagree?
Their answers will tell you whether you have human-in-the-loop by design, or by checkbox.



