Why does a human review step often fail?
A human review step often fails because the person sees only the AI's verdict, has too many items to check and is never told what to look for, so approving becomes the fastest path. This is called rubber-stamping, and it gives the appearance of oversight without the substance.
Picture a fictional 25-person wholesale distributor, Oakbridge Supplies. An AI tool matches supplier invoices to purchase orders and marks each one 'match' or 'mismatch'. The finance clerk, Priya, approves the matches for payment. Her queue on a Thursday reads: '212 invoices, 205 match, 7 mismatch'. Each line shows supplier, amount and a green tick. Last month a supplier invoice for 400 units was paid when only 40 were delivered, because the tool matched the order number and Priya approved the tick.
Where should a human be in the loop?
A human should be in the loop wherever an AI output changes money, a person's access to something, a customer commitment or a legal or safety outcome, and before that change becomes hard to undo. Routine, reversible outputs can be checked by sampling instead.
The EU AI Act's article on human oversight for high-risk systems makes the same point in regulatory language: the people overseeing a system should understand its limits, stay aware of automation bias, and be able to disregard, reverse or stop its output. Even if the Act does not apply to you, those are good design goals.
Which human in the loop review designs are there?
There are four common review designs, and the right one depends on the cost of an error and how many items there are.
Approve before action
A person approves each output before anything happens. Use for high-cost, low-volume work such as payments over a threshold or customer refunds.
Exception review
The AI handles clear cases and routes uncertain or unusual ones to a person. Works only if the routing rules are tested and some 'clear' cases are still checked.
Sampled audit
Outputs go out, and a person checks a random sample afterwards. Use for reversible, high-volume work such as internal summaries.
Independent first
The person forms their own view before seeing the AI suggestion. Slower, but it protects against anchoring on the AI answer for important judgements.
What does a weak and an improved review design look like?
Weak design: every invoice goes to Priya with a green or red mark. She approves 205 ticks in 20 minutes. Nothing tells her why each tick was given, and nothing tests whether she is still looking.
Improved design: invoices under a set amount where order number, quantity and unit price all match the delivery note are paid automatically, and Priya checks a random ten each week. Anything above the threshold, any new supplier and any quantity difference goes to her with the three fields side by side and the AI's reason in one line, such as 'quantity on invoice 400, delivered 40'. Once a month, the finance lead quietly includes two known errors in the queue to confirm they are caught. Priya also records each time she overrides the AI. A sudden drop to zero overrides is treated as a warning, not a success.
How do you check that review is real and not rubber-stamping?
Check that review is real by measuring it: time spent per item, how often reviewers override the AI, and whether planted test errors are caught. A reviewer who never disagrees with the tool is either reviewing very good output or not reviewing at all, and you need evidence to know which.
Show evidence, not just verdicts
Put the source fields beside the AI answer so checking is fast.
Limit the queue
If a person cannot properly review the volume, change the design rather than asking them to try harder.
Give authority
Reviewers must be able to reject or pause without needing permission.
Test the loop
Include known errors from time to time and record whether they are caught.
What are the common human in the loop mistakes?
The most common mistake is adding a reviewer to a process without changing anything else, then treating their approval as proof the output was correct. Others: asking one person to approve hundreds of items a day, hiding the evidence behind extra clicks, reviewing after the action has already happened for irreversible decisions, and celebrating a falling override rate without asking why. Automation bias affects careful people too, which is why the design matters more than good intentions.
Where does human in the loop fit in the Tosheo.ai catalogue?
Human in the loop sits in the Responsible AI and Policy category, in the AI literacy course, after ethical AI and responsible AI controls. It returns in the agentic AI and automation lessons, where tools act rather than suggest. The hands-on task asks you to redesign a review step from your own work, and the rubric feedback checks that the review point, evidence and authority are all clear.