Put a person in front of a queue where ninety-nine of every hundred items are correct and they will start approving without reading — not through carelessness, but because that is what attention does when the base rate is that lopsided. The review still exists on the process diagram. It has simply stopped functioning, and now it provides false assurance, which is worse than having no step at all.
This is why "a human checks everything" is usually a warning sign rather than a safeguard. It sounds responsible and it converts, within weeks, into a bottleneck that also does not catch anything. The useful question is not whether a person is involved, but which specific cases reach them and what they can actually see when one does.
Below the confidence threshold. The system knows when it is unsure, and those cases go to a person with the uncertain field highlighted. This is the workhorse, and it is the one that keeps the queue small.
Above a materiality line. Anything over a certain amount, or affecting a certain customer tier, is checked regardless of confidence. Not because the system is likelier to be wrong, but because the cost of being wrong is higher.
Anything irreversible. Payments leaving, contracts sent, records deleted, messages to customers who are already unhappy. If undoing it means an apology, a person presses the button.
A sample of the routine. A small random share of the cases that passed cleanly, reviewed to detect drift. This is the only step that catches the system getting quietly worse, and it is the first one people cut.
Together these usually add up to a small fraction of volume. That is the point — a queue a person can genuinely read is worth more than a queue they scroll past.
A reviewer needs to answer one question fast: is this right? That requires the source next to the answer — the invoice image beside the extracted total, the customer's actual message beside the drafted reply, the three signals beside the risk score. If checking requires opening another system, the review will be skipped under pressure, and pressure is the normal state.
It also needs the disagreement to be visible. Show what the system was unsure about and why it flagged this case, rather than presenting a finished answer for silent approval. And make rejecting cheap: one click, no mandatory justification field. A reviewer who has to write a paragraph to disagree will stop disagreeing.
Change rate. How often the reviewer alters what the system produced. Near zero means the threshold is too cautious and you are wasting the person; high means the system is not ready for that case type.
Time per item. If it is a few seconds, nobody is really checking. If it is many minutes, the screen is not showing enough.
Escaped errors. Wrong outcomes that got through review. This is the number that tells you whether the step is real oversight or a formality.
Queue age. How long flagged items wait. A review step that adds two days to every exception has costs the business case did not include.
The first weeks in production should have more review than the design calls for — wider thresholds, more sampling, a person looking at things that will later pass automatically. That period is how you earn the numbers that justify narrowing it. Start tight and loosen with evidence, rather than starting loose and tightening after an incident.
Then narrow deliberately. Each category that has run clean for a defined period moves out of mandatory review and into sampling. Some categories never leave — the irreversible ones, the large ones — and that is a design decision rather than a failure to automate. A process where twenty per cent of cases reach a person, and that twenty per cent is genuinely read, is in much better shape than one where a hundred per cent are nodded through.
Whoever does the reviewing should have been involved in defining the exceptions. The people who know which cases are dangerous are the ones doing the work today, and they will tell you in the first conversation if asked.
We use cookies for analytics — to see which pages bring enquiries. Nothing else, and nothing before you agree. Cookie Policy