Consequence first, confidence second
The common design routes low-confidence cases to a person. Necessary, insufficient. The case that hurts is the confident wrong answer on a large amount, and confidence alone will wave it through.
Start from what the action does. Reversible and small, let it run. Irreversible or large, gate it whatever the confidence. Then layer confidence on top, then novelty: a vendor never seen, a category never used, an amount far outside the distribution.
A question that gets a real answer from a client: what is the largest amount you would let a new joiner post unsupervised in their first week? That number is a defensible starting threshold and they can answer it in a meeting.
The review has to be faster than the work
The whole pattern rests on reviewing being much cheaper than doing. A gate showing a decision and two buttons trains people to approve everything inside a week, and then you have the cost of a gate and the risk of none.
- The decision and the reason in one sentence.
- The evidence positioned, not linked. The page with the figure highlighted.
- What it will do if approved, concretely, including the amount and the target.
- What was unusual about this case, since that is why it is in front of them.
- One click to approve, one to correct, and correcting must not mean opening another system, or the correction is never recorded and you lose the feedback.
Design it to loosen
A gate that never moves is a permanent tax on the workflow. Agree the ratchet before go-live, while everyone is calm.
- Gate everything for the first fortnight. Treat it as a supervised run that produces evaluation data.
- Measure the override rate per category. Categories running at 99% approval are candidates.
- Release one category at a time, and keep sampling one in twenty so you retain a signal.
- Write the threshold down in advance: "below 2% override across 500 cases and this category automates" is a decision made calmly rather than under pressure.
Be equally willing to re-gate. A category that has a bad week goes back behind the gate, and treating that as a normal operational action rather than a failure is what keeps people honest about the numbers.
The four ways it goes wrong
- Rubber stamping
- Approval near 100%, review time near zero. Sample it. If it is theatre, either remove it or make it informative; leaving it is the worst option.
- The unowned queue
- Items waiting days. Worse than no gate: work stalled and everyone assuming somebody else has it. Alert on age and name an owner per shift.
- Everything escalates
- Thresholds set from fear rather than policy, so most cases gate and the business case evaporates.
- No record
- An approval not stored with who and when means you added a step and gained nothing auditable.
Want us to run this with you?
The Audit is this method pointed at your systems, with a costed build plan at the end of it.
Schedule call