Approval gates: designing the human step in an autonomous workflow
Put the human in the wrong place and you get rubber-stamping. Here is how to choose which steps genuinely need a person.
“Human in the loop” is easy to say and easy to get wrong. Ask a person to approve everything and within a fortnight they are clicking approve without reading, which is worse than no gate at all — now there is a signature on a decision nobody made. The design problem is choosing the few steps where a human genuinely changes the outcome.
Gate on consequence, not on uncertainty
The instinct is to gate wherever the model is unsure. That produces gates in the wrong places: lots of interruptions on trivial calls, none on the confident-but-wrong ones. Gate instead on what the action does. Anything that leaves the company, moves money, changes a customer record, or cannot be undone gets a person. Everything else runs, and gets logged.
Make the approval cheap to give
A good gate shows the decision, the evidence behind it, and what will happen next — on one screen, in a few seconds. If approving requires opening three systems to check the agent’s work, the gate will be bypassed, informally, by people under time pressure. The quality of a gate is measured by how quickly a competent person can reject a bad one.
Batch where you can
Twenty individual approvals scattered through a day cost more attention than one review of twenty items with the outliers surfaced at the top. Batching also gives the reviewer comparison, which is where people are genuinely better than models — spotting the one that does not look like the others.
Log the rejections
Approvals tell you little. Rejections are the most valuable data an agent produces: they are labelled examples of the system being wrong, generated by the people who know. Feed them into the evaluation set. A workflow that has been running six months should be measurably better than at launch, and if it is not, the rejection log is where to look.
Plan for removing gates
A gate that never rejects anything is a tax. Review them: if a step has run for months with a rejection rate near zero, either loosen it to spot-checking or remove it and rely on the audit log. Autonomy should be earned incrementally and visibly, not granted at launch because the demo went well.
The line we hold
Every agent gets a defined scope of authority, approval gates on anything consequential, a complete action log, and a way to revert. Not because autonomy is dangerous in the abstract, but because those four things are what let a business expand it later with a straight face.