Add a human approval gate to an AI automation
How to keep a person in the loop without making them the bottleneck they were before.
Before you start
- A working automation
- Someone who will do the approving
What you will be able to do
- Route only the cases that need a human
- Design an approval that takes seconds rather than minutes
- Detect rubber-stamping before it costs you
Once an automation does something visible — sends a reply, issues a refund, posts publicly — someone reasonably wants a human check.
Done naively that means approving everything, which reinstates the bottleneck and, within two weeks, produces approvals that are not really being read.
Decide what a gate is protecting against
Irreversible, public or expensive. Everything else can be corrected after.
Gate the actions where being wrong is costly and hard to undo: money out, anything customer-visible, anything deleted.
Internal drafts, tags and summaries usually do not need one — a wrong tag is fixed in seconds. Gating those is what trains people to click approve without reading.
Route on confidence, not on everything
Ask the model how sure it is and send only the uncertain cases.
Have the AI step return a confidence alongside its answer and route only the low-confidence cases to a person. Self-reported confidence is imperfect and still correlates usefully with being wrong.
Combine it with consequence: high confidence on a small refund goes through; high confidence on a large one still gets seen. Two axes, four boxes, and only two of them need a human.
- Trusting a confidence score you have never calibrated. Sample fifty auto-approved items after a month and check the ones it was sure about.
Make approving take five seconds
The decision, the reason, and both buttons — in one message.
The approval message needs the proposed action, the input it was based on, why the model chose it, and approve/reject in place. If a reviewer has to open another system to judge, they will approve without doing so.
Include a one-line reject reason field. It is the only data you will ever get about why the automation is wrong, and it is worth the extra field.
Watch the approval rate as a signal
98% approval means the gate is theatre or the routing is wrong.
Track approvals against rejections. A rate near 100% means either your routing is sending safe cases to a human, or people have stopped reading. Both mean the gate is not doing anything.
A healthy gate rejects a real fraction. If it does not, tighten the routing so fewer, harder cases reach the person — that is what makes their attention worth having.
- Set a timeout: if nothing is approved within a set window, escalate or fail closed. Silent queues are how a gate becomes an outage.
Route by confidence and consequence, show enough to judge in seconds, and watch your approval rate. A gate approving everything is not a gate.
Common questions
Was this guide useful?
93% of readers found this useful
Read next
Connect two apps with an AI step in the middle
The genuinely useful automations are not the clever ones. They are a trigger, one AI step that makes a small judgement, and a writ…
Automate a repetitive task with AI in an afternoon
Not every repeated job is worth automating, and the ones that are follow a pattern. This walks through choosing one and building i…
Know when an agent is the wrong answer
A surprising share of agent projects would work better, cheaper and more reliably as a script with one model call in it.
Write SQL with AI against a schema it has never seen
A model will write syntactically perfect SQL against a schema it is guessing at. The query runs, returns numbers, and the numbers…
