How many gates is too many?

The threshold is behavioural rather than numerical: too many is the point at which the approver develops a routine for clearing them. That is usually obvious to the person doing it and invisible in every report, so it is worth asking them directly.

What should an approval screen show?

What prompted the action, what the system is about to do, what happens next, and what the alternative was. Without those an approver can only confirm that something is about to happen, which is a signature rather than a decision.

What does a very high approval rate mean?

One of two opposite things: the automation is reliably right and the gate could move somewhere more useful, or nobody is really reading it any more. Those need distinguishing, and the cheapest way by far is to ask the approver what the last thing they rejected was.

Glossary

Approval gate

A point where a person must confirm before an automated process continues, placed where a mistake would be costly rather than everywhere.

In plain terms

A stop where somebody has to say yes. It only works if the person is asked rarely enough to still be paying attention and shown enough to have an opinion, and most gates fail one of those two tests within a month.

Why it matters

Because a gate nobody reads is worse than no gate at all. It is visible in the design, it appears in the audit trail, and it persuades everybody that a decision is being made at a point where one has stopped being made.

How it works

Frequency decides whether it survives. A person asked a handful of times a day considers each one; a person asked two hundred times develops a rhythm for clearing them, and the rhythm is indistinguishable from approval in every record.

What the approver sees decides whether they can judge. A gate showing only the action about to be taken asks for a signature; one showing what prompted it, what the alternative was and what will happen next asks for a decision.

Place them by consequence rather than by step. Gating everything is the instinct and it guarantees the fatigue, whereas gating the actions that are expensive or hard to undo leaves few enough that each one is genuinely read.

Rejection has to be as easy as approval, and it usually is not. Where saying no means writing an explanation, finding somebody or blocking a colleague, the design has quietly made approval the path of least resistance.

Watch the approval rate, because it tells you what the gate has become. A rate approaching everything means either that the automation is right almost always, in which case the gate can move, or that nobody is reading, and those two need to be told apart.

Timing matters as much as placement. A gate that arrives while somebody is already doing the related work is read; one that arrives as a notification during unrelated work interrupts a person who then clears it to get back to what they were doing, which is the same failure arriving by a different route.

A gate is not a substitute for narrowing what the system can do. Confirmation covers the actions somebody thought to gate; access covers everything else, and the two are complements rather than alternatives.

Two gates that look identical in the log

The third line on each side is the one most often overlooked and the cheapest to fix. Approval and rejection are presented as symmetric choices and almost never are: approving continues the process silently, and rejecting means writing a reason, notifying somebody, or leaving a colleague waiting while the case sits unresolved. Faced with a borderline item at the end of a long day, a reasonable person takes the path that does not create work, and they are not being careless, they are responding to what the interface made easy. Making rejection cheap is usually a small change - a default reason, a route that sends the case somewhere sensible rather than nowhere - and it does more for the quality of the decisions than tightening any of the wording on the screen. The general form is worth remembering beyond this one control: whenever a system asks a person to choose, look at what each answer costs them, because that is what will decide the answer.

Seen in the wild

An agent that drafts a reply and waits before anything reaches a customer.
Sierra
An automation that pauses only where a record would be changed irreversibly.
Make
A coding agent that proposes a change and requires a person to apply it.
Claude Code

Common misconceptions

People assume

More gates means more safety.

In fact

Past a certain frequency each additional gate reduces the attention available for all of them. A person clearing two hundred confirmations a day is not exercising judgement two hundred times; they have developed a habit.

People assume

A gate in the log means a decision was made.

In fact

It means somebody clicked. Whether a decision happened depends on how often they are asked and what they were shown, and neither of those is recoverable from the record afterwards.

Questions

How many gates is too many?: The threshold is behavioural rather than numerical: too many is the point at which the approver develops a routine for clearing them. That is usually obvious to the person doing it and invisible in every report, so it is worth asking them directly.
What should an approval screen show?: What prompted the action, what the system is about to do, what happens next, and what the alternative was. Without those an approver can only confirm that something is about to happen, which is a signature rather than a decision.
What does a very high approval rate mean?: One of two opposite things: the automation is reliably right and the gate could move somewhere more useful, or nobody is really reading it any more. Those need distinguishing, and the cheapest way by far is to ask the approver what the last thing they rejected was.