An approval step has a 99.8% approval rate. What does that tell you and what would you change?

What is being tested

Whether you read the override rate as a diagnostic, and whether you know both directions it can point.

What it tells you

One of two things, and you must determine which:

1. The control is theatre. The human cannot meaningfully evaluate what they are approving — too much volume, insufficient context, or no realistic ability to disagree. They are a latency source with a compliance function, and the control provides no protection while creating the belief that it does.

2. The automation is good enough to trust. In which case the human step is unnecessary friction and should be removed or narrowed.

Both conclusions are actionable and they are opposite, so the diagnosis matters.

How to tell which

Sample the 0.2% that were rejected. Were those genuine catches that mattered? If yes, the control works and the question is whether it can be narrowed to the cases that resemble them. If they were trivial or arbitrary, the control is theatre.

Ask the approvers what information they have. If they cannot see what they would need to decide, the interface is the problem, not the person.

Check the time spent per approval. A hundred approvals an hour is not evaluation.

What to change

Escalate by exception, not by default. Automate the confident majority; route only the uncertain minority to a human — who then has time to actually think.

Give the human what they need to decide: the reasoning, the evidence, the alternatives, and what happens if they approve or reject.

Present evidence rather than a recommendation with a confirm button. Automation bias is real: people over-trust automated recommendations, particularly under time pressure, and the interface determines the behaviour.

Make disagreement easy and record it. Those cases are the training signal, and if overriding is difficult or discouraged the human is not in the loop.

Measure the override rate continuously. Near zero means theatre or trustworthy automation; high means the automation needs work. Either way it is informative, and it is rarely measured.

Where the human genuinely belongs

Irreversible or high-consequence actions, low-confidence decisions, novel situations, where law requires it, and where accountability must rest with a person — including confirming consequential actions taken by an agent that processes untrusted content, where the control is primary rather than procedural.