Skip to content

Stockertim Dictionary · Human resources · Technology

Override rate

Also called human override rate, reversal rate

In short

The share of automated recommendations that human reviewers change or reject; read alongside review times, it shows whether oversight is real.

  • Advanced
  • Ratio or metric
  • Checked

In detail

The override rate is a simple indicator of how a human-in-the-loop process actually works. Count the recommendations a system made in a period, count how many reviewers changed or rejected, and divide. It is usually tracked by reviewer, by team, by decision type and by group of people affected.

There is no single right level. A rate close to zero over thousands of decisions suggests reviewers may be deferring to the tool, the pattern called automation bias, or that they lack the information or authority to disagree. A very high rate suggests the tool is not useful or is badly matched to the task. A rate that differs sharply between reviewers, or between groups of applicants, invites a closer look.

Because overrides can be wrong too, organisations look at outcomes: were overridden decisions later confirmed as better? Combining the override rate with time spent per review and a periodic sample of cases checked by a second person gives a fuller picture of whether oversight is meaningful.

Formula

Override rate = recommendations changed / recommendations reviewed

recommendations changed
outputs the reviewer altered or rejected
recommendations reviewed
all outputs that reached a reviewer in the period

Worked example

Reviewers at a made-up firm, Fernhill Services, checked 4,000 automated shortlisting recommendations and changed 20. After training and a redesign, they changed 240 of the next 4,000.

  1. Before: 20 / 4,000 = 0.005
  2. After: 240 / 4,000 = 0.06

The override rate rose from 0.5% to 6%, a sign that reviewers now engage with each case.

In practice

AI governance dashboards report override rates monthly next to review times, and a sudden fall towards zero triggers a check of reviewer workload and training.

Watch out

  • A low rate is not proof the tool is right: it may mean nobody is checking.
  • Counting overrides without checking their quality can reward random disagreement.

Learn it properly

Official sources

Facts last checked on 3 November 2026.

Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.

English