Spoken script
Someone says an AI reviewer is 90% accurate. Before deciding whether that is useful, ask what was counted.
Were they counting documents, findings, sentences, or user ratings? Did the test include clean documents? Were difficult cases excluded? Were the most consequential misses visible?
Here is a fictional issue-detection example. There are 40 reference issues. The assistant finds 32 and misses 8. It also produces 6 false findings.
Its recall is 32 out of 40, or 80%. Its precision is 32 out of 38 proposed findings, or about 84%.
Now suppose 10 of the reference issues were critical and the assistant missed 2. The critical-issue recall is also 80%, but that result needs its own attention. The ordinary findings do not make those misses disappear.
The small test set also limits what we can infer about future performance. And the model result is still not the final human-review result.
Define the task, count consistently, report severity and uncertainty, and evaluate the combined workflow. A percentage becomes useful when you can explain its denominator and its relationship to the decision.
Viewer exercise
Would a shorter finding list necessarily be better? It could improve precision while reducing recall. Examine both and consider what users do with the result.