In one sentence

Precision is how many of the things the AI flagged were actually correct; recall is how many of the real cases it managed to catch.

The simple model

Imagine a step that flags suspicious invoices. It will make two kinds of mistake:

  • It flags an invoice that was fine. That hurts precision: not everything it flagged was a real problem.
  • It misses an invoice that was actually suspicious. That hurts recall: it did not catch every real case.

Precision asks: when it says yes, how often is it right? Recall asks: of all the real yes cases, how many did it find?

Why you care

These two are usually in tension. You can catch almost every real case (high recall) by flagging generously, but then you flag many innocent ones (low precision). You can be very sure each flag is right (high precision) by flagging only obvious cases, but then you miss subtle ones (low recall).

Which matters more depends entirely on the cost of errors:

  • If missing a real case is dangerous, a missed fraud, a safety issue, favour recall, even at the cost of more false flags for people to check.
  • If false flags are expensive or annoying, favour precision, accepting that some real cases slip through.

Tune the balance deliberately

You often adjust this balance with a confidence threshold: flag more aggressively for higher recall, or more cautiously for higher precision. The right setting is a business decision, not a technical one, because it depends on which mistake costs more.

A small example

A team flagging urgent support tickets first tuned for high precision, so every flag was reliable, but missed a third of the genuinely urgent ones. Since a missed urgent ticket was the costly error, they shifted toward recall, catching almost all urgent tickets and accepting a few extra flags for a person to dismiss.

What this does not solve

Precision and recall describe how a detection step errs; they do not tell you the cost of each error or the right trade-off. That comes from the business context. And they apply to yes-or-no style tasks, not every kind of AI output.