Precision vs Recall: How to Measure AI Accuracy in a Workflow

01Precision vs recall in one business example

For example, suppose a 15-person software company uses an AI step to flag incoming support tickets as urgent so they jump the queue. Two different things can go wrong. The AI can flag a routine ticket as urgent, so someone drops other work for nothing. Or it can miss a ticket that really was urgent, so a customer with an outage waits in the normal queue.

Precision tracks the first problem by asking what share of flagged tickets really were urgent. Recall tracks misses: out of all truly urgent tickets, it gives the share the AI caught. Scikit-learn's documentation describes precision as a classifier's ability not to label a negative sample as positive, and recall as its ability to find all the positive samples. You need both numbers, because improving one usually costs you some of the other.

02The confusion matrix behind precision and recall

Every evaluation starts from four counts. Compare the AI's label for each ticket with the correct label a person assigned, and sort each ticket into one of four boxes. Google's Machine Learning Crash Course calls this layout a confusion matrix. The numbers below are invented for illustration: 500 labeled tickets, of which 40 were truly urgent.

• True positive: an urgent ticket the AI correctly flagged.
• False positive: a routine ticket flagged as urgent, which is a false alarm.
• False negative: an urgent ticket the AI missed.
• True negative: a routine ticket correctly left in the normal queue.

03Precision and recall formulas, with F1

Precision divides true positives by everything the AI flagged, while recall divides them by every ticket that was truly positive. Both formulas appear in the scikit-learn and Google references listed below.

F1 is the harmonic mean of precision and recall. Google's course notes that when the two are far apart, F1 lands close to the worse one, so a system can't earn a good F1 by excelling at one and failing at the other. If misses matter more than false alarms, the F-beta score weights recall more heavily. Scikit-learn's documentation explains that a beta of 2 makes recall twice as important as precision, and a beta of 0.5 does the opposite.

• Precision = TP / (TP + FP) = 34 / 50 = 0.68. About two in three flags were real.
• Recall = TP / (TP + FN) = 34 / 40 = 0.85. The AI caught 34 of 40 urgent tickets and missed 6.
• F1 = 2TP / (2TP + FP + FN) = 68 / 90 = 0.76, rounded.
• Accuracy = (TP + TN) / all tickets = 478 / 500 = 0.956.

04Precision vs recall vs accuracy: why accuracy can mislead

In the example, accuracy is 95.6 percent, which sounds excellent. Now picture a system that never flags anything. It would be right about all 460 routine tickets and wrong about all 40 urgent ones, for an accuracy of 92 percent and a recall of zero. Its precision can't even be calculated, because it made no positive predictions.

Google's course makes the same point: when one class is rare, a model that always predicts the common class can score well on accuracy and still be useless. Urgent tickets, duplicate invoices and contracts that need legal review are often a small share of the total. For cases like these, report precision and recall, and treat accuracy as a supporting number at most.

05How the threshold trades precision against recall

Many AI classifiers produce a score, and someone picks the cutoff above which an item counts as positive. Google's course explains that raising this classification threshold usually means fewer false positives and more false negatives, and lowering it does the reverse. That's why precision and recall often move in opposite directions.

The table runs the same 500 tickets, 40 of them urgent, at three made-up thresholds. The F2 column weights recall twice as heavily as precision. If your tool returns labels without scores, you can still shift the balance by changing its instructions, its examples or the rule for what goes to review, then measuring again.

06When to favor recall and when to favor precision

Google's guidance is brief: use recall when false negatives cost more than false positives, and precision when it's very important that positive predictions are right. In a business workflow, the deciding question is what happens after the label.

If a flagged item goes to a person who can dismiss a false alarm in a minute, false positives are cheap and misses are the expensive error, so favor recall. If a flag triggers an action that's hard to undo, such as an automatic refund, a closed ticket or a merged customer record, a false positive is the expensive error. Favor precision there, or keep a person between the flag and the action.

• Favor recall: urgent ticket triage, flagging possible duplicate invoices for review, finding contracts with an auto-renewal clause and spotting safety complaints.
• Favor precision: auto-closing tickets as spam, auto-approving matched invoices, sending customer messages without review and merging records.
• Balance both: routing tickets between departments when each misroute costs about the same time to fix.

07Turn the confusion matrix into review time and cost

F1 counts a false alarm and a miss as equally bad, which often isn't true in a business workflow. Put a cost on each error type instead. In this illustrative case, assume a support lead needs three minutes to check and dismiss a false alarm, and a missed urgent ticket costs 45 minutes of extra recovery work once the customer escalates, before counting any harm to the relationship. Both figures are invented, so time your own.

On those assumptions, the high threshold that narrowly won on F1 is the most expensive option, and the low threshold that F1 ranked last costs the least. The workflow owner should choose the operating point from costs like these, with the metrics as inputs. Once you've chosen, fold these minutes into a full review-time measurement so you can compare the workflow with the manual baseline.

08Build a labeled test set before you trust any score

Precision and recall are only as reliable as the correct answers you compare against. NIST's AI RMF Playbook calls for documenting test sets and metrics, and for measuring performance in conditions similar to the deployment setting. For a small business, that means a labeled sample of your own recent work. A vendor's demo results describe someone else's data.

Watch the size of the rare class. With 40 urgent tickets, each miss moves recall by 2.5 percentage points, and a test set with only five urgent tickets can't tell you much. If urgent items are scarce, pull extra examples from older records instead of labeling thousands more routine ones, and note that this changes the mix, so precision on the test set won't match live traffic exactly.

• Pull a recent, representative batch, including busy periods, messy inputs and the edge cases that cause disagreements.
• Have someone who knows the work assign the correct label before seeing the AI's output.
• Write down the labeling rule, such as what counts as urgent, have a second person label a subset and settle disagreements against the rule.
• Keep the test set separate from any examples used to write prompts or tune the system.
• Run the test again after any change to the model, prompt, threshold or input source.

09Precision and recall for document field extraction

The same idea applies when AI pulls fields from invoices or forms, but you have to define what counts. One workable convention: a correct value is a true positive, a filled-in value that's wrong or absent from the document is a false positive, and a value present in the document that the system left blank is a false negative. Decide in advance how to count a wrong value where a correct one existed. Counting it as both a false positive and a missed value is the stricter choice.

Measure each field separately. For example, invoice number and total might score well while due dates lag. In extraction, a blank is usually safer than a confident wrong value, because a reviewer can see a blank and fill it in. That argues for favoring precision on fields that post to your books and sending uncertain fields to review.