The workflow at a glance
Define the positive class, then reconstruct the four outcome counts. With TP=8, FP=12, FN=2 and TN=78, accuracy is 86%, precision 40% and recall 80%. Correct arithmetic does not establish deployment suitability; missed cases, false alarms and independent acceptance evidence still matter.
Define urgent as the positive class
This is an original editorial support-routing exercise, not a product benchmark or real customer dataset. Human reference labels identify ten urgent and ninety routine tickets. The classifier sends twenty to the urgent queue, eight of which are actually urgent.
The Google classification metrics lesson supplies the metric definitions. All calculations here use one fixed threshold and the same sample, without mixing counts from another threshold.
| Actual label | Predicted urgent | Predicted routine | Row total |
|---|---|---|---|
| Urgent | 8 (TP) | 2 (FN) | 10 |
| Routine | 12 (FP) | 78 (TN) | 90 |
| Column total | 20 | 80 | 100 |
Four metrics answer different questions
Accuracy=(8+78)/100=86% asks how many tickets were classified correctly overall. Precision=8/(8+12)=40% asks how many urgent predictions are correct. Recall=8/(8+2)=80% asks how many actual urgent tickets were recovered.
F1=2×8/(2×8+12+2)=16/30≈53.3%. It is not the arithmetic mean (40%+80%)/2=60%. F1 excludes TN and does not replace a full account of operational losses or system experience.
| Metric | Calculation | Question |
|---|---|---|
| Accuracy | 86/100=86% | Which proportion of all predictions is correct? |
| Precision | 8/20=40% | How trustworthy are urgent predictions? |
| Recall | 8/10=80% | How many actual urgent tickets are found? |
| F1 | 16/30≈53.3% | What is the harmonic mean of precision and recall? |
Ask AI to rebuild the counts before writing the report
Use an AI system you are permitted to use. In KitNelo’s AI productivity studio, choose Data insights & hypotheses and enter only the exercise data. Active generation sends the input to the server; availability and usage limits follow the page status. No model was called for this article.
This is original exercise data, not customer data. Positive class = urgent tickets. TP=8, FP=12, FN=2, TN=78.
Verify the total, actual positive count and predicted positive count, then calculate accuracy, precision, recall and F1 with every denominator.
Compare a baseline that predicts routine for every ticket. Explain why accuracy alone cannot select a workflow.
Do not infer error costs, deployment suitability or product rankings. List missing decision evidence.
Retain exact fractions and round only at the end.Explain the 90% baseline correctly
Predicting routine for all one hundred tickets gets ninety correct, producing 90% accuracy—higher than 86%. Yet urgent recall is 0/10=0%. When the task is to recover urgent tickets, this baseline exposes how accuracy can conceal task failure.
With no urgent predictions, precision has a zero denominator and is undefined under the raw fraction definition. Do not silently replace it with 100% or 0%. State any reporting convention and preserve the fact that no urgent predictions were made.
The other workflow finds eight urgent tickets while sending twelve routine ones into the urgent queue. Its value depends on capacity, missed-ticket consequences and reference-label quality. We have assigned none of those costs, so this exercise selects no deployment winner.
Repair an authoritative-sounding mistake
Accuracy is 86%; averaging precision and recall gives 60%, so urgent handling is reliable.
Accuracy is 86%, precision 40%, recall 80% and F1 about 53.3%. Twelve false alarms and two misses still require operational acceptance.
Accept the underlying evidence as well as the prose
Check that the hundred records are complete and unduplicated, labeling rules are consistent, and predictions align with reference labels. A join that repeats records would no longer count one hundred independent tickets. For aggregation risks, see the channel-versus-pooled reversal exercise.
Validate with an independent sample that did not determine the threshold, within the actual process. Record threshold, sample date and label rules. Do not invent uncertainty intervals, savings or rankings from these practice counts. The percent versus percentage-point denominator checks address another common reporting error.
A report another person can reproduce
- Positive-class definition, threshold, four counts and sample provenance remain available.
- The total 100, actual urgent 10 and predicted urgent 20 can be reconstructed.
- F1 uses a harmonic mean, and zero denominators have an explicit convention.
- Operational costs, independent evidence and label uncertainty remain visible.
Common questions
Is lower accuracy always worse?
No. The 90% all-routine baseline misses every urgent ticket. Compare metrics in relation to the task and error costs.
Can precision and recall be swapped?
No. Precision uses predicted positives as its denominator; recall uses actual positives. Swapping names answers a different question.
Does 86% establish an AI product’s performance?
No. It comes from original counts without running a classification product. It establishes no vendor capability, customer outcome or deployment guarantee.
Further reading & sources
Put the guide into practice
Open the relevant tool, preview the result and check your exported file.
