Classifying Thousands of Records Cheaply and Accurately

Bulk classification is the least glamorous AI task and often the most valuable. It is also where naive implementations burn the most budget.

1. Define categories with boundaries

Ambiguous labels produce ambiguous results. For each category, write a rule and one borderline example. If two reviewers disagree on a sample, the definition is not finished.

2. Label a gold set by hand

Two hundred examples is enough to start. Hand-label them carefully — this set is your only measure of whether anything works.

3. Start with a small model

Classification rarely needs frontier reasoning. Test a small, cheap model first. If it reaches acceptable accuracy on the gold set, stop there.

4. Add confidence routing

Ask the model to return both a label and a confidence value. Auto-accept high confidence, send the middle band to a larger model, and route the lowest band to human review. This usually captures most of the savings without losing accuracy.

5. Use batch processing

For non-interactive workloads, asynchronous batch endpoints cost substantially less. A nightly job is indistinguishable from a real-time one from the user's perspective.

6. Monitor drift

Re-run the gold set monthly. Input distributions change — new products, new phrasing, seasonal patterns — and accuracy degrades silently when they do.