An AI that says it is 90% sure is reporting how it feels, not how often it is right.
Auto-classification is now good enough to use at scale. Some practitioners report it can match or beat manual review when the categories are well defined and the content is clean. But using it responsibly means measuring it. Here is a practical way to do that.
Step 1: Let the AI report its confidence
When the AI assigns a category, such as a record type or sensitivity level, ask it for a confidence score with each decision. For example, “I think this is a lease amendment, and I am 80% sure.”
Step 2: Choose a threshold
- Below the threshold: a person reviews it.
- Above the threshold: accept it, but sample a portion to check.
Pick the threshold by risk. Sensitive records need a higher bar than low-risk operational ones. One speaker cautioned against automation that is right only about 72% of the time for sensitive records.
Step 3: Sample the accepted decisions
A confidence score can be wrong, so check the high-confidence results too. The number you need is smaller than most people expect.
For a large collection, reviewing about 385 randomly chosen documents lets you estimate the error rate to within plus or minus 5 percentage points, with 95% confidence. This holds whether you have ten thousand documents or ten million.
The math is the standard sample-size formula: 1.96 squared, times 0.25, divided by 0.05 squared, which is about 385. The 0.25 assumes the most uncertain case, so the result is conservative.
- To narrow the margin to plus or minus 3 points, review about 1,068 documents.
- To narrow it to plus or minus 2 points, review about 2,401.
- For a small collection, you need fewer. With 2,000 documents, about 323 is enough for the plus or minus 5 point margin.
Make the sample honest
- Pick at random. Do not review the first documents or the easy ones.
- Sample each category separately if you need to know accuracy for each. One overall sample of 385 does not tell you how well the rare categories perform.
- Use qualified reviewers, and define in advance what counts as correct.
Use two reviewers for high-stakes decisions
For decisions like disposing of records, have the AI and a person each make the call independently, then compare notes on the disagreements. This borrows the dual-review technique used with human reviewers, and it shows you where the AI and your people differ.
Keep testing
- Re-sample on a schedule, not once.
- Re-test after any change to the model, the categories, or the kind of documents arriving.
- Watch for drift. A document's sensitivity can change over its life, and a threshold that worked last year may not work now.
- Keep a small set of test questions, called evals, that you run after each update.
How we apply this
Our document pipelines extract and classify, then send low-confidence fields to a human reviewer so decisions that matter are not left to a guess. See intelligent document processing and the FAQ. For the data side, start with Is your data ready for AI?, and see why retention schedules decide AI accuracy.
Illustrations in this guide are drawn from talks at an industry information-governance conference and from general practice. They are anonymized and are not Horus client results.
Part of our guide: Information governance for AI