5 min read
Your AI feature needs an evaluation set, not a demo
A convincing demo proves that something can work once. An evaluation set tells you how often it works.
Language models are very good at producing convincing demos. Feed them a well-chosen invoice and they extract every field correctly. That proves the approach can work. It says nothing about the hundred invoices with handwritten notes, rotated scans and unusual layouts that arrive next month.
Start with fifty real cases
Before writing a prompt, collect real examples from the workflow — anonymised where necessary — and write down the correct answer for each. Fifty cases are enough to start; include the awkward ones on purpose. This set becomes the yardstick for every decision that follows.
Measure, then decide
With an evaluation set, questions like “which model?” or “is this prompt better?” stop being matters of taste. You run the set, compare the results, and look at the failures. Often the most valuable outcome is learning which kinds of input the system should not handle automatically at all.
Design for the uncertain cases
No extraction or classification will be perfect. A production system needs a defined path for uncertain results: a review queue, a confidence threshold, a person who sees the original document next to the extracted data. Designing that path is engineering work, and it is what separates a pilot from a system.
The demo is still useful. It just should not be the last test.