Write the evaluation before the prompt
Most AI projects start by tuning prompts against a few examples. Ours start by agreeing what "good" means on a hundred real ones.
A demonstration that works on five documents tells you nothing about the five-thousandth. The only reliable way to know whether an AI system is good enough is to define "good enough" and measure it — before the first prompt is written.
Build the set from real work
Collect representative cases from the actual workflow, including the ugly ones, and have the people who do the job today label the correct outcome. Disagreements between labellers are not noise; they are the specification you did not have.
Agree the threshold with the business
What accuracy makes this useful? What error is unacceptable? These are business decisions, and they belong to whoever owns the process — not the engineer.
If you cannot say what would make you switch it off, you are not ready to switch it on.
Keep measuring in production
Corrections made by reviewers go back into the set. Model updates, prompt changes and vendor migrations are then a measurement, not a leap of faith.