08 / Trust & Governance
Evals
Structured tests that measure whether an AI system does its job well, run on a fixed set of real examples every time something changes.
How you’ll hear it
“Is the new version better?”
“The evals say better on refunds, worse on complaints.”
Avoid
Judging an AI system from a handful of impressive examples.
The longer version
Evals are to AI products what tests are to software: a repeatable set of examples and a way of scoring the output. Because AI behaviour varies, evals are how a team knows whether a change helped, and how an organisation shows quality to leadership and regulators.
In depth
Evals are structured tests for AI systems: a fixed set of realistic examples, a way of judging each answer, and a score you can compare over time. They are the AI equivalent of a test suite, and the main way to know whether a change made things better or worse.
They matter because AI output varies. A system that impressed everyone in a demo can fail on the cases that matter most, and a new model version can improve one task while quietly degrading another. Without evals, those shifts are found by customers.
Good evals are built from your own work: real questions, real documents, the awkward edge cases your team knows about. Some can be scored automatically; others need a person, or another model checked against human judgement. Starting with fifty well-chosen examples beats waiting for five thousand.
Evals also turn quality into something you can report. Leaders and regulators rarely want to read transcripts, but they can follow a score that moves.
What to do
- Ask for the eval results before approving any AI feature, not a demo.
- Build the first set from real cases, including the ones that went wrong.
- Re-run the evals every time the model, the prompt or the data changes.