What Are Evals (in AI)?

Evals (short for evaluations) are the tests that measure how well an AI system does its job. Much as traditional software has unit tests, an AI application has evals: a set of example inputs paired with expected or graded outputs, run automatically so you can tell whether a model, prompt, or agent is working, and whether a change made it better or worse.

Evals matter because AI does not behave like ordinary software. The same prompt can give different answers, quality is a matter of degree rather than pass or fail, and a small change to a prompt or model can quietly break things that used to work. You cannot eyeball that at scale. Evals turn “this feels about right” into a number you can track, so teams can improve an AI system on purpose instead of by guesswork.

What evals measure

It depends on the job, but evals usually check one or more of these:

• Accuracy: did the AI get the right answer?
• Format: did it return the structure you asked for, such as valid JSON or a specific layout?
• Task completion: for an agent, did it actually finish the job, not just talk about it?
• Safety: did it avoid harmful, biased, or off-limits responses?
• Faithfulness: did it stick to the source material instead of making things up?

Why evals matter for AI products

Three reasons. First, AI is non-deterministic, so you need many examples to judge real quality, not one lucky test. Second, evals catch regressions: when you swap a model or tweak a prompt, a good eval suite tells you immediately if something got worse. Third, they build trust. A team that can show its AI passes a clear set of evals can ship changes with confidence instead of crossing its fingers.

Evals versus traditional software tests

A normal software test is usually pass or fail: the function returns 4 or it does not. Evals are often graded, because language has shades of better and worse. An eval might score an answer for helpfulness on a scale, or use another AI model as a judge to rate the output. So evals borrow the idea of automated testing but adapt it to the fuzzy, probabilistic nature of AI.

Who works on evals

Writing good evals is a core skill for AI Engineers, and it is becoming specialized enough that some people expect a dedicated “Evals Engineer” role to emerge, the way software engineering split into many specialties. If that interests you, our explainer on the AI Forward Deployed Engineer and AI Engineer roles covers where evals fit into the wider set of new AI jobs.

Common misconceptions

• “One good test is enough.” Because AI varies, you need many examples to trust a result.
• “Evals are only for big labs.” Even a small AI feature benefits from a handful of example inputs you re-run after every change.
• “A high score means it is perfect.” Evals measure what you chose to test. They miss anything you did not think to include.

You may also like

• AI Forward Deployed Engineer (and AI Engineer jobs)
• What is an AI agent?
• How to use Claude AI
• The AI Glossary
• The daily Beginners in AI newsletter