AI evals
AI evals (evaluations) are repeatable tests for an AI system. You collect examples of the task, decide what a good result looks like, run the model, prompt or agent against them, and score the output. The point is to replace “it seemed better when I tried it” with a number you can compare from one version to the next.
Why it matters
AI output varies from run to run, and a change that fixes one case often breaks another. Without evals, every prompt tweak or model upgrade is a guess, and problems show up in front of customers. With them, you can switch to a cheaper model, rewrite a system prompt or add a tool and know within minutes whether quality held.
Public benchmarks don’t answer this for you. A model that tops a leaderboard can still be wrong for your contracts, your codebase or your support tickets. The evals that matter are built from your own work.
Building useful evals
- Start small. Twenty to fifty real examples, including the awkward ones that went wrong before, beat a thousand synthetic ones.
- Check what you can check. Did the code pass its tests, did the answer cite the right document, did the agent call the right tool? Exact checks are cheap and reliable.
- Use a model as judge carefully. For writing quality or tone, a second model can grade against a rubric, but spot-check its grades against a person’s.
- Test the whole agent. For agents, score the outcome and the path: did it finish, how many steps it took, whether it did anything it shouldn’t have.
Where it goes wrong
Teams build evals once and stop updating them, so the tests drift away from what users actually ask. Add every real failure to the set. Teams also over-fit: prompts get tuned until the eval passes while real quality stays flat. Keep some examples back that you never tune against.