What are AI evals?

AI evals, short for evaluations, are tests that check how well an AI system does a specific job. The system is run on a set of cases where the correct answer is already known, and each result is scored. A person who knows the work can do the scoring, or software, or another AI model following written criteria. The same tests are run again whenever the AI, its instructions or its data change.

Why it matters for a business

A demo shows AI working on inputs somebody else chose. Evals show what it does on the business’s own inputs, including the awkward ones. 20 to 30 real past cases, read carefully, answer most of what a team needs to know before relying on a tool, and they show where the wrong answers cluster, which matters more than how many there are. After launch, the same set is what catches a regression when a provider ships a new model version.

Example

Let’s say a finance team runs 25 past invoices, already coded and approved, through an AI step that assigns each one to a cost centre. 20 match. 3 of the 5 misses are invoices billed through a reseller, the same mistake each time, and one added line in the prompt clears all 3.

Go deeper

Testing AI before trusting it: how to run an eval on your own work

On Atlas: LLM observability and evaluation tools, compared

Related terms

• Golden dataset
• AI benchmarks
• LLMOps
• Model drift

All terms, A to Z