What is Model evaluation?
A model evaluation consists of three parts: a set of real test cases, defined criteria for a good output, and a measurement that can be repeated. The difference from the widespread "we tried it and it looked fine" lies in repeatability: the same cases, the same criteria, a documented result - which makes two prompt versions, two models or two points in time comparable. The test set is built from real transactions, not invented examples, because nobody invents the edge cases systems actually fail on.
What gets measured is what matters for the deployment: factual correctness, the rate of `hallucinations`, adherence to format and tone, and for agents the choice of the right tools. Methods range from human spot checks through automated assertions to using a second model as judge - which has blind spots of its own and needs checking itself. For the boundary: an A/B test measures the effect on users in live operation, model evaluation measures the quality of the output before and during it - one does not replace the other.
Evaluation shows its real value after launch. Models get swapped, prompts get adjusted, knowledge sources grow, and every one of those changes can cost quality that used to be there. Whoever runs the test set on every change trades known quality only deliberately for new - whoever does not, discovers regressions through complaints. That is why the test set belongs to the product: it grows with every real failure case and has an owner who maintains it.
Why does Model evaluation matter?
McKinsey has measured how rare this discipline is: fewer than one in five companies track well-defined KPIs for their GenAI solutions - and that same KPI tracking is among the strongest predictors of measurable bottom-line impact. Evaluation is not bureaucracy; it is the difference between a tool and a bet.
Model evaluation in practice
- 01Quoting assistant: 50 real enquiries from the last quarter form the test set - every prompt change runs against it first, then into production.
- 02Support chat: every week one person rates 20 random conversations on three criteria - correct, complete, right tone - and the curve of those scores sits next to the usage numbers in the dashboard.
- 03Model swap: before the cheaper model takes over, both run on the same test set - the decision is made on the result, not on the price list.
