Evaluation Tooling Has Become a Category Companies Buy Rather Than Build

A recognisable pattern in enterprise software is that capabilities begin as something every team builds badly for itself, and become a purchased category once enough teams have built enough versions to agree on what the thing should do. Continuous integration went through it, error monitoring went through it, feature flagging went through it. AI evaluation and observability tooling is going through it now, and the shift from homegrown harness to bought product is the clearest available evidence that shipping AI features has become routine engineering rather than experimentation. The homegrown harnesses had a common shape and common weaknesses. A folder of test cases, a script that ran them against the current prompt and model, and a human reading the output to decide whether it got better. That works at small scale and fails in specific ways as usage grows: the test set stops representing real traffic, judgement of quality drifts between reviewers and over time, and nobody can say whether last Tuesday's prompt change helped, because the comparison was never systematic enough to answer. What the purchased tooling adds is mostly discipline rather than intelligence. Production traces captured and sampled into evaluation sets automatically, so the tests track what users are actually doing. Regression comparison across model and prompt versions, with the same cases scored the same way each time. Structured human review workflows for the cases automated scoring cannot settle. Online monitoring that catches quality degradation in production rather than waiting for the next offline evaluation run — which matters because the underlying model can change beneath a product that changed nothing. The build-versus-buy calculation still favours building in one situation worth naming: when the quality criterion is genuinely domain-specific and requires expertise the tooling vendor cannot encode — clinical accuracy, legal correctness under a particular jurisdiction, safety in a regulated industrial process. In those cases the scoring logic is the hard part and general tooling only supplies the plumbing. For the far larger set of products where "is this response accurate, appropriately formatted, and on-tone" is the actual question, writing that infrastructure in-house has stopped being a defensible use of engineering time.