Your AI Needs a Test Suite More Than It Needs a Better Model

When an AI feature underperforms, every team's first instinct is a better model. Nine times out of ten what is actually missing is a test suite. Software engineering learned this decades ago; AI teams are relearning it with production incidents.

Why AI regressions are silent

Ordinary code fails loudly — exceptions, 500s, red dashboards. A prompt edit that degrades refund-policy answers fails silently: the system still returns fluent, confident text. Wrong answers do not throw. Without evaluation, your detection mechanism is customer complaints, which is to say your brand is the canary.

The triggers are constant, too. Providers update models (sometimes behind the same version name). Someone "improves" a prompt on a Friday. Your catalogue changes shape. Retrieval indexes drift. Each is a deployment to your AI's behaviour, and most teams ship them untested.

What a real harness looks like

• A labelled set of 150–500 real cases — pulled from production traffic, not invented. Include the ugly ones: ambiguous requests, angry phrasing, questions touching policy edges. Twenty synthetic happy-path examples is a demo, not a test.
• Graders appropriate to the task. Exact-match for extraction. Rubric-scored LLM-as-judge for prose, with the judge itself spot-audited by humans monthly — judges drift as much as workers.
• Thresholds wired to CI. Below the line, the prompt change does not merge. Same as unit tests, same muscle memory, same argument-ending authority in code review.
• Segment reporting. Overall accuracy hides the segments that hurt: fine on English electronics queries, bad on French returns. Slice by language, category and intent, because your customers arrive pre-sliced.
• A drift sample from production. Grade a daily sample of live outputs against the rubric. This is how you learn a provider changed something before their changelog says so — and how you know your accuracy claim is still true this quarter.

The economics

A harness for a support agent costs one to two engineer-weeks including labelling. A single bad-model-week in season costs more — in refunds mis-issued, in escalations, in the meeting where someone proposes turning the whole thing off. And the harness compounds: it is what lets you adopt cheaper models confidently, trim prompts aggressively and expand agent authority with evidence. Every optimisation in our cost-cutting playbook leans on it.

We do not ship AI systems without one — the eval suite is deliverable one, before the feature, and it transfers to you with the code. If a partner proposes an AI build with no line item for evaluation, that line item exists; it is just been moved into your incident budget.