How we make AI reliable enough for legal work

Every AI product goes through the same moment: someone tweaks a prompt, runs three queries, and announces that the system "seems better." Sometimes it is. Sometimes the tweak quietly broke a case nobody tried that day, and you find out from a customer. The difference between teams that compound quality and teams that oscillate is not model choice. It is whether "better" is a measurement or a mood.

What a working harness looks like

Running CaseScout, where a wrong answer is not an inconvenience but a professional risk for the lawyer reading it, forced us to treat evaluation as core infrastructure. The shape that emerged has four parts, and none of them are exotic.

• •Versioned prompts. Every prompt in the pipeline is a versioned artifact, dozens of stages, each with history. A prompt change is a diff you can review and roll back, not an edit in someone's notebook.
• •A golden set. Real questions with known-good answers, curated by domain experts, covering the ugly cases: ambiguous questions, thin corpus areas, questions the system should refuse.
• •LLM-as-judge, spot-audited by humans. Judge models score outputs for groundedness, completeness, and citation accuracy on every run. Humans audit the judges on a sample, because an unexamined judge drifts just like an unexamined model.
• •A daily quality report. Scores, regressions, and cost per query broken down by pipeline stage, delivered automatically, so quality is something the team reads about every morning rather than debates in retrospectives.

Evals change what you're allowed to do

The payoff is not a dashboard; it is velocity with a safety net. Want to swap the embedding model? Run the harness, compare retrieval hit rates head to head. New model release from a lab? An hour later you know whether it helps your workload. Not the benchmark, your workload. We run A/B comparisons on retrieval configurations the way web teams A/B test landing pages, because regressions get caught by the harness instead of by customers.

Scale is a multiplier. Multiply a measured system and you get growth. Multiply an unmeasured one and you get incidents.

If your firm is evaluating an AI tool, ask the vendor one question: how do you measure answer quality, and can we see the report? If the answer is a shrug, keep shopping. If you want to see what a real answer looks like, ask us for ours.