Why 'AI for Legal' Means Nothing — And What to Ask Instead

Every legal technology vendor claims to use AI. This is true in the same way that every restaurant claims to use fresh ingredients — technically defensible, operationally meaningless, and designed to avoid a more uncomfortable conversation.
The uncomfortable conversation is: what does your AI actually do, and what happens when it gets something wrong?
Here are the five questions that cut through the noise.
01 — What data was the model trained on? A model trained on publicly available legal documents performs differently from one trained on negotiated commercial agreements. A vendor who cannot answer this question clearly is telling you something important.
02 — How does the system handle uncertainty? Legal AI systems make mistakes. The relevant question is whether the system knows when it is uncertain. A well-designed system flags clauses it is less confident about differently from ones it has reviewed thousands of times. A system that presents all outputs with equal confidence is designed to look impressive in a demo.
03 — What is the human review process? AI-assisted review is not AI-replaced review. The AI surfaces, flags, and prioritizes. The attorney decides. Any vendor suggesting their system eliminates attorney review is creating liability for the firms that believe them.
04 — How is client data handled? Two components: does the vendor retain client documents after processing, and does the vendor use client documents to train their model? Both require a clear answer before any client matter touches the system.
05 — What does failure look like? This separates vendors who understand legal work from vendors who understand software. A vendor who has thought seriously about this will tell you exactly what their system is most likely to miss and under what conditions. A vendor who hasn't will tell you their system is highly accurate. Accuracy is not a safeguard. It is a statistic.
The useful questions are never about the technology. They are about the vendor's relationship with uncertainty, with failure, and with the specific demands of legal practice. The vendors who understand that will still be the right partners when the technology improves.