"We use AI" has become one of the least informative sentences in legal services. It can mean a rebuilt process with controls around it — or a chat window someone opens between two memos. As a buyer you cannot audit the technology, and you do not need to. Five questions, and the shape of good and bad answers to each, will sort serious practice from marketing in a single meeting.
Who supervises the work?
Good answer: a specific, named lawyer reviews every piece of work before it reaches you; that lawyer's name is on the deliverable, and their workload makes real review plausible. When you ask who will supervise your matter, you should get a name, not an org chart.
Red flags: "quality is assured by the system", supervision described as spot-checking a sample of the output, or the admission — sometimes a cheerful one — that the volume is too high for anyone to read everything the client receives.
The follow-up is just as telling: ask how much time the supervising lawyer typically spends on a matter like yours. The answer does not need to be large — it needs to be consistent with genuine review.
Is every finding traceable to its source?
Good answer: yes — and here is a demonstration. Every statement in the deliverable links to the exact clause or document it comes from; coverage is reported, including what could not be read; quoted text is verbatim. This is a question best asked with a screen in front of you.
Red flags: accuracy statistics offered in place of traceability. "The model is highly reliable" answers a different question — it tells you the findings are usually right, and at the same time that you have no way of knowing which ones.
Who answers for the result?
Good answer: we do — with the same professional responsibility as for any other mandate. The engagement letter should read no differently because AI was involved; the firm's liability, professional secrecy and duty of care do not change with its tooling.
Red flags: AI-specific disclaimers in the engagement terms, language that shifts responsibility toward a software vendor, or any suggestion that AI-assisted deliverables come with a different standard of care. If the firm will not stand behind the output, you are the one standing behind it. This question deserves to be settled in writing, not in conversation — read the engagement terms with it in mind before you sign.
How is confidential client data handled?
Good answer: precise and unhesitating. Where the data is processed, under which contracts, whether it is used to train models (it should not be), how professional secrecy is preserved, and what happens to the data when the matter ends. A Swiss firm should answer in terms of its own obligations, not its supplier's brochure.
Red flags: "our provider is fully compliant" without being able to say with what; not knowing in which country processing takes place; or treating the question as unusual. Firms that have done the work find this the easiest question of the five.
What happens when the system is uncertain?
Good answer: uncertainty is surfaced, not smoothed over. Ambiguous clauses, unreadable documents and conflicting sources are flagged in the deliverable and escalated to the responsible lawyer — and the firm can show you what such a flag looks like in practice.
Red flag: "it is very rarely wrong." Every system of this kind is sometimes uncertain; a firm that has never seen its system express uncertainty has a system that hides it — which is the most dangerous configuration on this list.
None of these five questions is about technology. They are about supervision, evidence, accountability, confidentiality and honesty — the questions you would ask about any legal team, sharpened for a new way of working. That is the point: AI does not change what you are entitled to expect from a law firm; it changes how visibly a firm meets those expectations, or fails to. Firms doing this work seriously enjoy the questions, because the answers are their differentiator. We are one of them — and if you want to put the five questions to us, we would welcome it.