Legal AI in Practice

Compare the order inside a column. Never compare the numbers between columns. Each survey wrote its own task list, asked its own population and counted in its own way. Censis allowed one answer per respondent and counted every avvocato, so its column tops out at 19.9%. Thomson Reuters let respondents pick as many as applied and counted only people whose organisation already uses the tools, so its column tops out at 77%. Those two numbers are not two measurements of the same thing.

What survives all of that is the ranking. Finding the law, summarising and reviewing documents sit at the top of every column, drafting follows, and the bottom row is the one nobody asks about: three of the four questionnaires carry no strategy or outcome prediction item at allThomson Reuters · Generative AI in Professional Services, Apr 2025 (opens in a new tab)and the one that does put it at 1.2%Censis and Cassa Forense · Rapporto sull'Avvocatura 2025, table 35, Apr 2025 (opens in a new tab). A sponsor writing a task list is already telling you what the profession treats as a plausible job for a model, which is why the blank cells are worth as much here as the dark ones.

Size and difficulty are not the axis. Summarising a 300 page bundle is a big job and every sentence in the summary can be traced back to a page. Deciding an indemnity cap takes a minute and no document in the file tells you afterwards whether the answer was right. Verifiability is the axis.

Trust falls down the rungs in that order wherever it is measured. A 2025 survey of 4,457 French avocatsCNB and Viavoice · L'IA et la profession d'avocat, p. 29, 2025 (opens in a new tab)put trust at roughly 66% for finding the most relevant source, 37% for answering a legal question, 32% for writing submissions, about 25 to 27% for building a strategy and 9% for deciding a case. Frequent users trusted it more at every rung, and the order never changed.

The surveys are consistent with a verifiability filter. They do not prove one. Professional rules, personal liability, plain habit, and the fact that clients pay for a lawyer's judgment rather than for text would each produce the same ranking on their own.

Keeping what you tell it confidential is on the chart but is not a task. It was in the same question, so leaving it out would be choosing the rows that suit the argument. It asks whether the model can be trusted to respect confidentiality, which is a property of the tool rather than a job you hand it, and it is marked rather than removed for that reason.

The survey was run by Viavoice for the Conseil national des barreaux among 4,457 avocats between 17 April and 27 May 2025CNB and Viavoice · L'IA et la profession d'avocat, volet avocats, page 29, 2025 (opens in a new tab). The same deck reports that 92% of avocats using a general purpose model in their practice say they check its answers, rising to 93% among lead users, which is the behaviour you would expect from people who trust it most where checking is cheapest.

The honest limit. These surveys are consistent with a verifiability filter. They do not prove one. Professional rules, personal liability, plain habit, and the fact that clients pay for a lawyer's judgment rather than for text would each produce this same ranking on their own, and nothing in a trust question can separate them. What the data shows is the ordering and its stability. The explanation is an argument laid on top of it.