AI and professional secrecy: what a law firm can do, and what it cannot

The Barreau du Québec has published a practice guide on the responsible use of generative AI, updated since its first release. On professional secrecy, it is unusually clear. The guide is in French, so the quotes below are our translation.

Everything turns on that distinction. The guide defines an open system as one that is publicly accessible and does not store data internally. A consumer chatbot used with a personal account is the obvious example.

The guide goes straight on: “By contrast, closed systems, developed and deployed within an organization and keeping data internally, reduce these risks.” It does not say they remove the risk. It says they reduce it, which is the honest way to put it.

Between those two poles sit vendor-hosted enterprise offerings with contractual commitments. The guide does not classify them by name. It asks firms to choose solutions with strong protection for data and professional secrecy, and to make sure any third-party collaboration meets the highest confidentiality standards. So each tool is the firm’s call, contract in hand.

The guide does not ban AI. It sets the terms for using it. Some uses never touch professional secrecy, and others open up once the tool runs as a closed system.

Pasting a client’s contract into a consumer assistant to get a summary is exactly the act the guide calls a breach. The information does not have to be reproduced or disclosed later. Entering it is enough.

The guide also reminds lawyers that judgment cannot be delegated. A reference produced by AI must be checked against recognized, authoritative sources before it is used, and generated content does not go to the client or the court without validation.

The guide is blunt: the legal professional “remains fully responsible for the mandate.” The lawyer must ensure the quality and compliance of the work, whether people or technology tools did it, “by exercising close supervision.” The duty to supervise expressly extends to information technology, generative AI included.

In practice, that means a written use policy, clear rules on what may go into which tool and verification procedures. Our AI acceptable use policy template is a reasonable starting point.

The guide recommends documenting the instructions given to the AI and the steps taken to validate the results. It says this record must be kept for at least 7 years after the file is closed.

A well-built tool makes this easier: every query, every document consulted and every sign-off leaves a time-stamped trace tied to the file. When the tool is hosted at the firm, that log stays under the firm’s control too.

The guide asks lawyers to inform clients and obtain free and informed consent, explaining the risks of generative AI and the mitigation measures planned. It suggests putting it in writing, for instance through a clause in the retainer agreement.

That obligation belongs to the lawyer. No technology vendor, Cogio included, can discharge it on the lawyer’s behalf. What a vendor can do is describe exactly where data goes and who can reach it, so the lawyer can explain the situation to the client with full knowledge.

Before connecting a tool to the firm’s files, a few questions tell you which category it falls into. The answers should be in writing, and specific enough for the lawyer to explain them to a client.

Start by listing the AI tools already in use at the firm, including the ones staff picked up on their own. For each, note what information goes into it. That is often where the first risk turns up, well before any technology project.

Next, pick one low-risk, high-volume workflow, such as sorting incoming email or preparing file openings. Measure the time before and after, keep the lawyer’s sign-off at every step and document it all as the guide asks. The rest comes more easily once that first workflow runs smoothly.

At Cogio, we install AI agents on a server at the firm or with a hosting provider in its region, under the firm’s own contract. The model runs on that server, without going through an outside AI vendor. The agents read email, documents and the firm’s practice management software, prepare the work, and the lawyer signs off. That is what we call sovereign AI.

Our page for law firms describes the possible workflows, each with its own lawyer checkpoint. If the firm runs Microsoft 365, our article on Copilot and Law 25 explains where data goes in that case. Both approaches can coexist, as long as the firm decides which information goes where.