Most firms ask the wrong question about AI. They ask whether they are allowed to use it. They are. The useful question is shorter and less comfortable: when text goes into the tool, where does it go?

What the ABA opinion asks of a law firm

The ABA generative AI ethics opinion, Formal Opinion 512, tells lawyers that their existing duties follow them into these tools: competence, confidentiality, communication, supervision, candor and reasonable fees. The practical test comes first: know where client information goes when it is typed into a tool, and get informed consent when the tool can reuse it.

The sentence that changes daily practice

The opinion was issued in July 2024 by the ABA’s Standing Committee on Ethics and Professional Responsibility. Most of it restates duties lawyers already know. One passage goes further, and it is the one a managing partner should read twice:

“many of today’s self-learning GAI tools are designed so that their output could lead directly or indirectly to the disclosure of information relating to the representation of a client, a client’s informed consent is required prior to inputting information relating to the representation into such a GAI tool.” — ABA Formal Opinion 512, as quoted in The Bar Examiner, National Conference of Bar Examiners

In plain words: if the tool learns from what it is given, the client must agree before anything about their matter goes in. And the same commentary is clear that a boilerplate waiver does not count as informed consent. A line buried in the engagement letter will not do the job.

The opinion is advisory. It interprets the ABA Model Rules, and each state bar applies its own rules and its own guidance. But it is the reference point most of them started from, and the reasoning is hard to argue with.

Where the profession stands

Adoption is real, and still a minority. According to the ABA’s 2024 Legal Technology Survey Report, as reported by the ABA Journal, 30% of responding lawyers are using AI. Three-quarters of those surveyed named AI-generated hallucinations as a reason to hesitate.

Accuracy gets the headlines. Confidentiality gets less attention because nothing visibly breaks. A wrong citation is caught by a judge. A paragraph of client information sent to the wrong place is caught by no one.

What to check before a tool sees a file

The opinion asks lawyers to understand a tool before relying on it. That comes down to a handful of questions. None is technical, and a firm is entitled to a written answer to each.

What leaves the firm. The whole file, or only the passage the task needs?
Where it goes. Which vendor, in which country, under which contract?
Who can read it. Do people at the vendor have access to what is sent?
What is kept. For how long, and for what purpose?
What is reused. Does the text train a system that other customers will use?

One more question is for the firm itself: afterward, can it say what was sent, by whom and for which matter? Without that record, it cannot answer a client, and it cannot answer the bar.

The open browser tab

The most common risk is not an IT project. It is a browser tab. An associate in a hurry opens a consumer AI assistant, pastes in a complaint to get a summary, and closes it. Nobody decided it and nobody knew. The terms of that assistant can change without the firm being told.

The firm’s practice management software is not to blame. It holds the matter, and holds it well. But it does not summarize and it does not sort, and that gap is exactly what the tab fills.

A ban alone does not work. A firm that forbids AI and offers nothing in its place gets hidden use, which is the worst kind for confidentiality.

What an installed agent changes

An agent built to measure and installed at the firm answers the same questions differently. Files stay on the firm’s own machines. For each task, only the text it needs is sent to the AI model.

That is not nothing, and we do not claim that nothing leaves. It is a known amount, limited to the task, that a lawyer can describe to a client when asking for consent. The difference with the browser tab is not one of degree. In one case the firm knows. In the other it does not.

On one engagement at a Paris law firm, over 28 days, the agents handled 144 qualified inquiries and carried out 8,090 actions. Each of those actions read or wrote something. At that volume, confidentiality cannot rest on one person checking every move. It has to be settled beforehand, in how the tool is built. We have not delivered an engagement to a US law firm; what we describe here is what the opinion requires and what we have seen of the work.

What the agent never does

It gives no legal advice. It takes no position on a matter, does not speak for the lawyer and never poses as one. Anything that goes out in the firm’s name has been approved first.

It does not decide what is confidential either. That judgment belongs to the lawyer, along with the responsibility that comes with it. The agent works inside a perimeter the firm has set, and it stops at the edge.

Read more: what an agent takes over at a law firm and what happens to the firm’s data. Pricing starts from $1,700 per agent, then $170 per month: see what a custom agent costs.

What an AI tool sees of the client file