Lawyers have spent the last few years worrying about how lawyers use AI. We wrote policies, negotiated enterprise agreements with no-training, zero data retention, and confidentiality protections, trained our people, and read nervously about every new story about a brief filled with cases that do not exist. That work was necessary and it is not finished, but while we were busy putting our own house in order, our clients moved in with a consumer AI account and a lot of confidence.

Clients are now using consumer AI tools in ways that create real problems for them and for the law firms that represent them. These problems have been simmering quietly for a while. They are about to come to a head, and in-house counsel, clients, and law firms would all be better served by talking about them now rather than after a privilege fight or a malpractice claim, no matter how far-fetched either may be.

The first issue is the simplest. A client takes confidential information, whether its own trade secrets, customer data, or personal information about employees, and pastes it into a consumer AI tool. Many consumer tools reserve the right to retain prompts and outputs, use them to train their models, and share them with third parties. Even where a tool does not train on the data, the client has handed confidential information to an unrelated third party that has made no promise to keep it confidential. That confidentiality gap can undercut trade secret protection, violate the client’s own contracts and privacy obligations, and create an argument that the information is no longer confidential at all. The client did not think of it as a disclosure. The privacy policy it clicked through without reading certainly did consider it a “fair game” disclosure.

The second issue is where lawyers should start paying very close attention. Clients are taking our advice, our emails, our draft pleadings and our work product and feeding them into consumer AI tools to summarize them, ask follow-up questions, or get a “second opinion.” The impulse to use AI is understandable. No client wants to pay for a call to ask what the third paragraph of a memo means. But privilege depends on confidentiality, and sharing privileged communications with a third party is the classic way to waive it.

The concerns of clients using consumer AI is no longer hypothetical. A federal court in the Southern District of New York answered the question earlier this year: documents a defendant generated with a consumer AI chatbot were not privileged and were not protected as work product. The court reasoned that the chatbot is not a lawyer, that the tool’s privacy policy disclosed that inputs and outputs could be used for training and shared with third parties (destroying any reasonable expectation of confidentiality), and that the materials were not prepared at the direction of counsel. The court also made clear that sending the AI-generated material to counsel afterward did not make it privileged. Another federal court reached a more forgiving result on work product for a pro se litigant, reasoning that waiver requires disclosure to an adversary or in a way likely to reach one. So, while courts are not yet in lockstep, nobody wants to be the client whose privilege depends on which judge drew the case.

The third issue catches even the careful clients. Some clients have done the right thing and are using AI tools with contractual confidentiality and no-training protections. At least that is something. That helps a great deal with the first two issues. However, it likely does not solve the discovery problem. Every one of those tools keeps a chat history, which is a document. If a client asks an AI tool whether its contract position is weak, how bad its exposure is, or how to word an email to a terminated employee, those prompts may well be relevant, nonprivileged and discoverable. In the copyright litigation against one major AI developer, the court ordered production of a sample of 20 million de-identified user chat logs. Your client’s chats are not going to be anonymized and lost in a crowd of 20 million, they are going to be Exhibit A.

Clients need to understand that a conversation with an AI tool is not a conversation in their head. It is a written record, often stored by a third party, and it will be subject to litigation holds, preservation obligations, and document requests just like email. Retention settings, who is permitted to use these tools for legal topics, and how chat histories are preserved should be part of every company’s AI governance program.

The fourth issue is less about legal doctrine and more about the invoice. Clients are increasingly sending their lawyers AI-drafted agreements, policies, letters, and even court filings with a request to “just take a quick look.” Some of this work product is fine, but much more of it is not okay. This so-called “work product” often confidently wrong, internally inconsistent, built on the wrong jurisdiction’s law or missing the provisions that actually matter. Reviewing a bad draft is harder than writing a good one, because the lawyer must first figure out what the draft is trying to do, then determine everything that is wrong with it, then fix it without losing whatever the client was attached to. More often than both clients and their lawyers would like, that process takes longer than drafting from scratch would have.

Clients understandably expect AI to make legal work faster and cheaper, and when the lawyer uses AI, it often does. When the client uses consumer AI and hands over the result, it frequently does the opposite, and the client is then surprised by the bill. Lawyers should be candid about this at the outset, and should feel comfortable telling a client that the most efficient path is to start with the firm’s own drafting.

The fifth issue is the one I believe will generate the most friction. Clients are using AI tools to basically practice law on their own, then asking a lawyer to review and sign off on the result. The request is usually framed as a limited review, a request for a few hours, a modest fee, a quick blessing. What the client is really asking for is an insurance policy. If the AI-generated contract, filing, or compliance program fails, the client will point to the lawyer who reviewed it.

That is a mismatch between the fee and the exposure, and lawyers should not accept it. A lawyer who signs off on work product owns that work product in the eyes of the client, the court and, eventually, the malpractice carrier. A limited-scope engagement is permissible, but only if the limitation is reasonable and the client gives informed consent. An hour of review cannot reasonably support a full endorsement of a document the lawyer did not draft and may not fully understand. Firms should either scope these engagements narrowly and in writing, with clear language about what was and was not reviewed, or decline to provide the blessing at all.

None of this means clients should stop using AI. They will not, and they should not. Used properly, these tools are enormously valuable, and I spend a good part of my days helping make that happen. But clients and in-house counsel need to treat consumer AI the way they would treat any other unrelated third party that has made no promise of confidentiality, because that is exactly what it is.

There are several practical steps that clients, in-house counsel and law firms can take now.

• Companies should adopt AI policies that expressly prohibit entering privileged communications, attorney work product, and confidential legal matters into consumer AI tools, and should provide approved enterprise tools with confidentiality and no-training protections for legitimate needs.
• In-house counsel should treat AI chat histories as records subject to retention policies and litigation holds, and should think carefully about whether employees should be using AI tools to analyze legal exposure at all.
• If AI-assisted analysis is part of a legal strategy, it should be done by or at the direction of counsel, in tools selected by counsel.
• Law firms should consider addressing client AI use in engagement letters, including the risk of privilege waiver when firm communications are shared with AI tools.
• Law firms should define in writing the scope of any review of client-generated or AI-generated work product, and should price that work based on the exposure assumed, not on the number of pages handed over.
• Lawyers should have frank conversations with clients early, before the client’s AI experiment becomes the firm’s problem.

Law firms have been asked, rightly, to prove that they use AI responsibly. It is fair, and overdue, for the profession to ask the same of clients. Raising these concerns is not about protecting legal fees or resisting technology, but rather it is about protecting the privilege, protecting confidential information, and making sure that when a lawyer’s name goes on something, the lawyer actually stands behind it. While your client’s new co-counsel may be fast, cheap and available at 2 a.m., it is also not bound by any duty of confidentiality, cannot be sanctioned, and will never be named in the malpractice complaint. Somebody should explain that to the client. It might as well be us.

Reprinted with permission from the October 30 issue of The Legal Intelligencer. © 2026 ALM Media Properties, LLC. Further duplication without permission is prohibited. All rights reserved.