AI for Law Firms: Where It Is Safe to Use
AI for law firms means using AI in legal practice: researching the law, reviewing and drafting documents, summarising files and correspondence, managing intake and running the business of the firm. The useful question is not whether it helps, since it plainly does, but where it is safe to use. Legal work has an unusual combination of properties: confident errors are easy to produce, hard to spot, and publicly punished when they reach a court.
That combination is why a law firm needs a map rather than a policy of yes or no. Some uses are safe with ordinary care. Some are safe only with a specific verification discipline. Some are not ready. The frame for the wider sector is set out in AI for professional services firms; this briefing applies it to law.
What the evidence says about errors
The most careful study so far is from Stanford’s RegLab, published in 2025 in the Journal of Empirical Legal Studies. It tested the AI research tools of the two market-leading legal research providers, which are built on their own case law databases specifically to reduce errors. It found they still produced hallucinated or misleading answers between 17% and 33% of the time. The best-performing tool answered 65% of queries accurately. Grounding the model in real law reduced errors against a general chatbot. It did not remove them.
The consequences are now on the public record. The AI Hallucination Cases database maintained by researcher Damien Charlotin had logged 2,046 court and tribunal decisions involving AI-generated fabricated or misrepresented material by September 2026, 815 of them involving lawyers rather than self-represented litigants. Sanctions range from costs orders and fines to bar referrals, and the lawyer’s name is in the judgment.
The lesson is not that legal AI is useless. It is that every tool, including the specialist ones, produces text that looks exactly as reliable when it is wrong as when it is right.
A map of where it is safe
Safe with ordinary care. Work where an error is visible and cheap: summarising the firm’s own documents and correspondence, producing first drafts of routine letters, structuring intake information, preparing chronologies from documents in the file, and administrative tasks such as scheduling and billing narratives. A lawyer reading the output would notice most mistakes.
Safe only with a verification discipline. Work where an error looks plausible: legal research, drafting submissions and advice, reviewing contracts against a standard, and due diligence. These are among the most valuable uses and the ones behind almost every sanctioned filing. They are safe when every authority is opened and read in the primary source, every quotation is checked word for word, and a named lawyer signs off on having done so.
Not ready. Anything where AI output reaches a client or a court without a lawyer reviewing it, anything that asks a tool to decide rather than prepare, and any use of a tool whose terms allow it to train on or retain client material without adequate protection.
The professional rules already apply
Lawyers do not need new rules to know what is expected. The existing duties cover it.
In the United States, the American Bar Association’s Formal Opinion 512, issued in July 2024, applied the existing Model Rules to generative AI: lawyers must understand the tools well enough to use them competently, protect client confidentiality, which for some tools means obtaining the client’s informed consent before entering their information, communicate with clients about AI use where it matters to them, supervise the people and systems doing the work, and charge reasonable fees, which generally rules out billing clients for time spent learning a tool.
In Germany, lawyers are bound by professional secrecy under Section 203 of the Criminal Code, and may give service providers access to client information only under Section 43e of the Federal Lawyers’ Act, which requires a confidentiality obligation agreed in text form. An AI vendor processing client material is such a provider. Other jurisdictions differ in detail and agree in substance: the lawyer remains responsible, and the confidentiality chain extends to the tool.
The practical questions about where data goes, who can see it and what happens when you leave are set out in data sovereignty, and the register of approved tools belongs to AI governance.
The verification discipline, concretely
Firms that use AI for research and drafting without incident tend to share the same habits.
- Every authority is opened. Not searched for, opened and read, in an authoritative source, to confirm it exists, says what it is cited for and is still good law.
- Every quotation is checked word for word. Invented or altered quotations appear in hundreds of the logged decisions, often attached to real cases.
- The reviewer is named. A specific lawyer confirms the checks were done, on the file.
- The prompt and output are kept. If a question arises later, the firm can show what the tool produced and what the lawyer changed.
- Juniors do the checking, supervised. It is the modern version of the research training they used to get, and it teaches the same judgment.
None of this is new. It is what a careful partner always expected of a trainee’s research note. AI makes it compulsory rather than aspirational.
The commercial pressure
Clients are pushing from the other side. Thomson Reuters’ Future of Professionals report 2026 found 77% of clients consider AI-enabled quality improvements from their law firms very important or essential, while only 5% say most or all of their firms deliver them. And 71% of in-house legal professionals expect their outside firms to change commercial models as AI use grows, against 28% of firms that have changed pricing so far.
Taken together, that means a firm that avoids AI risks losing work, a firm that uses it carelessly risks sanctions, and the firm that uses it with a visible verification discipline has something to say in a pitch that most competitors cannot.
Where to start
Begin in the safe zone, where the gains are real and the risk is low: summarising files, drafting routine correspondence, structuring intake. Put an approved tool with proper confidentiality terms in place first, so nobody needs a consumer account. Write the verification rules down before anyone uses AI for research or drafting, and treat them as a condition of use rather than a guideline.
Then tell clients what you do and how it is checked, because they are already asking. As with every firm, the harder work is bringing the people with you, not choosing the software.
If you want to know where AI belongs in your firm’s work, tell Daniel which matters you would start with, what is already in use, and what the firm needs it to do. Book 30 minutes with Daniel.
Frequently Asked Questions
Is it safe for law firms to use AI?
For many tasks, yes, with the right controls. Summarising the firm’s own documents, drafting routine correspondence and structuring intake are low risk. Legal research and drafting submissions are valuable but safe only when every authority and quotation is checked in the primary source and a named lawyer signs off. Output should never reach a client or court unreviewed.
Do legal AI research tools still hallucinate?
Yes, less than general chatbots but not rarely. A Stanford RegLab study published in 2025 found the AI research tools of the two leading legal research providers produced hallucinated or misleading answers between 17% and 33% of the time, with the best tool accurate on 65% of queries.
How often have lawyers been sanctioned for AI errors?
Often enough to be tracked systematically. Damien Charlotin’s AI Hallucination Cases database had recorded 2,046 court and tribunal decisions involving AI-generated fabricated or misrepresented material by September 2026, 815 of them involving lawyers. Sanctions include fines, costs orders and bar referrals, with lawyers named in the judgments.
What do professional rules say about lawyers using AI?
Existing duties apply. The American Bar Association’s Formal Opinion 512 of July 2024 addresses competence, confidentiality, client communication, supervision and reasonable fees. In Germany, lawyers are bound by professional secrecy under Section 203 of the Criminal Code, and service providers with access to client data must be bound under Section 43e of the Federal Lawyers’ Act.
Can law firms bill clients for AI-assisted work?
They can bill for the work, but fees must be reasonable. ABA Formal Opinion 512 indicates that lawyers billing hourly should charge for time actually spent and generally not for time learning a tool. Many firms are reconsidering pricing, and Thomson Reuters found 71% of in-house legal teams expect outside firms to change commercial models as AI use grows.
Where should a law firm start with AI?
In the low-risk zone: file summaries, routine correspondence and intake. Put an approved tool with proper confidentiality terms in place first, write down the verification rules before AI is used for research or drafting, and tell clients how AI is used on their matters and how it is checked.
About the Author
Daniel Förster is the Managing Partner of Thane Alaric. He has built nine businesses, delivered dozens of projects and worked with more than 2,000 founders, across AI and technical systems of every kind, funnels, marketing, sales and leadership. He works owner to owner, and Thane Alaric runs on the same systems it builds. Business-to-Life. We build the business your vision needs.
