AI for law firms: use cases, rules and an AI policy (2026).
What AI does well for law firms, which kind of AI fits which legal task, how to stop fake citations, what the SRA and ABA rules say, and what a firm AI policy should cover.
For law firms, AI works best on high-volume reading and first-draft work: due diligence, contract review, drafting from the firm’s own precedents, and research with a citation for every statement. A lawyer reviews everything before it leaves the firm. Start with one practice group and one matter type, check every authority against an authoritative source, and agree a short AI policy before anyone pastes a client document into a tool.
The workspace we built for a law firm’s legal research and drafting cut routine drafting time by 90% and made contract due diligence three times faster, inside firm-specific guardrails. But law is unforgiving of errors. A system that invents a case or misses a clause causes real harm, so how the system is built matters more here than almost anywhere.
What does AI do well for law firms?
Due diligence and contract review
Reading data rooms and contract sets, extracting the clauses that matter (change of control, termination, liability caps, exclusivity, assignment), flagging deviations from your positions, and producing a report with every finding linked to its source page. Construction contracts and subcontracts raise their own questions, covered in AI in construction.
First drafts from your own precedents
Drafting letters, clauses, memos and standard agreements from the firm’s own templates and past work, in the firm’s style, for a lawyer to review and finish.
Research with citations
Answering questions from your knowledge base, practice notes and the sources you subscribe to, with a citation for every statement, and saying plainly when the sources don’t support an answer.
Matter and knowledge management
Summarising long matter files and correspondence chains, and making the firm’s accumulated know-how searchable in plain language.
Client intake and time narratives
An assistant on the website, email or WhatsApp can collect the facts of a new enquiry, gather the details a conflict check needs and route the enquiry to the right team, without giving advice. At the other end of a matter, a system can draft time-entry narratives from the activity on the file for the fee earner to approve and correct.
Which kind of AI fits which legal task?
Not every legal task needs a language model, and some are safer without one. A useful way to decide is to match the task to the kind of system, the data it needs and the way it fails.
| Task | Right kind of AI | Data it needs | Main risk |
|---|---|---|---|
| Data room and contract review | Document AI to extract clauses, plus a language model to compare them with your playbook | The contracts, your playbook and past reports | A missed or misread clause, so test on a sample before relying on it |
| First drafts | Language model with retrieval over your precedents | Templates, past matters, house style | Drafting from an outdated or wrong precedent |
| Legal research | Retrieval over authoritative sources, with a citation check | Legislation, case law, practice notes you are licensed to use | Invented or misapplied authority |
| Client intake | Conversational agent with strict limits on what it may say | Intake questions, practice areas, referral rules | Giving advice, or taking in confidential detail it shouldn’t |
| Conflict checks | Deterministic matching against the client and matter database | Client, matter and party records | A fuzzy match that misses a name, so the final call stays with a person |
| Limitation periods and court deadlines | Rules written as code and checked by a lawyer | Court rules, dates, jurisdiction | A confident but wrong date from a language model |
| Time narratives | Language model drafting from activity on the file | Time entries, emails, documents touched | Narratives that don’t match the time actually spent |
Buy a legal AI product, or build?
Ready-made legal AI products are strong general tools, and many firms should start there. A custom system makes sense when:
- Your know-how is the advantage. The system should draft from your precedents, playbooks and positions, not generic ones.
- The workflow is specific. A particular kind of diligence, a regulatory filing, or a high-volume matter type with its own steps and checks.
- It has to live in your systems, such as the document management system, practice management platform and data rooms, with your permissions and ethical walls respected.
- Clients require control over where their data is processed and who can see it.
Many firms do both: a general tool for everyday drafting, and a custom system for the workflows that define their practice. Whichever route a firm takes, the Law Society’s generative AI guidance for solicitors (30 June 2026) asks it to check whether a tool is a closed system or also trains models for third parties, and to document inputs, outputs and errors where the tool keeps no history.
How do you stop AI from citing cases that don’t exist?
This is the failure that has already reached the courts. In England and Wales, the Divisional Court’s June 2025 judgment in Ayinde v Haringey and Al-Haroun v Qatar National Bank arose from fake citations and quotations put before the court. It states that freely available generative AI tools are not capable of reliable legal research, and that lawyers who use AI for research have a professional duty to check it against authoritative sources, such as the government’s legislation database, the National Archives judgments database, the official law reports and the databases of reputable legal publishers.
The consequences have continued. The Solicitors Regulation Authority’s warning notice on the misuse of AI (17 August 2026) says that using AI does not reduce or transfer a professional’s responsibilities, and that failing to have regard to it risks disciplinary action. The Law Society’s guidance reports that in April 2026 a judge referred two solicitors to the SRA after AI-assisted court documents contained incorrect citations and cases that did not support the submissions.
In the US, ABA Formal Opinion 512 cites a Stanford study that found leading AI legal research tools hallucinated between 17% and 33% of the time. A legal AI system has to be designed so that a fabricated authority can’t reach a lawyer’s signature:
- Grounded answers only. Every statement comes from a retrieved source and links to it. No source, no statement.
- Citations resolved in code. Each case name, neutral citation and quotation is looked up in the source library. A citation that doesn’t resolve, or a quotation that isn’t in the cited text word for word, fails the check and the draft is not released as final.
- Abstention. When the evidence is weak, the system says so rather than filling the gap. Our note on retrieval that knows when to say “I don’t know” covers how.
- Evaluation on real matters. A test set of real questions and documents, with answers checked by lawyers, run before every change. See evals before features.
- Review built into the workflow. Drafts and findings are clearly marked as AI-generated until a lawyer approves them. ABA Opinion 512 suggests marking AI material in the file so later users know it can be wrong.
In legal work, an AI that says “I couldn’t find support for that” is more valuable than one that always has an answer.
What do the professional rules say?
This section describes what regulators and courts have published as of October 2026. It is not legal advice, and a qualified professional should decide how the rules apply to your firm.
England and Wales
- Accuracy and duty to the court. The SRA’s notice says solicitors should be assured that every case authority they cite is genuine, relevant, has a verifiable citation and advances the argument put forward. Supervisors remain accountable for the work carried out through them, and reliance on AI output is not a suitable defence if false citations reach the court.
- Confidentiality. The notice says client information should go into an AI system only where contractual, technical and organisational safeguards exist: kept in a secure environment, not used to train models unless explicitly authorised, and not kept longer than necessary. It records that the SRA has seen confidential client information entered into public AI tools.
- Rules still moving. The Law Society reports that the Civil Justice Council consulted in February 2026 on whether civil procedure rules should require transparency about AI use in court documents. Watch for the outcome.
- The bench. Judges have their own AI guidance (31 October 2025), which covers hallucinations and keeping private information out of public tools.
United States
- Competence and supervision. ABA Formal Opinion 512 (29 July 2024) says lawyers need a reasonable understanding of a tool’s capabilities and limits, must review its output to an appropriate degree, and remain fully responsible for the work. Managerial lawyers must set clear policies on permissible use and make sure people are trained.
- Confidentiality and consent. Because many self-learning tools can expose one client’s information to other users, the opinion says a client’s informed consent is required before client information is put into such a tool, and that boilerplate in an engagement letter is not enough.
- Candor to the court. Outputs must be checked before filing, and the opinion tells lawyers to consult each court’s local rules on AI use.
Italy and the EU
- Telling clients. Italy’s AI law (Law 132/2025, in force since 10 October 2025) requires professionals to tell clients about the AI systems they use, in clear, simple and complete language, and limits AI to instrumental and supporting activities, with the professional’s intellectual work prevailing (Article 13).
- The EU AI Act. The European Commission’s AI Act page says it entered into force on 1 August 2024 and became generally applicable on 2 August 2026, with some exceptions and later dates for high-risk systems. It matters for firms with clients in the EU. Our EU AI Act guide for companies building AI covers the builder’s side, and the UK rules for businesses cover the UK.
Documents are untrusted input
A contract or email can contain hidden instructions aimed at the AI. Treat every document as data, never as instructions (see prompt injection is an input-validation problem), and enforce ethical walls and matter permissions in the system itself, as they are in the firm.
Do we have to tell clients, and how should we bill?
It depends on the market. In Italy, telling clients is a legal requirement. In the US, ABA Opinion 512 says lawyers must disclose AI use if a client asks, if the engagement terms require it, when client information goes into a self-learning tool, and when the tool is relevant to the basis of the fee or influences a significant decision. In England and Wales, the Law Society says the SRA had no specific guidance on disclosure for client care when it wrote (June 2026), and advises agreeing with clients whether and how AI is used.
On billing, ABA Opinion 512 is plain: a lawyer who bills by the hour must bill the time actually spent. If drafting with AI takes fifteen minutes of input plus review, those minutes are what can be charged. The engagement letter is the natural place to say how AI is used and what the client has instructed.
Where AI is the wrong tool, and a lawyer stays in charge
- Legal advice and judgment. ABA Opinion 512 says lawyers may not leave it to AI alone to advise clients or negotiate their claims. The system prepares; the lawyer decides.
- Anything filed or sent to a client. A lawyer reads and signs it, with citations verified.
- Deadlines and limitation periods. Calculate them with rules in code, then have a lawyer confirm.
- Conflict decisions. The system can search and match, but a person clears the conflict.
What should a law firm AI policy cover?
A small firm doesn’t need a long document. The sources above point to a short policy that answers these questions, written before people start experimenting on their own:
- Which tools are approved, and who checked their terms on retention, training on your data and who can see inputs. ABA Opinion 512 says to read the terms of use and privacy policy of any tool before using it.
- What may go into which tool. Separate public tools, approved tools under contract, and the firm’s own system, and say whether client-confidential or privileged material is allowed in each.
- What must be verified. Every citation and quotation, every figure and every date, before anything goes to a court or client.
- Who supervises and who is trained. Supervisors stay accountable for work done with AI, and everyone who uses it is trained on its limits.
- What is recorded. Inputs, outputs and errors, and a mark on AI-generated material in the file.
- What clients are told, in the engagement letter and when asked, and how time is billed.
- What happens when something goes wrong. A named person, a way to report a mistake quickly, and a rule on when to tell the client or the regulator. The SRA notice mentions self-reports from solicitors who relied on inaccurate AI output.
- When the policy is reviewed. The tools change quickly, so set a date.
How to start and what to measure
Due diligence and first-draft generation from precedents are usually the best first projects. They’re high-volume and repetitive, easy to check, and the time saved is easy to measure. A first project can run like this:
- Pick one practice group and one matter type, and a lawyer who owns the result.
- Collect real examples: past contracts, precedents and the reports your team produced, with the answers a lawyer would accept.
- Test before trusting. ABA Opinion 512 describes testing a tool on a smaller set of documents by comparing its summaries with a manual review, which is the right shape for the first evaluation.
- Run alongside the current process with a lawyer reviewing every output, then reduce review only where the numbers support it.
- Write the AI policy in the same weeks, not afterwards.
Track a handful of numbers against how the work is done today:
- Lawyer hours per first draft or per diligence report.
- Review time per document.
- Errors found in review: missed clauses, wrong precedents, citations that failed the check.
- Share of flagged issues that a lawyer confirms as real.
- How often the system abstained instead of answering.
If you want help designing a system like this around your own documents and systems, our services page describes how we work.
Questions law firms ask
Can AI review a whole data room?
Yes. Extraction and flagging across thousands of documents is where AI saves the most time. Lawyers review the flagged issues and a sample of the rest, instead of every page, once the system has been tested against a manual review of a smaller set.
Will it work with our document management system?
Usually, through its APIs, with your existing permissions carried through so people only see what they’re allowed to. Integration with an older document or practice management system is often the biggest piece of work.
Who is responsible for AI-drafted work?
The lawyer who reviews and signs it, exactly as with work drafted by a junior. The SRA’s notice and ABA Opinion 512 both say the responsibility stays with the professional. The system’s job is to make that review fast and well-informed.
Can we paste client documents into a public chatbot?
Not without the safeguards regulators describe. The SRA says client information should go into AI systems only where contractual, technical and organisational protections exist, and ABA Opinion 512 requires informed client consent for many self-learning tools.
Does a small firm need an AI policy?
Yes, and it can be short. Managerial lawyers in the US are expected to set clear policies on AI use, and the SRA expects effective governance and supervision of AI risks in England and Wales. A policy also gives staff a clear answer on which tools they may use.
Modulus Labs AI builds research, drafting and due-diligence systems for law firms and legal teams, grounded in your own documents and guardrails. Tell us about the work you want to speed up, and we’ll reply within one business day.