AI for law firms: where does it help, and where can’t it be trusted?

AI for law firms is useful for first passes: reviewing a document against a checklist, drafting from your own precedents, summarising bundles and correspondence, organising client intake and searching your firm’s know-how. It is not safe as a source of law. Every authority it produces must be checked against an authoritative source, and a solicitor remains responsible for everything that leaves the firm. This guide is for small and mid-sized firms in England and Wales that want to use AI without creating a new risk.

AI for law firms: where it helps first

• First-pass document review. Read a contract or lease against your own checklist and flag clauses that look unusual, missing or one-sided, for a lawyer to judge.
• Drafting from precedents. Start a first draft from your approved precedents and the facts of the matter, rather than from whatever a general tool has seen on the internet.
• Summarising bundles and correspondence. Turn a long chain of letters or a large bundle into a chronology, a list of issues and the points still open, each tied back to the page it came from.
• Client intake. Gather the facts of a new enquiry in a structured form, spot what is missing and route it to the right team for a conflict check and a decision.
• Internal know-how search. Answer fee earners’ questions from your own guides, templates and practice notes, pointing to the source document.

In each case AI does the gathering and the first draft. The legal judgement, and the decision about what the client is told, stays with a lawyer.

Where it must not be trusted

Generative AI produces text that reads well and can still be wrong. It can cite cases that do not exist, attach the wrong outcome to a real case or misstate what a statute says, with the same confidence as when it is right. This is not a theoretical risk. In June 2025 the Divisional Court, in Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin), dealt with two cases in which citations to authorities that do not exist had been put before the court. The judgment warns that AI tools may cite sources that do not exist, says lawyers have a professional duty to check research against authoritative sources, and expects those with leadership responsibilities, such as managing partners, to take practical steps.

The working rule is simple. Treat any authority, quotation or statutory reference that came from AI as unverified until a person has read it in an authoritative source. A tool that links each answer to its source makes checking quicker. It does not remove the need to check.

AI is also weak where the answer turns on facts that are not in the documents, on the commercial context of a deal, or on a recent change in the law. Those are exactly the places where a client is paying for judgement.

Confidentiality and privilege

Client confidentiality is a professional duty, not a software setting. In August 2026 the SRA published a warning notice, Misuse of AI, which says client information should only be entered into AI systems where that is appropriate. It warns that public tools may store or retain what is entered, or use it to improve the tool, and that sharing confidential client information through AI tools may mean legal professional privilege is waived and cannot be recovered.

Before any client material goes near an AI tool, settle these points in writing:

• Which tools are approved for client data, on business terms, and which are not.
• Where each provider stores and processes data, how long it keeps it and whether it is used to train models.
• Who in the firm can use each tool, and on which kinds of matter.
• What must never go into a general tool, such as privileged advice, special category data or material covered by a court order.
• How clients are told about your use of AI, in your terms of business or privacy notice.

The guide to using ChatGPT with customer data covers the basics that apply to any business, and the free AI acceptable use policy template gives you a staff policy to adapt for the firm.

Supervision and professional duties

Using AI does not change who is responsible. The SRA’s warning notice makes clear that solicitors remain accountable for their work however it was prepared, including AI output and work done by people they supervise, and that supervisors can be in breach if inaccurate material reaches a court without adequate review.

• Name a partner who owns AI use in the firm and keeps the approved tool list current.
• Treat AI output like a trainee’s first draft: useful, and reviewed by someone qualified before anyone relies on it.
• Make checking citations a fixed step in your filing and sign-off process, not a matter of personal habit.
• Train everyone who uses the tools, including support staff, on what they are for and where they fail.
• Keep a record of which tool was used on which matter, so you can answer questions later.

The Law Society’s guide Generative AI: the essentials is a useful primer. This article is practical guidance, not legal advice. Your obligations depend on your firm and your work, so read the SRA’s material directly and take advice where you need it.

Choosing tools or building

The options fall into three groups: legal-specific AI products such as Harvey, AI features inside the practice management and document software you already use, and general assistants such as ChatGPT or Microsoft 365 Copilot. Some firms also build their own tools around their precedents and know-how.

For most small and mid-sized firms a sensible order is to use what is already in your systems, under terms you have checked; add a legal-specific product where it clearly beats that on work you do often; and build only when the task is specific to your firm. A review checklist for your own practice area, a drafting flow from your own precedents or a know-how assistant over your own notes are good examples. That kind of build is what agents and automation work covers. Whichever route you take, ask how the tool shows its sources, where client data goes and how you would switch it off.

What I have built

I designed and built The Counsel, a legal assistant for the laws of England and Wales. It brings together specialist skills and agents that help people read, draft and check legal documents: catching risky clauses, preparing a clean first draft, giving a second opinion before signing, and reviewing one document from several angles at once.

The skills engine behind it is published on GitHub as UK Legal Skills, so anyone can read how it works. It runs in Claude Code as /legal commands: 38 skills, each for a specific job such as contract review, lease and tenancy checks, employment contracts, AML and KYC reviews or a UK GDPR audit, plus 12 specialist agents that the larger reviews run in parallel. For a law firm, five points stand out:

• It checks the law it relies on. Skills that depend on recent reforms check on legislation.gov.uk whether a provision is in force before treating it as binding, and its case law search runs against the National Archives’ Find Case Law service rather than what the model remembers. A lawyer still reads every authority before anyone relies on it.
• It is narrow on purpose. It covers England and Wales only, and every output opens with a statement that it is AI-generated analysis, not legal advice.
• One skill reviews a firm’s own AI use. Run as /legal ai-compliance, it reads an AI usage policy or a client-facing AI disclosure and checks it against the SRA Standards and Regulations, ICO guidance on AI, the UK’s AI principles and the EU AI Act, then lists the gaps to close.
• Every step can be read. Each skill is a plain Markdown file, so a partner or compliance lead can see exactly what it is told to do before anyone uses it.
• The licence is source-available, not open source. The repository’s summary says you can use it inside your firm and when serving your own clients, but not to build a competing product or hosted service. Read the licence file for the exact terms.

The Counsel is a public product rather than a tool built for one firm, but it reflects how I approach legal work: narrow skills for specific jobs, clear review and drafting flows, and plain-English explanations of what the tool has done.

If you are weighing up AI for your firm and want a second opinion on where to start, book a free 30-minute AI consultation. Bring one task that takes up fee earners’ time and we can look at whether AI would help, with no obligation.