Most firms that get into difficulty with AI do so because nobody set rules the team can follow. The tools arrived by the side door, in a browser tab or a personal account, and the duties they touch were left to each person's judgement. This guide is for the partner who has to fix that. It explains what the regulators and the courts now expect a firm to have written down, the twelve sections a policy needs, how to write one in a day, and how to keep it from going stale. A free template that follows the same structure is on the template page.
Why the policy became urgent in 2026
For two years the position was guidance and encouragement. The Solicitors Regulation Authority published compliance tips on using AI, the judiciary issued its guidance for judges in December 2023 and refreshed it on 31 October 2025, and the Information Commissioner kept a hub of AI and data protection guidance up to date. None of it was law, and a small firm was entitled to treat it as reading for later.
That changed in the second half of 2026. On 17 August the SRA issued a warning notice on the misuse of AI, which is a different kind of document. Guidance explains what the rules mean. A warning notice sets out the conduct the regulator expects and states that a firm failing to have proper regard to it risks disciplinary action. The SRA published it alongside the figure that explains the timing, being 42 reports of possible AI misuse received between July 2025 and July 2026, with investigations open. Three days later the regulator's annual assessment of continuing competence recorded that it now looks at whether solicitors are aware of and use warning notices when planning their learning. On 3 September 2026 the Solicitors Disciplinary Tribunal published its judgment in the case of Abhishek Kumar, the first time a lawyer's use of AI in proceedings had been litigated before it. He had put fabricated authorities before the tribunal, corrected them with the same kind of tool and introduced fresh errors, and told the regulator he lacked the expertise to verify the output. He was struck off.
Earlier in the year the SRA had rewritten its supervision guidance and expanded it from nine pages to twenty four, with the core statement that firms may use AI tools provided an authorised individual remains accountable for the process and the work produced. Put those documents side by side and the message is consistent. Nobody is asking you to stop. Everybody is asking you to show what sits behind the use, and the policy is where that showing begins.
What the rules require of the policy
The Standards and Regulations do not mention large language models and do not need to. The SRA Principles require you to act with integrity, in a way that upholds public trust, and in the best interests of each client, and none of that changes because a machine produced the first draft. The warning notice names the paragraphs of the Codes already in play, being competence, effective systems for supervising client matters, governance and controls, and confidentiality, and ties each to a failure the regulator keeps meeting. False material reaching a court comes first, with the flat requirement that every authority put before a judge is checked as genuine and accurately cited before filing. Client information comes second, with contractual, technical and organisational safeguards expected before client material enters any tool. Supervision comes third, and the notice shuts a door some firms had leant on, because the supervisor of a junior who used the tool answers for the output as fully as the person who wrote the prompt.
The SRA's compliance tips add two expectations the policy should carry. A risk and impact assessment belongs on file before a tool is adopted rather than after something has gone wrong, with the compliance officer answerable for it. And a firm should make clear where a client is interfacing with AI, which a plain line in the client care information covers for the general position.
The ICO's guidance reaches the same tools from the data protection side. Matter files are dense with personal and special category data, so a tool that touches them sits at the sharp end of the UK GDPR, and where the processing is likely to pose a high risk to individuals the ICO expects a data protection impact assessment before it starts. The Data (Use and Access) Act 2025 received Royal Assent on 19 June 2025, so expect the ICO's pages to keep moving. The assessment the ICO expects and the one the SRA expects are much the same exercise, and the policy should say who does it and where it is kept.
The courts complete the picture. The judiciary's guidance tells judges to stay alert to the use of AI by those appearing before them and lists the tell tale signs, from authorities that cannot be found to American spellings in an English filing. The Civil Justice Council's working group signalled in July 2026 that it saw no need for a new AI declaration on professional drafting, because a document bearing a solicitor's name already carries that person's responsibility for its content. Responsibility was never going to move to the software. The policy's job is to make sure the person who carries it knows they do.
The twelve sections your policy needs
The Adliora template runs to twelve short sections. Here is what each one is for and the decision behind it.
1. Purpose
One paragraph saying why the firm has the policy and what it protects. Clients, staff and the firm, in that order, and the firm's professional obligations underneath. A reader who stops here should know the policy is about duties rather than about software.
2. Who it covers
Everyone. Partners, solicitors, paralegals, trainees, consultants, administrative staff and anyone working for the firm on a temporary basis, for any use of AI on firm business, whether the tool sits inside a system the firm pays for or is a free service on the open web. The gap most policies leave is the personal account on a personal phone, and this section closes it.
3. Principles
Four lines that the rest of the document expands. A person stays responsible, so the named fee earner answers for every piece of work that leaves the firm. Client information is protected, so no client data goes into a tool until the firm knows where it goes and who can read it. Output is checked before anyone relies on it. The firm is open with clients about the part AI plays. If a member of staff remembers nothing else, these four are enough to keep them out of trouble.
4. Approved tools
A list, by name, of the tools staff may use and what each is approved for, such as research support, first drafts and summaries. Then the process for adding one, being a short assessment of what the tool does with client data, where it falls short and who signs it off, recorded and kept. This is the section that satisfies the SRA's expectation of assessment before adoption and the ICO's expectation of an impact assessment, and it is the section the warning notice has in mind when it speaks of governance and controls. Staff must not use a tool for firm work unless it is on the list.
5. What you must not do
Five prohibitions in plain words. No client names, case detail or confidential information in a tool not approved for that use. No reliance on output without checking it against a reliable source. No AI drafted work presented as final without a competent person reviewing it. No use of AI to reach a decision about a client that a person should make. No signature on a statement of truth for a document whose contents have not been verified. The instruction to use your judgement is not a policy. These five are.
6. Confidentiality and client data
Every prompt that carries client information is a disclosure to whoever runs the service behind the tool. Before client data touches any AI tool, staff confirm that the tool is approved for that data, that the provider does not use inputs to train its models, and that access stays limited to people entitled to see the information. Special category data, such as health or criminal information, gets extra care and a check with a named person first. The AI note taker that joins a video call deserves its own line here, because it records a confidential and often privileged conversation and sends it to a provider to process, and a consumer tool on a personal account answers none of the questions above.
7. Checking what AI produces
Every output is a first draft. Legal content is checked against a reliable source before it is used, figures are checked against the file, and a competent person reviews the work before it reaches a client or a court. The Irish High Court's practice direction of 1 September 2026 put a point here that English firms should adopt as their own, that running an output through the same or another AI tool does not amount to checking it. If your people cannot open the report and read the passage, they have not verified anything.
8. Telling clients
The client care information explains the firm's general use of these tools. On a matter where AI does more than sit in the background, the client gets a short note. The client should always know that the firm, and not a tool, carries responsibility for their work, and who to speak to if something goes wrong.
9. Supervision and responsibility
Supervisors know which of their staff use which tools, and for what. A named authorised person stays accountable for AI assisted work in the same way they are accountable for the work of a junior. Where a supervisor has any doubt about a tool or an output, they stop and raise it before the work goes out. This is the section that answers the SRA's expanded supervision guidance, and the difference it draws is between supervision that is real, meaning a person who looked at the work and understood what the tool did, and supervision that is a name on a file.
10. Training
Every member of staff who uses AI for firm work understands enough about the tools to know what they do well and where they fail. The firm provides training on the policy and on safe use, and repeats it when tools or duties change. Competence with these tools is part of keeping knowledge and skills up to date, which is why this section should be read beside each solicitor's continuing competence record.
11. Reporting concerns and breaches
If anyone thinks a tool has been misused, that client data has gone somewhere it should not, or that AI output has led to an error, they report it at once to a named person. The firm treats reports openly and without blame, because catching a problem early protects the client. Where a data breach has happened, the firm follows its breach procedure, including whether the ICO and the affected client must be told.
12. Review
At least once a year, and sooner when the tools or the rules change. A named owner, the current version, the approval date and the next review date recorded at the top of the document.
Writing it in a day
The policy does not need a project. It needs a day with the right people and the right inputs. Start with an amnesty rather than a hunt. Ask every member of staff, without consequence, what tools they use and for what, so the picture you write from is honest. Most firms find more tools in use than the partners knew about, and a policy written for the tools the partners imagined will be ignored in favour of the ones people have.
Gather the retention and training terms of everything already in use, because that is where the confidentiality half of the warning notice bites hardest in a small practice, and it is what section four is built from. Have the client care letter to hand for section eight, and the firm's breach procedure for section eleven. Then sit the partners and the compliance officer down with the template and make each bracketed decision in turn. Which tools are approved, for what. Who assesses a new one. Who is the named person for special category data, for doubts and for reports. When the review falls.
Write in plain English and keep it to two or three pages. A new paralegal should read it on a Monday morning and know by lunchtime what they may and may not do. Have it approved by the person who owns compliance in the firm, record the date, and put the approved version where every member of staff can find it.
Rolling it out
A policy nobody has read protects nobody. Put it in front of the whole firm with an hour of training, working through the approved tools, the five prohibitions and the checking step with examples from your own practice. Add it to induction so good habits form from the first week. Brief supervisors separately on what to look for in AI assisted work, being confident wording that misstates the law, citations that do not exist and summaries that miss the point, because a polished draft is not a checked one.
Update the client care information in the same week, and decide how the firm records that a person checked each piece of AI assisted work before it went out. A line on the file, a sign off field in the case management system, or a note in the drafting record all serve, provided the firm can point to it when asked. The ability to demonstrate supervision is what turns a private habit into something you can put in front of the regulator.
Keeping it alive
The tools move faster than any policy, so the review is where most policies fail. Fix the annual date and add three triggers that bring the review forward, being a new tool or a material change to one already in use, a new warning notice or guidance from the SRA, the ICO or the courts, and any incident reported under section eleven. Each time, ask whether the approved list still describes what people use, whether the checking step still works, and whether the training record shows that the people using the tools have been taught what changed.
That last point ties the policy to each solicitor's continuing competence record. The SRA's 2026 assessment reviewed 123 training records across 66 firms and found that almost half failed to explain how a learning need was identified and why it mattered. A firm that put AI into drafting or research this year has changed how the work is done and what goes wrong in it. An hour spent writing down, for each person, what changed in their work, what that exposed as a gap and what they did about it turns the policy's training section into evidence rather than intention.
The mistakes we see most often
The first is a policy that tells people to exercise judgement without saying what the firm has decided. The second is a ban, which drives use out of sight rather than out of the building. The third is a policy written and filed, never trained and never mentioned again, which the regulator will read as evidence that the firm knew the risk and did nothing about it. The fourth is the assumption that a paid product is safe by default, when the warning notice is explicit that paid products store, retain and train too. The fifth is checking AI output with AI, which the Kumar judgment shows ending a career. Each of these is avoided by a policy that names tools, names people and names the checking step, and by a firm that reads its own policy once a year.
Common questions
Does a small firm with two partners need a written AI use policy?
Yes. The SRA's warning notice of 17 August 2026 asks every firm to show the oversight and controls behind its use of these tools, and a firm holding that in the owner's head has nothing to show. Two pages, signed off and known to the staff, meets the point. A thick manual nobody reads does not.
Can we ban AI instead of writing a policy?
You can write that down, but it will not describe what happens in your firm. Fee earners paste drafts into free chatbots at home and paralegals run note takers in client calls whether or not the partners have decided anything. A ban without approved alternatives drives that use out of sight, which is the worst place for it. A policy that names the tools people may use, and for what, brings it back under the firm's control.
Is a paid legal AI product safe to use without a policy?
No. The warning notice is blunt that paid products as much as free ones store what you type, keep it longer than you assume, or use it to train a model. The policy is what records that someone read the retention and training terms, decided the tool was fit for the work, and named who checks its output.
How often should the policy be reviewed?
At least once a year, and sooner when the tools you use or the rules that govern them change. A new warning notice, a new model handling client data, or a change of supplier each count as a trigger. The SRA now assesses whether solicitors are aware of and use warning notices when planning their learning, so the review date and the training record belong together.
Where to go next
Take the free template and adapt it in an afternoon. If your people already use tools the partners have not approved, the AI Readiness Audit finds out what and where, and the Staff AI Literacy Session is the half day that puts the finished policy into practice. Adliora is a management consultancy, not a law firm, and where a question needs formal legal advice you should take it from your regulated adviser.
