The SRA's warning notice on the misuse of AI, published 17 August 2026, tells firms what it is concerned about. It does not set out what an AI policy should say. This post offers one working version. Below is a complete one-page policy a firm can adapt and adopt, followed by a clause-by-clause note on what each part is doing and which regulatory expectation it answers.
The short version: the notice is outcomes-focused. In the SRA's own words, it sets "the standards we expect solicitors and firms to meet, but we do not prescribe exactly how those standards should be met in different circumstances". So no regulator will ever hand you a template. What the notice does demand is evidence of control: effective governance under paragraph 2.1 of the Code of Conduct for Firms, competent and supervised work under paragraphs 3.2 and 3.5 of the Code of Conduct for Solicitors, and confidentiality under paragraph 6.3 of both Codes. One page, signed, dated and enforced, is a simple, low-cost way to show control. Here it is.
This sits with the SRA AI warning notice, which covers the order of operations: named owner, shadow AI audit, policy, verification rule. This post fills in the policy itself.
Updated 9 October 2026. We corrected the name of the Upper Tribunal case, brought the privilege wording into line with the SRA's notice, and added clauses on compliance roles, DPIAs, personal data, telling clients, training and breach reporting. The breach clause now names the Information Commission (still known as the ICO), which took over the Information Commissioner's functions on 30 September 2026.
Why one page is enough
A firm does not need a forty-page AI governance framework to start. It needs a page every fee earner has actually read. The notice asks whether the firm can demonstrate that AI use is owned, controlled and checked. A framework nobody opened demonstrates nothing. A one-page policy that is signed, dated, circulated and enforced demonstrates exactly what the notice asks for.
Two qualifications before the template. First, a policy on top of usage you have not inventoried is decoration. If you do not know which tools your fee earners actually use, run the shadow AI audit in Part 2 of the implementation kit first; the audit's output is the approved tools register this policy points to. Second, this is a general template, not legal advice on your firm's position. Your firm decides what goes in it and owns the result. Your COLP and whoever leads on data protection should read it before it is signed.
The template
Copy everything between the lines and fill in the brackets. It fits on one page of A4 at a normal font size. If you add to it, cut words rather than clauses.
[FIRM NAME] AI USE POLICY
Version [1], approved [date]. Owner: [name, role]. Next review: [date, no more than 12 months out].
1. Who and what this covers. Everyone who works for or with the firm: partners, employees, consultants, contractors and temps. It covers every AI tool, paid or free, on firm or personal accounts and devices.
2. Ownership and roles. [Name] owns AI use in the firm. They keep the approved tools register, approve new tools, and answer for the firm's AI use to clients, insurers and the SRA. Our COLP, [name], checks that AI use fits our regulatory duties and decides what must be reported to the SRA. [Our COFA, [name], approves any tool that touches client account or billing records.] Data protection lead: [name].
3. Approved tools only. Only tools on the approved register may be used for client work. For each tool, the register records what it is approved for, the account type, the data terms reviewed, where the data is stored and the DPIA decision.
4. New tools and DPIAs. Before a tool is approved for any work involving personal data, [data protection lead] decides whether a data protection impact assessment (DPIA) is needed and writes down why. If one is needed, it is done before the tool is used. No tool is approved without written processor terms from the supplier.
5. What never goes into an AI tool. Unless the tool is on the register with data terms covering confidentiality and a commitment not to train on inputs, none of the following is entered into any AI tool: client or matter identity, client documents or data, anything privileged, anything confidential to a third party. When in doubt, it does not go in.
6. Personal data. Personal data goes into an approved tool only for the purpose the tool is approved for, and only as much as the task needs. Health data, other special category data and criminal records data need [data protection lead]'s approval each time. Our privacy notice names the AI suppliers who receive personal data, or the types of supplier.
7. Firm accounts only. AI use for firm work happens on firm-controlled accounts, never personal ones.
8. Verification. Nothing AI-assisted leaves the firm (to a court, a client, a counterparty or a regulator) until a named person has checked every citation, quotation and factual claim against the primary source. The checker's name goes on the file. Reviewed is not the standard. Verified is.
9. Telling clients. We tell clients [in our client care letter / when we open the matter] which kinds of AI tool we use on their work, and we answer their questions about it honestly. If a client's terms or instructions rule AI out, we don't use it on their matter.
10. Training. Nobody uses an approved tool for client work until they have been trained on it, covering both how to use it and their professional duties when they do. [Owner] keeps a dated record of who was trained and refreshes the training at least once a year.
11. Exceptions. Only the owner approves an exception, in writing, recorded on the register with a date and a reason.
12. Raising a problem and data breaches. Anyone who thinks AI output was wrong, or that client or personal data went somewhere it shouldn't, tells [owner] and [data protection lead] the same day. There is no blame for raising it. Concealment is the disciplinary issue, not the mistake. [Data protection lead] decides whether it is a personal data breach. If it is, and it is likely to put people at risk, we notify the Information Commission (still known as the ICO) without undue delay and, where feasible, within 72 hours of becoming aware of it. If the risk to people is high, we tell them too. The COLP decides whether anything must be reported to the SRA. [Owner] checks whether our insurer needs to be told.
13. Breach of this policy. Deliberate breach of this policy is a conduct matter.
Signed: [managing partner]. Date: [date].
Clause by clause: what each part is doing
Clause 1 (scope). The notice's confidentiality concern applies to "both paid for and free-to-use AI tools". Personal accounts and home devices sit outside the firm's control in a way a sanctioned platform does not. If the policy only covers firm tools on firm machines, it misses the use the firm can least see.
Clause 2 (ownership and roles). Paragraph 2.1 of the Code of Conduct for Firms requires effective governance structures, systems and controls. Paragraph 9.1 requires the COLP to take all reasonable steps to ensure the firm complies with the SRA's rules and that serious breaches are reported promptly. Paragraph 9.2 gives the COFA the same duty for the SRA Accounts Rules, which is why the COFA line only matters for tools that touch client money or billing. Governance that belongs to everyone belongs to no one. In some firms the owner and the COLP are the same partner. That works as long as the policy says so.
Clause 3 (approved register). This is what "control" looks like when an insurer, a client or the SRA asks. It also serves paragraph 7.2 of the solicitors' Code: you must be able to justify your decisions and actions. A register with dates and reviewed data terms is a justification. A verbal understanding is not.
Clause 4 (DPIAs and supplier terms). UK GDPR Article 35 requires a DPIA before processing that is "likely to result in a high risk to the rights and freedoms of natural persons", particularly when using new technologies. The regulator's list of processing that needs a DPIA includes innovative technology "(including AI)" when it is combined with another risk factor, such as special category data. Not every AI tool will need one. Where you decide a tool doesn't, the regulator's guidance says you "should document your reasons". Article 28 separately requires a binding contract with any supplier that processes personal data for you. The regulator's DPIA guidance is under review after the Data (Use and Access) Act 2025, so check the current version before you rely on it.
Clause 5 (confidentiality). This is the clause the notice cares most about. Its language: client information "should only be entered into AI systems where appropriate contractual, technical and organisational safeguards are in place". The notice cites UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC), where the Upper Tribunal observed at paragraph 21 that "to put client letters and decision letters from the Home Office into an open source AI tool, such as ChatGPT, is to place this information on the internet in the public domain". (The decision was reported together with a second case, R (Munir) v SSHD, but the confidentiality point arose in the first.) The Tribunal added that doing this waives privilege. That was an observation, not the issue the case turned on, and the SRA puts the consequence more carefully: privilege "may be permanently waived and unable to be recovered". Either way, there is no undoing it, which is why the default in this clause is exclusion, not permission.
Clause 6 (personal data). Client files are full of personal data, so UK GDPR applies whenever they go into a tool, as well as the duty of confidentiality. The clause ties each use to an approved purpose and the minimum data the task needs. Special category data (Article 9) and criminal offence data (Article 10) carry extra conditions, so they get their own sign-off. Article 13 requires a privacy notice to tell people who receives their data, which is why the notice should name your AI suppliers or the types of supplier.
Clause 7 (firm accounts). Consumer accounts carry consumer data terms. This clause is the cheap companion to clause 5: it removes the ambiguity about which terms govern the input.
Clause 8 (verification). The notice is explicit that accountability does not move: "AI has no separate legal personality; solicitors and regulated individuals who use AI in the course of delivering legal services remain accountable for their work and outputs, regardless of how that work has been prepared." In R (on the application of Ayinde) v Haringey LBC [2025] EWHC 1383 (Admin), fabricated AI citations were put before the court, the lawyers faced a wasted costs application, and they were referred to their regulators. The judgment notes at paragraph 29 that it is "likely to be appropriate for the court to make a reference to the regulator" where false citations are placed before the court. The named-checker rule exists so that the verification is not a culture or an aspiration; it is a name on a file.
Clause 9 (telling clients). The SRA hasn't prescribed a form of words. Its 2023 Risk Outlook on AI said "Tell clients when you will be using AI with their case, and how it will operate", and added that how you do this "is a decision for your firm to make". The Law Society's guidance says to tell clients when and how AI tools are used in their matters "where appropriate to do so". The clause turns that into a recorded decision. Check your engagement terms too, because some clients rule out AI use altogether.
Clause 10 (training). Paragraph 4.3 of the Code of Conduct for Firms requires firms to ensure their managers and employees are competent and keep their "understanding of their legal, ethical and regulatory obligations, up to date". The Law Society's guidance treats training in how to use a tool and training in using it ethically as separate, and says "both are equally important". A dated training record is the evidence.
Clause 11 (exceptions). A policy with no exception route gets quietly worked around, and then you have shadow use again. A written exception with a date and a reason keeps the exception inside the governance instead of outside it.
Clause 12 (problems and breaches). One AI mistake can trigger three separate reports, so the clause names who decides each. A personal data breach goes to the Information Commission, which took over the Information Commissioner's functions on 30 September 2026. UK GDPR Article 33 requires notice "without undue delay and, where feasible, not later than 72 hours" after the firm becomes aware of the breach, unless it is unlikely to result in a risk to people. Article 34 requires telling the people affected when the risk to them is high. Serious breaches of the SRA's rules must be reported promptly to the SRA (Code of Conduct for Firms, paragraphs 3.9 and 9.1). And paragraph 3.5 requires the firm to be honest and open with a client if things go wrong. The no-blame line matters as much as the deadlines. The notice records that the SRA has received self-reports from solicitors who relied on AI tools that generated inaccurate content, and firms find problems early only when raising them is safe.
Clause 13 (breach of policy). Without a consequence line, the document is guidance. With one, it is a policy. Keep it short.
What the notice requires, and what this template adds
Be straight with yourself about the difference, because it is the difference between compliance and comfort.
The notice requires: effective governance, supervision and controls (Firms Code 2.1, 4.3, 4.4); competent, supervised service (Solicitors Code 3.2, 3.5); confidentiality (6.3 in both Codes); properly arguable submissions to the court (2.4); and the ability to justify your decisions (7.2). Its summary line is the one to remember: "The use of AI does not diminish or transfer your professional responsibilities."
The template goes further than the letter of the notice, deliberately. The named-checker verification rule extends the notice's explicit discussion of verifying legal authorities to anything AI-assisted that leaves the firm. The blanket exclusion of client material from unregistered tools is stricter than a case-by-case risk assessment. The DPIA, personal data and breach clauses come from data protection law rather than the notice, and the training and client clauses come from the Code and Law Society guidance. Each is a cheap bright line a firm can actually enforce, and each leaves the evidence trail the notice expects firms to be able to show. Proportionate does not have to mean vague.
Putting it into practice
First, the managing partner names the owner, confirms who the COLP and data protection lead are, and fills in the brackets on the template. If you have not inventoried actual usage, run the anonymous five-question survey from the implementation kit at the same time.
Next, build the approved tools register from what the audit found. For each tool, record the account type, the data terms you reviewed, where the data is stored and the DPIA decision. Arrange a short training session for anyone who will use an approved tool. If the audit finds unapproved tools in real use, the register is how they get handled rather than hidden.
Finally, the owner signs, dates and circulates the policy with a two-line covering note, and it goes on the intranet banner or the shared drive where people actually look. Diary the review date.
If a client, an insurer or the SRA later asks how your firm controls AI use, the answer is a document, a register and a name.
The US parallel, in one paragraph
For US-qualified colleagues, the shape is identical. The American Bar Association's Formal Opinion 512 (July 2024) reaches the same destination through existing duties, including competence, confidentiality, communication with clients, supervision, candour toward the tribunal and reasonable fees. The same operational answer falls out: named ownership, controlled tools, verified output. One page works there too.
Where tooling fits
Nothing in the template requires buying anything. Tooling enters later, when the firm has a named workflow, real numbers from its register and audit, and a decision to make. The questions we collect in Answers cover that stage. On the product side, Margo, our contract-review tool, is in controlled development and not generally available. The verification and data-handling rules above are what we are building it to satisfy. You can read the approach at /margo/.
Frequently asked questions
What does this law firm AI policy template cover?
The template has thirteen short clauses. It names the person who owns AI use, plus the COLP and the data protection lead, and it covers every AI tool, paid or free, on firm or personal accounts. Only tools on the approved register may be used for client work, and firm work happens on firm-controlled accounts. Client and matter identity, client documents, anything privileged and anything confidential to a third party stay out of any tool unless it is on the register with data terms covering confidentiality and a commitment not to train on inputs. The other clauses cover DPIAs, personal data, checking AI-assisted work before it leaves the firm, telling clients, training, exceptions, raising problems and data breaches. The policy is signed, dated and reviewed at least every twelve months.
Is there a free AI policy template for law firms?
Yes. The template on this page can be copied and adapted at no charge. It is general operational guidance, not legal advice on your firm's position. Your firm decides what goes in it and owns the result, and your COLP and data protection lead should read it before it is signed.
Does the SRA require a written AI policy?
No SRA rule requires one in those terms. The notice is outcomes-focused and prescribes no format. What the Code of Conduct for Firms requires is effective governance structures, systems and controls (paragraph 2.1), and the solicitors' Code requires you to be able to justify your decisions (paragraph 7.2). A signed, dated one-pager plus an approved tools register is a simple, low-cost way to evidence both.
Can we just ban AI instead?
You can, and the notice does not stop you. The practical problem is enforcement. A ban nobody follows produces exactly the shadow use the notice is worried about.
Does the policy need to cover free tools like ChatGPT?
Yes. The notice says that both paid-for and free-to-use AI tools may not provide the contractual and technical safeguards needed to maintain client confidentiality. The Upper Tribunal's comments in UK v SSHD [2026] UKUT 81 (IAC) were about client and Home Office letters put into ChatGPT. Free and personal accounts are where a firm has least control, so the template puts them in scope from the first clause.
Who should own AI use in a law firm?
One named person. Usually a partner, such as the managing partner or the COLP. In a two-partner firm that means one of the two, not "both of us", because a question with two owners has none when an insurer or the SRA asks it.
How often should the policy be reviewed?
At least every twelve months, and sooner if the firm adopts a new class of tool, has an incident, or the regulatory position moves. Keep old versions dated rather than overwriting them; the version history is itself evidence of governance.
Do we need a DPIA before using an AI tool?
Not always. UK GDPR Article 35 requires one before processing that is likely to result in a high risk to people. The regulator's list says AI and other innovative technology need a DPIA when combined with another risk factor, such as special category data. If you decide a tool doesn't need one, write down why. The DPIA guidance is under review after the Data (Use and Access) Act 2025, so check the current version.
Who do we report a personal data breach to now the Information Commission has taken over?
The Information Commission (still known as the ICO), which took over the Information Commissioner's functions on 30 September 2026. UK GDPR Article 33 still requires notice without undue delay and, where feasible, within 72 hours of becoming aware of a breach, unless it is unlikely to result in a risk to people. Its guidance and breach reporting are still on ico.org.uk. Whether the SRA also needs to know is a separate question for your COLP.