AI for UK Law Firms: SRA Rules, Confidentiality and What to Automate

The Short Answer

UK solicitors can use AI, and most already do. What changed in 2025 and 2026 is that the regulator and the courts said plainly who carries the risk: the solicitor and the firm, not the tool. Three things follow from that:

• Every authority AI produces gets checked against an authoritative source before it goes anywhere. Courts have referred solicitors to the SRA for not doing this.
• Client material only goes into closed systems whose terms rule out training on your data, because an open tool can breach confidentiality and risk privilege.
• A written policy is not a control. The Divisional Court said guidance on its own is insufficient. Supervision, an approved-tools list and checking have to actually run.

Position as at 26 October 2026. Several of the sources below were published or updated in the last six months, and a Civil Justice Council report on AI in court documents is still due. Check the dates before relying on any of it.

What the SRA and the Law Society Say

The SRA warning notice (August 2026)

On 17 August 2026 the SRA published a warning notice on the misuse of AI. Its central line is that AI “does not diminish or transfer your professional responsibilities.” It expects client data not to be used to train AI models unless the client has explicitly authorised it, and it cites the recent cases where fabricated citations reached a court. Alongside it, the SRA reported 42 reports of potential AI misuse between July 2025 and July 2026, and said it had added AI sections to its effective-supervision guidance in June 2026.

The notice does not say whether you must tell clients you use AI. That is a firm decision, covered below.

The Code of Conduct already covers it

There is no AI-specific rule. The obligations come from the Code of Conduct for Solicitors as it already stands:

• 3.2 and 3.3 - the service must be competent and your competence kept up to date, which now includes knowing what your AI tools get wrong.
• 3.5 and 3.6 - you remain accountable for work you supervise and must make sure the people you manage are competent. A trainee using AI unsupervised is a supervision question.
• 6.3 - you keep the affairs of current and former clients confidential. This is the rule an open AI tool puts at risk.
• 8.6 - you give clients information in a way they can understand, which matters if AI shapes the advice.

The SRA's earlier Risk Outlook on AI (November 2023) made the same point: you “remain responsible and accountable for the outputs from AI”, and you must protect confidentiality and legal privilege.

What the SRA has allowed

In May 2025 the SRA authorised Garfield.Law, the first purely AI-based law firm. Its scope is narrow: debt recovery for small businesses through the small claims process. The SRA noted that it will not propose case law and will only take a step once the client has approved it. That is a useful pattern for any firm: a narrow, well-defined task, no open-ended legal research, and a human approval at each step that matters.

Law Society, Bar Council and the judiciary

• The Law Society's Generative AI: the essentials was updated in June 2026. Its checklist starts with establishing whether a tool is a closed system.
• The Bar Council updated its guidance in November 2025: responsibility for the work stays with the barrister.
• Judges have their own AI guidance (October 2025), which tells them not to put private information into public AI tools, and a position statement (October 2026) that AI must not determine the outcome of a case.
• The Civil Justice Council is consulting on AI in court documents. For pleadings and skeleton arguments it is currently minded that no extra AI-specific requirements are needed. The final report is expected later in 2026.

In Scotland the Law Society of Scotland published its own guide to generative AI in October 2024. This article follows the position in England and Wales.

What the Courts Have Done

The rules became concrete through judgments, not guidance.

Ayinde v Haringey; Al-Haroun v Qatar National Bank ([2025] EWHC 1383 (Admin), June 2025). The Divisional Court dealt with two cases where citations that did not exist reached the court. It said lawyers have a professional duty to check the accuracy of AI research against authoritative sources, that heads of chambers and managing partners must take “practical and effective measures”, and that promulgating guidance on its own is insufficient. Two solicitors were referred to the SRA. One of them had relied on material from the lay client, and the court said a lawyer is not entitled to rely on the client for the accuracy of citations.

Cork v Smith ([2026] EWHC 1199 (Ch), May 2026). A letter from a national firm quoted an insolvency rule that does not exist, an unchecked AI hallucination. The judge said lawyers cannot outsource legal research or legal reasoning to an AI, and the matter was referred to the SRA.

The lesson for a firm is that the failure was never the tool. It was a missing check between the tool and the court, and nobody senior making sure the check ran.

Confidentiality and Privilege

This is the risk most firms underestimate, because it happens before anyone sees an output.

In April 2026 the Chancellor of the High Court, Sir Colin Birss, gave a speech on privilege in the age of AI warning that “one cannot assume that confidentiality is preserved”. A speech is not binding law, but both it and the SRA notice point to a 2026 Upper Tribunal decision (UK v SSHD [2026] UKUT 81 (IAC)) where putting client documents into an open AI tool was found likely to breach confidentiality and potentially waive privilege.

Before client material goes into any tool, a firm needs answers to four questions:

1. Is it a closed system? Does anyone outside the firm see the inputs, and does the provider's contract prohibit training on them?
2. Where is it processed and stored? Several legal AI vendors state EU processing options; check what your client engagement terms and your DPIA require.
3. What is kept, and for how long? Prompts, uploaded documents and outputs are all client material.
4. Can you show what was sent? A log of which documents went to which model, on which matter, is what lets you answer a client or the SRA later.

When I built a clause analysis system for an investment fund law practice, these four answers were design requirements, not a compliance appendix: zero-retention processing with every model provider, client documents isolated per organisation at the database level, and a full audit trail of user, document and time on every interaction.

Data Protection: UK GDPR and the 2025 Act

Legal documents are full of personal data, so UK GDPR applies to most AI use in a firm. Two points are specific to AI:

• Automated decisions changed in 2026. Section 80 of the Data (Use and Access) Act 2025 replaced Article 22 of UK GDPR with Articles 22A to 22D, mainly in force from 5 February 2026. The restrictions bite on significant decisions with no meaningful human involvement, and are tighter where special category data is used. Intake triage, conflict checks or eligibility screening that decide something about a person without a human in the loop need looking at.
• The ICO's guidance is under review. Its guidance on AI and data protection was last updated in March 2023 and says it may change. Treat it as the baseline, not the last word.

Firms with EU clients should also note the EU AI Act timeline moved: the AI Omnibus put the Annex III high-risk obligations back to 2 December 2027. I cover what that means for UK companies in the EU AI Act for UK and US companies.

What Is Safe to Automate

The useful line is not “AI or no AI”. It is whether a qualified person checks each output against a source before it leaves the firm, and how bad an unchecked error would be.

The lower-risk rows are also where the time is. On the fund law system, the work was finding six clause types across 150-200 page partnership agreements and analysing each against the firm's own library. It saves 4-6 hours per agreement reviewed, and the lawyers still make every judgement call.

What an AI Policy Needs to Contain

Ayinde makes a policy necessary and not sufficient. These are the parts that turn it into a control:

1. An approved-tools list, with what each tool may be used for and on which kinds of client material. Anything not on it is not used on client work.
2. A checking rule that every case, statute and quotation is verified against an authoritative source before filing or sending, with the name of who checked.
3. Supervision written into how trainees and paralegals use AI, since Code 3.5 and 3.6 put that on the supervisor.
4. A matter-level record of which tool was used for what, so you can answer questions later.
5. A client information decision. The SRA notice does not require disclosure, but expectations are moving. In Clio's 2026 UK and Ireland survey (a vendor survey), 81% of firms said they disclose AI use while 7% of clients recalled being told, and 79% of the public wanted disclosure.
6. A DPIA for any tool that processes client personal data, revisited when the tool or its terms change.
7. A route for mistakes: who is told when an AI error is found, and how it is corrected with the client or the court.

Most firms are not there yet. The same Clio report found 17% of firms had no formal AI policy at all, and in LexisNexis's September 2026 survey (also vendor-run) 94% of 543 legal professionals said they use AI for legal work, while 83% worried about inaccurate or fabricated information and 53% about confidential data leaking.

Buying a Tool or Building One

Most firms start by buying, and that is usually right for general drafting and research. The vendors differ more on where they process data, what they publish about customers and how they ground answers than on their demos. I compare the main ones, with what each publicly states, in Harvey vs Legora vs Luminance for UK firms.

Building makes sense when the document type is specialised, the work has to happen inside Word, or the firm needs to control exactly where client data goes. The architecture for that is in my implementation guide for law firms.

Frequently Asked Questions

Next Steps

I'm an independent AI consultant and fractional CTO based in the UK, and I've built production legal AI for a law firm. If your firm is choosing a tool, writing its controls or deciding whether to build, I offer a 45-minute call at no charge to work out where AI fits and where it doesn't. Book a call, or see the legal AI case study first.

This article explains the regulatory position for planning purposes. It is not legal advice; check your own obligations with your compliance officer (COLP) or the SRA.