The SRA's misuse of AI warning notice: what small firms must do

The SRA's Misuse of AI warning notice, published on 17 August 2026, is presented by the SRA as guidance on existing obligations, not a new rule. The SRA says it will have regard to the notice when exercising its regulatory functions. For a firm of one to ten people the part that bites is client data. The notice expects contractual, technical and organisational safeguards before client information goes into an AI system, and below is a four-point check you can run against any provider's terms.

What the notice is, and what it is not

The SRA frames it as guidance it will have regard to, and not as a rule. Its notice of 17 August 2026 says that failing to give it proper regard puts you at risk of disciplinary action. The notice applies to what it calls all firms and individuals the SRA regulates, and it includes a section for in-house solicitors.

The duties it leans on are old ones. For solicitors, Code for Solicitors paragraphs 3.2 and 3.5 (competence and supervision). For firms, Code for Firms paragraph 2.1 on governance, systems and controls, plus paragraphs 4.3, 4.4 and 9.1 on supervision and the COLP's duty. It also names SRA Principles 1, 2, 4, 5 and 7. If you have read our piece on the SRA's wider guidance on AI, nothing here will shock you. What is new is how specific the notice gets about where client information goes.

The client-data paragraph a small firm should print out

The paragraph to keep on your desk is the one on client information. The notice says it should go into AI systems only with appropriate contractual, technical and organisational safeguards, and that the firm should satisfy itself of four things: the data stays in a secure environment, unauthorised third parties cannot access it, it is not used to train models unless explicitly authorised, and it is not kept longer than necessary.

Those are outcomes. The notice does not tell you where to find the evidence, so here is where we would look. These are checks against the notice, not statements about any vendor's actual terms. Your answers will depend on the document in front of you.

A four-point vendor-terms check for a small firm

Secure environment. Look in the data processing agreement, not the security page. You want a stated hosting location and a named list of sub-processors, with a security commitment written into the contract and not offered as a description of current practice. If there is no DPA to read, you have your answer. Our guide to the data processing agreement for AI tools shows what a usable one contains.

No unauthorised access. Two questions. Who at the provider can read your inputs, and under what conditions? And what are the sharing defaults inside your own account? A tool that lets staff share conversations by link, or that puts everything in one open workspace, can fail this point even when the provider behaves perfectly. In a firm of five, the account settings are your job and nobody else's.

No training without authorisation. Find the sentence that says inputs are not used to train models. Then find the exceptions. Wording about using data to improve the service is the warning sign, and the notice itself says provider terms may allow entered data to be used to improve the tool. Check whether the position is a contractual default or something you had to switch off, and whether that switch applies to every user on the account.

Retention. You need a stated retention period, deletion on request, deletion when the contract ends and a position on backups. Look for carve-outs for abuse monitoring or logs, which can keep your inputs longer than the headline period. If the terms say nothing about how long data is kept, the provider has not given you what you need to be satisfied about under the notice.

Free and paid tools: the notice does not let either off

The notice says both free and paid tools may lack adequate safeguards, and that provider terms may allow entered data to be stored, retained or used to improve the tool. Read that against every vendor pitch that says its business tier is safe. We will not tell you any tier is compliant, because compliance is the result of the four checks above on the terms you actually signed. A price tag is not a safeguard.

The stakes are not abstract. The SRA says privilege may be permanently waived where client information goes into a tool without adequate safeguards, and the notice cites the Upper Tribunal's remarks on putting client letters into an open source AI tool. Our piece on AI and privilege waiver covers the US case law that runs the same way.

Court material and supervision

On the courts, the notice is blunt. Authorities must be genuine, relevant and verifiable (Code paragraphs 2.2 and 2.4), and it says false material put before a court may amount to contempt. It cites R (Ayinde) v Haringey LBC [2025] EWHC 1383 (Admin), which the notice describes as involving fabricated citations, wasted costs and regulatory referrals, and says reliance on AI output is no defence. It also cites BCP v A Mother [2026] EWFC 71 (B), where it says an unregistered barrister misled the court through AI hallucinations without deliberate intent; Cork and another v Smith [2026] EWHC 1199 (Ch), on the responsibilities of authorised persons when unchecked AI errors occur; and Brett v SRA [2014] EWHC 2974 (Admin), for the point that misleading the court is among the most serious offences. Sanctions elsewhere are in our hallucination sanctions piece.

On supervision, the notice says Authorisation of Firms Rule 9.4 requires at least one person who has practised as a lawyer for at least three years to supervise regulated work. It adds that the three-year requirement does not apply to all supervisors, so firms should check that every supervisor has appropriate experience. Our reading is that this matters most when a trainee or paralegal is the one using AI, because the supervisor owns the output either way. That is our inference, not the SRA's statement.

Enforcement is live. The SRA's news release of 17 August 2026 reports 42 reports of potential AI misuse between July 2025 and July 2026, with several investigations ongoing. No outcomes are stated, and we will not guess at them.

Want the next SRA change in your inbox before it hits the trade press? Subscribe to the newsletter and we will send you the next one.

FAQ

Our free guide sets out what to lock down before client material goes near a model, and the tool picks are where to start once your checks are done.