The SRA’s warning notice of 17 August 2026 puts two familiar obligations under fresh scrutiny: accuracy and client confidentiality. For firms using AI, both depend on knowing how work is actually done. An approved subscription and a policy will only take management so far.
A recent disciplinary decision illustrates the consequences of failing to check. In SRA v Abhishek Kumar, case 12884-2026 (currently subject to appeal), a registered foreign lawyer used AI to defend proceedings arising from a criminal conviction. When errors in his submissions were challenged, he used AI again; his response contained further false material.
The tribunal ordered his removal from the register, stating that the AI-related misconduct would independently have justified that sanction. Dishonesty was not alleged, and the tribunal distinguished citation typos from fabricated authorities and unsupported propositions.
The wider lesson concerns the arrangements around the work. Who chooses the tool, what information goes into it and who checks the result? Those questions lead directly into supervision, confidentiality, billing and the service clients receive. They also expose a potential gap that a written AI policy can easily conceal.
Who is doing the work?
A fee-earner has a demanding client, a deadline and a difficult document to draft. The firm’s approved AI tool is unavailable or clunky, but their personal AI tool works perfectly well. They intend to check the answer afterwards.
That is a plausible route into so-called ‘shadow AI’: work carried out using tools beyond the firm’s approved arrangements or oversight. The same problem can arise through meeting recorders, browser extensions and new features within familiar software. A policy that refers only to ChatGPT will miss much of it.
This is not only a junior-lawyer problem. A senior partner may know exactly what advice to give while knowing little about the unsanctioned tool being used.
Start by asking people what they use and why. You may have to offer some form of temporary amnesty to elicit complete responses. If admitting to an unapproved tool is likely to result only in disciplinary action, expect people to hold back. The exercise may reveal both a risk and a sound case for investment: staff repeatedly finding their own way to summarise correspondence or meetings are identifying work the firm could help them do safely.
But approval of the tools is only the beginning. Someone still has to check the work.
In R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin), the Divisional Court addressed verification and leadership responsibilities. The decision is useful not because it creates a special ‘AI rule’, but because it applies ordinary professional duties to a route by which unreliable material reached the court. In Al-Haroun, the client supplied research that included citations said to have been generated using publicly available AI tools. The solicitor admitted relying on that research without independently verifying it. The court made clear that the client’s errors did not absolve the legal representatives: responsibility for authorities put forward by a lawyer remained with the lawyer. It did not find the threshold for contempt proceedings met, but referred the solicitor and firm to the SRA.
Checking that a case actually exists is not enough. Does it contain the quoted passage? Does it support the proposition? Outside litigation, a lease summary might omit a crucial qualification, or a chronology might miss a changed instruction.
Effective review takes expertise, source material and time. If faster drafting encourages a firm to accept more matters, its supervisors face more work. Has their capacity increased too? Counting the minutes saved on a first draft tells management little about the cost of delivering sound advice.
AI could also help reviewers spot inconsistencies and unanswered questions. Used carefully, it can improve scrutiny. But firms must develop the judgment to challenge its answers. A junior asked to explain the reasoning and identify its sources will learn more than one asked simply whether the draft looks right.
Witness evidence needs particular care, even when the instruction is only to improve the wording. Does the revised account add facts, remove uncertainty or make the witness sound more definite? The restrictions on tampering with evidence and influencing its substance in paragraphs 2.1–2.2 of the Solicitors’ Code apply however the words were produced.
What leaves the firm?
An incorrect draft can sometimes be corrected before anyone relies on it. Disclosed information may be impossible to recover.
UK and R (on the application of Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC) adds a further dimension. In one of the two linked cases, a solicitor had uploaded draft client correspondence and Home Office decision letters to ChatGPT to improve the wording and prepare summaries. The Tribunal said that entering confidential material into a public AI service placed it in the public domain, breached confidentiality and waived legal privilege. It contrasted that position with a closed system such as Microsoft Copilot. This should not be reduced to a simple distinction between free and paid tools. The firm needs to understand the actual data journey: what the provider retains, who can access it, whether it is used for training, where it is processed and what happens when the service changes or ends.
The second case was principally about supervision. The Tribunal said it did not matter whether false citations had been inserted by a junior worker or generated by ChatGPT. The lawyer responsible for the matter had to ensure that the work was checked before it reached the court. Firms must understand how work is really being done.
The SRA’s warning notice says that even paid tools can lack adequate safeguards. An ‘enterprise’ label does not, by itself, make a tool suitable for legal work.
Firms therefore need to establish who can access information, where it goes, how long it is retained and what suppliers may do with it. A promise not to train a model on client data answers only one of those questions. Approval to summarise routine correspondence should not become permission to upload every document on every matter.
Internal access matters too. Connecting an AI assistant to the firm’s records and case-management system could expose information to colleagues who should not see it, undermining information barriers. Existing restrictions must remain effective when the method of searching changes.
Professional duties of privilege and confidentiality are distinct from data protection obligations. The ICO’s AI guidance addresses those separate obligations.
Clients need guidance as well. A client may upload advice to have it explained or translated without considering what happens to it. Explain the risk and consider whether clearer advice from the firm would reduce the need. A direct conversation with a chatbot should not be assumed to enjoy the privilege attaching to an exchange with a solicitor. You may also need to explain whether the client will incur additional charges if using AI as a ‘second opinion’ leads them to ask the firm to revisit its advice.
Who gets the AI savings?
A task used to take three hours. With AI, it takes one hour, including checking. What goes on the bill?
Under an agreement to charge for time actually spent, recording the old three hours would misrepresent the work. Paragraph 8.7 of the Solicitors’ Code requires the best possible information about pricing, likely overall cost and costs incurred. The firm cannot charge for time that has been automated away, so the client financially benefits from its investment in technology. The firm may, technically, be able to do more work to make up the difference, assuming that work is available.
An agreed fixed fee is different. Efficiency does not automatically require a reduction in fees: the firm still provides a valued service, supplies its judgment and legal expertise, and accepts professional responsibility for the work. In theory, the firm benefits financially from this type of fee agreement.
The harder question concerns incentives. A firm that encourages faster working while retaining unchanged hourly targets may penalise the very efficiency it wants.
There is a commercial opportunity here. Better knowledge of production and review costs could support fixed prices for defined services, giving clients certainty and the firm a return on its investment. Management should decide how much of the saving to retain, pass on or reinvest in quality.
It must also consider what clients need to know about AI’s role. A service sold on extensive personal attention from a named specialist deserves scrutiny if its delivery has substantially changed.
The client has AI too
Clients can use AI to challenge advice, inspect a bill or write a complaint. Some will uncover real shortcomings. Others will send pages of repetition, irrelevant authorities and unrealistic demands. The firm’s own use may also provoke complaints about inaccurate summaries or automated updates that suggest progress when none has occurred.
Complaint handlers already face rising demand. The Legal Ombudsman received 14,259 new complaints in 2025/26, up 37%, although those figures do not establish how much of the increase AI caused.
A complaint that is plainly AI-generated, with long arguments and repeated allegations, may still have merit. Its substance needs investigating. Dismissing the letter as AI-generated will not answer the grievance. Nor does its length suspend the firm’s complaints obligations: complaints must be handled promptly, fairly and free of charge, with the required information about escalation if unresolved after eight weeks. Solicitors’ Code, paragraphs 8.4–8.5.
There is a risk of both sides producing increasingly elaborate correspondence while achieving nothing. An early conversation may establish what the client actually wants and whether the matter can be resolved.
AI can help with the investigation by arranging a chronology, grouping allegations and finding correspondence. The Legal Ombudsman’s 2025/26 annual report describes its own pilots supporting investigations and service-complaint responses, with governance and quality controls. It reports reduced administration and improved consistency.
For firms, the greater opportunity may be learning across complaints. Repeated confusion about costs or missed updates at the same stage of a transaction gives management something concrete to fix.
Give compliance a say before buying new technology
These issues belong in management decisions about workload, procurement and pricing. The Code of Conduct for Firms requires effective governance and accountability for work carried out through others. The COLP can achieve little if consulted only after a supplier has been chosen.
The risks also extend beyond drafting. An AI-generated AML assessment tool needs to be tested against the underlying law, guidance and the firm’s own risk profile and appetite. A system allowed to send emails, process personal data or change records on the file requires much tighter decisions about permissions and approval than one producing text for review.
A useful response to the warning notice would be to examine a handful of completed matters where AI was used. Establish which tools were used, what information they received, who checked the result and how the work was billed. Compare that with the policy and the service promised to the client.
That exercise could expose an unsafe shortcut or identify a better way of working. Either finding would be more useful than another generic policy circulated without engaging with how the firm actually uses AI. Firms have an opportunity to improve quality and profitability together, provided management decisions account for the checking, training and client contact that good legal work still requires.
