How UK law firms adopt AI safely under SRA duties
How law firms adopt AI without breaching SRA confidentiality and privilege duties: the real risks, the controls that make it safe, and where it earns its place.
Ibrahim Mizi
· 9 min read
Updated 
OpenKit helps UK law firms adopt AI without breaching their SRA confidentiality and privilege duties. The short version: keep client data inside the firm’s own environment, verify every AI output against the source before it reaches a client, and record where AI was used. The tool is rarely the risk. The controls around it are what make adoption safe.
We give no legal advice and hold no legal accreditation. This is a practical guide to the technology and the governance, written for the partner or COLP who has to sign off on AI without exposing the firm.
Is it safe for a UK law firm to use AI?
Yes, when the firm treats AI output the way it treats a trainee’s first draft: useful, fast, and not to be sent to a client unchecked. The SRA regulates outcomes rather than tools, so the firm stays accountable for confidentiality and quality whether the work was done by a person or a model. Safety is a property of the workflow, not the software.
That is the whole argument of this post. The rest sets out what the SRA and Law Society actually expect, the three risks that sink firms who skip the controls, and the specific governance that makes adoption defensible at renewal.
What do the SRA and Law Society expect?
The SRA does not ban AI and does not tell firms which tools to use. Its risk outlook on the use of AI in the legal market puts it plainly: “Our regulation focuses on the outcomes that firms achieve, and not necessarily on the specific systems they use to achieve them.” The same report says solicitors remain responsible and accountable for the outputs of the AI they use, and it groups the risks it sees under headings that include errors and hallucination, confidentiality and privacy, bias, accountability, and crime. Supervision stays with the firm.
Two things follow from that, and it is worth keeping them apart, because a lot of AI guidance blurs the line. The binding duties are the ones your firm already has under the SRA Standards and Regulations, confidentiality under paragraph 6.3 of the Code of Conduct for Firms among them, and they apply to AI-assisted work exactly as they apply to a trainee’s draft. Everything else is guidance rather than rule: the SRA’s compliance tips for solicitors say it should always be made clear to clients where they are interfacing with AI, and the Law Society’s Generative AI: the essentials walks through the vendor-contract and data-protection questions worth asking before a tool goes anywhere near a matter. Keeping a record of where AI touched a matter is not a rule anyone has written down, but it is what lets the firm answer the question a year later, which is why we build the logging in by default.
The three risks that actually matter
Most of the danger in legal AI concentrates in three failure modes. Each is well documented, and each has a known control. The table below is the honest version of what goes wrong and what stops it.
| Risk | What goes wrong | The control that makes it safe |
|---|---|---|
| Hallucination | The model invents plausible but false case law, statutes, or contract terms that read as authoritative | Only trust outputs that cite a verifiable source; a qualified person checks every citation against the original |
| Confidentiality leakage | Matter data pasted into a public AI tool that may retain or train on the input, breaching privilege | Run AI under contracts that forbid training on your data; keep privileged material inside the firm tenant |
| Data residency | Matter data leaves the UK or EU, complicating UK GDPR transfers and failing client procurement checks | Put a lawful transfer mechanism in place for anything leaving the UK, and document where the model, the data and the logs each sit |
| Unrecorded use (shadow AI) | Fee-earners use consumer tools privately, so the firm cannot evidence what AI touched a matter | Provide a sanctioned tool, set an acceptable-use policy, and log AI interactions per matter |
The hallucination risk is not hypothetical. In June 2025 the Divisional Court dealt with two cases in which fabricated authorities reached the court, five non-existent cases in one and eighteen fictitious or misquoted authorities in the other, and referred the lawyers involved to the Bar Standards Board and the SRA (Ayinde v Haringey and Al-Haroun v Qatar National Bank). The court warned that in the worst cases misuse of AI can amount to contempt. That is why the working rule across the profession is that AI drafts and a person verifies, and a system that cites its source on every output makes that verification a thirty-second check rather than a research task.
How do firms stop AI inventing fake case law?
By refusing to trust any AI output that cannot point to where it came from. The practical control is citation-anchored retrieval: the model answers only from documents the firm has given it, and every answer links back to the exact paragraph it came from, so the fee-earner verifies against the source rather than against the model’s confidence.
This is the core of how we build. On the BAiSICS commercial-lease platform, every extracted field traces to a citation in the source document, and the model is held to refusing out-of-scope questions rather than guessing. That single design choice is what separates a tool a partner will sign off from one that quietly creates liability.
How do firms keep client data confidential and in the UK?
By keeping privileged material inside the firm’s own environment and never routing it through a tool that might train on it. The control is partly contractual and partly architectural: run against models deployed in a UK or EU region under terms that forbid training on your inputs, or run open-weight models on private infrastructure the firm controls.
Where a matter or a client’s procurement contract demands it, the data never leaves the firm tenant at all. The BAiSICS platform runs entirely inside an AWS UK region, which is the kind of concrete answer a client’s data-protection officer or your professional indemnity insurer will ask for before they are comfortable.
Where AI earns its place in a law firm
Not every task needs AI, and the safest adoption starts with the work where the controls are easy and the payback is obvious. Document-heavy review, where the answer always lives inside a source the firm already holds, is the natural first use case, because citation back to source is built in.
The clearest example we can point to is BAiSICS, a custom OCR plus bespoke LLM pipeline built for commercial-lease review. It took lease review from around two hours to roughly ten minutes at 96% accuracy against benchmarks marked up by senior partners, beating GPT-4 and the leading legal-AI products in the same test. The gain came not from a flashier model but from constraining the system to the firm’s documents and citing every field.
What governance does a firm need before it adopts AI?
Less than most firms fear, but it has to exist before the first matter touches a model. A short acceptable-use policy, a sanctioned tool so fee-earners stop reaching for consumer apps, a record of where AI is used per matter, and a named owner (usually the COLP) who can answer the SRA and the insurer. That is the floor.
OpenKit’s role is to build the tools and the audit trail that make this practical, not to act as your regulator. We hold ISO 27001, ISO 9001, and Cyber Essentials independently and operate to UK GDPR, so the security posture under the AI is documented from the start. We work alongside the firm’s COLP on supervision and design delivery to help the firm meet its own SRA obligations, never to discharge them on the firm’s behalf. For the full picture of how we approach legal work, see our legal AI page, and for where any engagement begins, our AI audit maps the workflows worth automating before anything is built.
Is it safe for UK law firms to use AI?
Yes, when the firm keeps client data inside its own environment, verifies every AI output against the source, and records where AI was used. The SRA regulates outcomes, so the firm stays accountable for AI work the same way it is for a junior. Safety comes from the controls around the tool, not the tool itself.
What does the SRA say about AI in law firms?
The SRA does not ban AI. Its risk outlook on AI in the legal market says its regulation focuses on the outcomes firms achieve rather than the specific systems they use, so the firm stays accountable for confidentiality and for the quality of its advice however the work was produced. That report groups the risks it sees under several headings, among them errors and hallucination, confidentiality and privacy, bias, and accountability. The SRA's separate compliance tips add that it should be clear to clients where they are interfacing with AI.
Can AI breach client confidentiality or legal privilege?
It can, if matter data is pasted into a public AI tool whose provider may retain or train on inputs. That is the shadow-AI risk most firms underestimate. The fix is to run AI against models in a UK or EU region under contracts that forbid training on your data, with privileged material kept inside the firm tenant.
What are the main risks of using AI in legal work?
Three dominate: hallucination, where the model invents plausible but false case law or contract terms; confidentiality leakage through public tools; and data residency, where matter data leaves the UK or EU. Each has a known control. Citation to source, private deployment, and UK-region hosting address them in turn.
How do firms stop AI inventing fake case law?
By only trusting AI outputs that cite a verifiable source and treating anything uncited as a draft to check. In June 2025 the Divisional Court referred lawyers in two cases to their regulators after fabricated authorities reached the court, and warned that misuse of AI can end in contempt proceedings. The working rule that follows is that the AI drafts and a qualified person verifies every citation against the original.
Does AI for law firms need to keep data in the UK?
UK GDPR does not require it. What it requires is a lawful transfer mechanism for personal data leaving the UK, which for most European destinations is covered by the UK's adequacy regulations and elsewhere usually means the ICO's international data transfer agreement. Most firms and their professional indemnity insurers ask for UK hosting anyway, because it settles the question instead of documenting it. OpenKit deploys legal AI inside a UK region by default, as on the BAiSICS platform.
Has OpenKit built AI for UK law firms?
Yes. The BAiSICS commercial-lease platform is a custom OCR plus bespoke LLM pipeline that took lease review from around two hours to roughly ten minutes at 96% accuracy against partner-graded benchmarks, beating GPT-4 and leading legal-AI products in the same test. It runs entirely inside an AWS UK region with citation on every extracted field.
Settle the rules before you buy.
If this raised a question about what your own team is allowed to do, the answer is an AI Charter. It is four decisions with names against them, short enough that people actually read it. One Charter, agreed once, governs every system we build with you and every one your team runs after we have gone.
Keep reading.
The closest pieces to this one, by shared ground.
