Guides
AI for law firms: what works, what the SRA expects
How UK law firms use AI, what the SRA's 2026 warning notice requires, and how to give Claude or ChatGPT safe, permissioned access to client files.

UK law firms get the most from AI when it reads firm knowledge and every answer traces to a source, with each person seeing only what they are cleared to see. The SRA’s warning notice of 17 August 2026 names two failures: fictitious citations reaching a court, and confidential client information entered into AI systems without proper safeguards. A good AI set-up for a firm answers both. We built Louvain as permissioned company memory for that job: it connects your email, documents, meetings, chats and calendar, and every answer is cited to its source.
TL;DR#
- Firms use AI today for summarising, drafting, document review, research and internal knowledge questions. Every output still needs a lawyer to check it.
- The SRA says solicitors stay accountable for AI-assisted work. Its main concerns are false citations and client confidentiality.
- Keep client data out of AI tools unless contractual, technical and organisational safeguards are in place.
- Write down your AI policy, your approved tools and your checking step. Code provision 7.2 expects you to show compliance.
- Louvain gives Claude, ChatGPT and your team cited answers with permissions enforced fact by fact. Plans start at £20 per person per month.
What UK law firms actually use AI for#
The use cases are less exotic than the headlines. The Law Society lists contract analysis, drafting or summarising documents, e-discovery, client chatbots and internal knowledge databases among them. It also lists predicting case outcomes and creating images, which are rarer.
A October 2025 special in Briefing, a UK legal technology magazine, gives a picture of day-to-day use. UK firms use generative AI for summarisation, drafting, rephrasing, proofreading, legal document review, standard playbooks and prompt libraries. Bird & Bird uses agentic AI for multi-step research, with structured reports and inline citations. (Agentic means the AI carries out several steps in a row, instead of answering one prompt.)
The same report describes one firm that automated 1,360 documents, saving 700 hours and about £150,000. Its NDA agent saved up to 60%. Treat those as one firm’s results, not a benchmark for yours.
Three larger moves show where the market is heading:
- Gowling WLG announced on 21 January 2026 a firm-wide rollout of a legal generative AI platform across its UK LLP business. More than 200 lawyers had piloted several solutions in autumn 2025.
- Addleshaw Goddard chose a platform for firmwide rollout on 9 November 2026, after early use of its own tool, AGPT.
- Keystone Law reported revenue growth of almost 18% in April 2026 while rolling out agentic tools. These help lawyers query its operating manual, find sources of funds for AML checks and improve conflict checking.
Notice what the Keystone tools have in common. Two of the three read the firm’s own information: its manual and its client records. That is the part of AI work most firms underestimate. A generic chatbot knows nothing about your clients, your matters or your internal rules. The value comes when AI can read your own material.
What the SRA expects from you#
The SRA’s warning notice is the document to read first. It applies to every solicitor, but supervisors and compliance officers carry the most weight.
The notice says solicitors remain accountable for AI-assisted work. It ties that to specific provisions of the SRA’s Principles and Code of Conduct:
| Provision | What it covers |
|---|---|
| 1.4 | Not misleading the court |
| 2.4 | Properly arguable submissions |
| 2.5 | Contempt |
| 3.2 | Competence |
| 6.3 | Confidentiality |
| 7.2 | Justifying decisions and demonstrating compliance |
All of this is in the SRA’s warning notice. The regulator is also watching. It received 42 reports of potential AI misuse between July 2025 and July 2026. Investigations concern inaccurate citations, supervision and confidentiality. In June 2026 it expanded its supervision guidance to cover AI.
Provision 7.2 matters most in practice. An analysis in The Legal Brief says it “does most of the practical work”. A firm that can produce a written policy, an approved-tools list, supervision records and a verification step is answering the question in advance.
If you are a barrister or work with chambers, the Bar Standards Board has its own AI guidance of May 2026. It asks barristers to consider risks to legal professional privilege and contempt, and to make sure tools comply with GDPR. CILEx Regulation encourages its firms to document AI use, safeguards and oversight, and to consider publishing an AI statement for clients.
What goes wrong: the Pinsent Masons case#
The clearest lesson comes from a court. In Cork & Anor v Smith (May 2026), two misleading AI-generated letters reached the High Court. Judge Mullen said the ruling should serve as a public admonishment of Pinsent Masons and senior lawyers. The firm referred itself to the SRA and met its former client’s costs, including the cost of finding new lawyers.
The judge’s findings are specific. The junior solicitor “almost entirely outsourced the thinking process” to the AI. They did not check references, even when the AI itself warned them to. They did not tell supervisors that AI had been used.
Three habits would have stopped this:
- Every citation is checked against its source before anything leaves the firm.
- Staff say when AI was used, and supervisors can see it.
- The tool shows where each statement came from, so checking takes seconds, not an afternoon.
The judge also said AI “does not, at least at present, do away with the need for proper research and thought” from a legal professional. That answers the question many partners ask: will AI replace lawyers? On current evidence it changes the work, and the accountable person stays a lawyer. Addleshaw Goddard’s Kennedy and Bird & Bird’s Santos both describe a lawyer-in-the-loop approach for quality control, risk management and client comfort.
Confidentiality: the part most firms get wrong#
Most firms do not start with a court filing. They start with a partner pasting a client email into a chat window. That is where the SRA’s second concern bites.
The SRA says client information should only go into AI systems where appropriate safeguards are in place. The notice says those safeguards must:
- protect confidentiality and keep data secure and away from unauthorised third parties
- prevent the data being used for training, except where explicitly authorised
- limit retention to no longer than necessary
The Law Society’s 30 June 2026 guidance adds practical steps. Define your use cases. Establish your rights over prompts, training data and outputs. Check whether a tool is closed or trains third-party models. Tell clients how you use AI. Assess whether the input data is appropriate.
There is a second confidentiality problem inside the firm. A matter may be ring-fenced because of a conflict. An HR file is for two people only. A trainee should not see a partner’s pricing notes. When you connect an AI tool to everything, it can surface any of these to anyone who asks the right question. A permission check on the folder is not enough, because the answer is built from many documents at once.
A safe rollout in six steps#
Here is the order we would follow in a firm of 10 to 200 people.

1. Write the policy first#
One page is enough. Say which tools are approved, which matters or data types are off limits, who checks output, and who signs off AI use in court documents. This is the record that provision 7.2 expects. CILEx Regulation does not require a standalone AI policy, but it does encourage documenting use and oversight.
2. Pick a small set of use cases#
Start with work where a wrong answer is caught easily: summarising a long email chain, finding what was agreed in a meeting, drafting a first-pass note. Bevan Brittan’s Roberts advises firms to look for tools that fit real use cases rather than adopt technology because everyone else has.
3. Decide who may see what#
List the groups: partners, fee earners, support staff, finance, and any matter-specific teams. Decide the rule for each. Then check that your AI tool enforces those rules itself instead of relying on people to behave.
4. Connect the tool to the firm’s own knowledge#
Generic models do not know your clients. Addleshaw Goddard’s research found that optimised retrieval and better prompting raised accuracy in commercial contract review from 74% to upwards of 95% on average. Retrieval means the AI looks up the relevant passages before it answers. How you feed the AI material matters as much as which model you pick.
5. Require a source on every answer#
If a statement has no source, it does not go to a client or a court. This one rule would have prevented the Pinsent Masons problem. CILEx Regulation says AI content should be reviewed and verified before it is relied on, shared with clients or submitted to a court, with the depth of review matched to the task and the cost of error.
6. Log it and review quarterly#
Keep a record of what was approved and what went wrong. Ask staff what they use. Update the policy. A partner who can show that record has an answer when the regulator asks.
Where Louvain fits#
Louvain is permissioned company memory for growing teams. It connects your email, documents, meetings, chats and calendar, and turns them into checked facts. Each fact is cited to its source.
Permissions are enforced on every individual fact. Each person, and each AI agent, sees only what they are allowed to see. That is the control described in step 3, built into the tool instead of left to habit.
For a firm, this has three practical effects:
- You can ask questions about clients, deals and decisions and get answers traced back to the source. That supports step 5.
- You get a daily brief on where each matter stands and what is owed, so partners do not need status meetings or chasing in chat.
- Claude and ChatGPT can use it as a connector. They get safe, permission-aware access to company knowledge. A model asked a question on behalf of a trainee sees what the trainee may see, not what the partner can.
The security and data handling page explains how Louvain handles data, and the subprocessors list names the third parties involved. Read both with your compliance officer before you decide, because they feed directly into the SRA’s contractual, technical and organisational safeguards.
Choosing a plan#
Louvain is priced for growing teams, not for firms with a large enterprise budget. The Team plan costs £20 per person per month. The Business plan costs £40 per person per month. Enterprise has no published price. Full details are on the pricing page.
For a 25-person firm on the Team plan, that is £500 a month. The Business plan for the same firm is £1,000 a month.

If you are rolling out Claude or ChatGPT across your firm, start with the pricing page and choose the plan that fits your team. The library has more reading on how Louvain works.
Why the firm’s own memory is the real asset#
Bird & Bird’s Santos observes that many clients prefer to reach generative AI through their law firm, because “it’s easier to justify higher legal invoices than new software budget allocations”. That makes firms AI implementation partners for their clients. It also raises the stakes. If clients rely on you to run AI safely, the standard you set internally becomes part of the service.
The Law Society lists the risks plainly: IP infringement, data protection and privacy, cybersecurity, training data, misleading or false outputs, bias, and human-resources and reputation risks. It also says generative AI can raise efficiency and cut costs where processes can be automated. Most of those risks come down to two questions. Where did this answer come from? And who was allowed to ask?
A firm that can answer both has what it needs to use AI with confidence. It can say yes to Claude and ChatGPT because the layer underneath decides what they read.
Beyond law firms#
The same logic holds for accountancies, consultancies and property managers, who also hold confidential client data. A founder who wants fewer status meetings needs the same thing: one place where commitments, decisions and deadlines are cited to their source. The regulator differs. The problem of permissions and proof does not.
FAQ#
Can a law firm use ChatGPT or Claude with client data?#
Yes, if the safeguards the SRA describes are in place. These cover contract, technology and organisation: confidentiality, no training use unless explicitly authorised, secure handling and limited retention. Check the terms of the specific product and record your decision in line with provision 7.2.
Will AI replace lawyers in the UK?#
The evidence so far points to changed work, not replaced lawyers. A High Court judge said in 2026 that AI does not remove the need for proper research and thought by a legal professional. Firms describe a lawyer-in-the-loop model for quality control and client comfort.
Do we need a written AI policy?#
The SRA’s provision 7.2 expects you to justify decisions and show compliance, so a written policy, approved-tools list and checking step are the practical answer. CILEx Regulation does not require a standalone policy but encourages documenting AI use, safeguards and oversight. Keep it short and review it regularly.
How much does Louvain cost?#
Team is £20 per person per month and Business is £40 per person per month. Enterprise has no published price. See the pricing page for what each plan covers.
Sources#
- SRA: Misuse of AI warning notice
- The Law Society: Generative AI, the essentials
- Briefing, AI special, October 2025
- Gowling WLG: rollout of a legal AI platform across its UK LLP
- Legal Technology: Addleshaw Goddard firmwide rollout
- Legal Futures: Keystone Law revenue and agentic AI
- New Law Journal: SRA reports and investigations
- The Legal Brief: SRA AI warning notice analysis
- Bar Standards Board: AI guidance, May 2026
- CILEx Regulation: Guidance on the use of AI
- Law Gazette: Pinsent Masons and the SRA
- Addleshaw Goddard: LLMs in legal document review