Before solicitors use AI on client work, a UK firm should confirm that the tool's terms and settings protect confidentiality, that a competent person verifies every output, how clients will be told, where data is processed and stored, and that the insurer and firm policies cover the use, then record the purpose, a usage log and an exit route.
None of this is new law. The Law Society's generative AI guidance for firms is explicit that professional obligations apply whether AI was used by the solicitor personally or by anyone under their supervision, and it was written for small and medium firms in particular. What the checks below add is a way to turn that into evidence: a short form, signed before the first matter, that shows the firm thought about each risk.
The eight checks on one sheet
The Law Society's lifecycle considerations for SMEs run through preliminary considerations (starting with purpose), data and training, procurement, implementation and usage, risk management, review, and communication. For a small firm deciding whether to switch a tool on, those compress to eight checks. Each has a question, a way to verify it and a piece of evidence to file.
| Check | Question to answer | Evidence to file |
|---|---|---|
| 1. Purpose | Which specific task is this tool for, and which tasks are excluded? | One-line use case and an excluded list |
| 2. Confidentiality | Can client information reach the supplier's staff, other customers or model training? | Terms and privacy settings, printed with the date |
| 3. Data location | Where is data processed and stored, and can we choose? | Supplier's written answer and the region setting |
| 4. Supervision | Who checks outputs, and how, before anything leaves the firm? | Named reviewer per task type |
| 5. Client care | How will clients be told, and have any asked us not to use AI? | Client care wording, opt-out record |
| 6. Insurance and contract | Does our insurer accept this use, and what liability does the supplier accept? | Insurer's written reply, contract clauses |
| 7. Records | Can we show what was put in and what came out, if a question arises later? | Log of inputs and outputs or tool history |
| 8. Exit | Can we remove our data and stop using the tool without losing records? | Export and deletion terms |
Check 1 and 2: purpose and confidentiality
Start narrow. "Drafting first versions of routine client updates from a fee earner's own bullet points" is a purpose. "Use AI for legal work" is not. Write the exclusions at the same time: no uploading of client documents until the data checks pass, no use for authorities or citations without independent verification, no use on matters under an information barrier.
For confidentiality, read three things and put the date on them. First, the supplier's terms for the exact plan the firm will buy, not the website's headline. Second, the data-use statement: does the supplier train models on your content? When checked in November 2026, OpenAI (for ChatGPT Business), Anthropic (for Claude Team) and Microsoft (for Copilot used with work accounts) each stated that business content is not used to train their models by default, while consumer plans may require you to switch training off; verify it for the plan you actually choose, because plan names and defaults change. Third, the settings: who can see conversations, whether history is kept, and whether admins can turn off sharing. The guide to stopping AI tools training on your business data walks through the settings.
The simplest way to avoid the whole problem on day one is to start with tasks that carry no client information, such as a template for a newsletter or a checklist of what to ask at a first meeting. For anything else, the question of whether solicitors can use ChatGPT without breaching confidentiality sets out the reasoning in more detail.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowCheck 3: where the data goes
The Law Society's guidance lists the geographic location of data among the contract terms to review, alongside data rights, intellectual property and liability, and says to confirm that the use and storage of data comply with data protection laws, including UK GDPR where it applies. Three questions turn that into something you can ask a supplier:
- In which country is our data processed, and in which is it stored at rest?
- Can the firm choose the region? Some suppliers offer it only on higher tiers: OpenAI's business pricing comparison, checked at the same time, lists data residency, with the UK among its regions, for ChatGPT Enterprise but not for ChatGPT Business. Confirm it for the specific plan, and whether the choice covers only storage or also processing.
- Which sub-processors handle the data, and how are we told when they change?
Keep the supplier's written answer, because a statement on a sales call is not evidence. If a supplier cannot say where data is processed, that is your answer. When last checked (an archived copy from October 2026), the ICO's guidance on AI and data protection carried a notice that it is under review because of changes made by the Data (Use and Access) Act, so treat any summary you find online, including this one, as a starting point and check the regulator's current pages.
Check 4: supervision that is more than a glance
The Law Society's guidance ties AI use to the duty to supervise work for clients and to the duty not to mislead clients or the court. Practically, that means each task type has a named checker and a described check. A table keeps it honest:
| Task type | Reviewer | What the review consists of |
|---|---|---|
| Client update draft | The fee earner who owns the matter | Read against the file; every fact, date and figure checked |
| Summary of a document | Fee earner | Spot-check three points against the original; read the sections the summary skipped |
| Research or authorities | Supervising solicitor | Every authority opened and read in a trusted source; none taken on the tool's word |
| First draft of a clause | Fee earner, then supervising partner on non-routine matters | Compared with the firm's precedent and the instructions |
The reviewer must be competent in the subject, not merely senior. A partner who does not practise in employment law cannot supervise an AI-drafted employment letter just by being a partner. For authorities in particular, the Law Society's guidance cites an April 2026 ruling in which two solicitors were referred to the regulator over AI-related conduct in court papers, which is a reminder of what an unchecked citation can lead to. The routine for checking is in how to stop AI inventing case law.
Check 5: client care and what clients are told
The Law Society's guidance says the regulator did not, when it was written, have specific guidance on generative AI use or disclosure for client care. It advises firms and clients to decide together whether and how AI tools might be used in providing advice and support, and notes that clear communication prevents misunderstandings about how information is produced. Verify that statement against the current page before you rely on it, because it is the sort of line that changes.
In practice, a short paragraph in the client care letter is the simplest way to do it. An illustrative version:
"We may use approved AI tools to help with tasks such as drafting and summarising.
Any such use is supervised by a solicitor, who is responsible for the final work.
We do not put your confidential information into any tool that has not been approved
by the firm, and we will not use AI on your matter if you ask us not to. Please tell
us if you have any concerns."Two practical points go with it. Keep a record of any client who opts out, and make sure fee earners can see it on the matter. And do not promise what you cannot guarantee: "we never use AI" is a risky sentence in a world where a spell-checker, a transcription tool and an email client may all contain AI features. The wording of client care letters is covered in more depth in AI and client care letters for solicitors.
Check 6 and 7: insurance, contracts and records
The Law Society's lifecycle checklist tells firms to assess the liability and insurance cover relating to a tool and to speak to their insurer about whether the use is covered under the existing policy. Do that by email, naming the tool and the task, and file the reply. On the contract side, read the supplier's liability clause and any indemnity. Some vendors' output indemnities cover only certain plan types: OpenAI's Service Terms, for instance, give the output indemnity to API customers and to Enterprise customers (a term that there covers Enterprise, Edu and Healthcare), not to ChatGPT Business, so confirm the position for your plan rather than assume.
For records, the same guidance advises that if a tool does not provide a history of use, you should document inputs, outputs and system errors so use can be monitored. Choose a tool that keeps a history, or add a simple log: date, matter reference, task, tool, who reviewed, outcome. Review the log monthly for the first quarter.
Check 8: exit
Ask what happens to your data if you stop. Can you export it, how long does the supplier keep it after you end the contract, and can you request deletion? Some business workspaces have no self-service export at all: OpenAI's help centre says it is not available in ChatGPT Business, Enterprise or Healthcare workspaces, where data access goes through the workspace owner. Confirm the position for your plan. A firm that cannot leave a tool cleanly has handed it a hold over its own records.
A five-solicitor firm decides on a drafting tool
Take an illustrative five-solicitor firm that wants a business AI assistant for routine client updates and first drafts of letters. The practice manager runs the sheet in an afternoon.
- Purpose. Drafting routine client updates from fee-earner bullet points, and summarising the firm's own template documents. Excluded: authorities, anything involving a barrier, anything for children's or vulnerable clients until reviewed.
- Confidentiality. The firm reads the business plan terms, finds training on content is off by default, and prints the page. Free accounts are barred by the AI policy.
- Data location. The supplier's help page offers a storage-region choice on the business plan. The firm selects it and files a screenshot, and emails the supplier to ask about processing location and gets a written reply.
- Supervision. The fee earner reviews every update; for letters to counterparties, a second solicitor reads it.
- Client care. A paragraph is added to the client care letter. Existing clients are told at their next update.
- Insurance and contract. The broker's reply says the use is acceptable on the stated terms. The liability clause is noted.
- Records. The tool's history is enabled, and the fee earner logs the matter reference for each use.
- Exit. Deletion on request is confirmed; export is limited, so nothing the firm needs lives only in the tool.
The result is a signed one-page form and a go decision for a single, narrow purpose. The tool is not approved for anything else until the sheet is run again for that task.
A go or no-go form you can copy
AI tool pre-use sign-off
Tool and plan: [name, plan] Date: [date] Prepared by: [initials]
Purpose (one line): [task]
Excluded uses: [list]
1 Confidentiality: terms read [Y/N] training off [Y/N] evidence filed [Y/N]
2 Data location: processing [country/region] storage [region] confirmed in writing [Y/N]
3 Supervision: reviewer named per task [Y/N]
4 Client care: wording added [Y/N] opt-out process [Y/N]
5 Insurance: insurer replied [Y/N] date [date]
6 Records: history on or log in place [Y/N]
7 Exit: export and deletion terms read [Y/N]
Decision: GO for this purpose only / NO-GO
Signed: [partner] Review date: [date, within 6 months]A single "N" is not automatically a no-go, but it needs a written reason. The six-month review date matters because terms, models and features change, and a sign-off from last spring may no longer describe the product.
Grading a supplier's answer
Suppliers answer security questions in ways that sound complete and are not. Here are three illustrative replies to the data-location question, with how to read each.
- "We take data security extremely seriously and use industry-standard encryption." This does not answer the question. Ask again, in the same words, and say you will record the reply.
- "Customer content is stored in the region you select in workspace settings. Some processing for abuse monitoring may occur elsewhere; details are in our data processing addendum, section 4." This is a real answer. It gives a mechanism, an exception and a place to verify. File it, read section 4, and decide whether the exception matters for your matters.
- "Data stays in our home region." It may be true, but ask whether it means storage, processing or both, and whether sub-processors are included. A short sentence is a reason to ask, not to relax.
Grade each reply as answered, partly answered or avoided, and put the grade on the sign-off form. Two "avoided" grades on questions about confidentiality or data location are, in my view, enough to stop. A supplier that is serious about law-firm customers will have answered them before.
A two-page policy skeleton
The Law Society's guidance expects firms to follow their wider IT, AI, confidentiality and data governance policies, which in a small firm is easiest with a short document everyone reads. A workable outline, with plain headings:
- Approved tools. A list of named tools and plans, with the approved purpose of each. Nothing else is approved.
- What may go in. Public and firm-owned template material always; client information only in tools marked "client data approved"; never in personal accounts.
- Checking. Every output is read and verified by the fee earner before use; authorities are verified in a trusted source.
- Recording. The log fields, and who reviews them.
- Clients. The client care wording and the opt-out process.
- Problems. How to report an error, a near miss or a suspected disclosure, to whom, and how fast.
- Review. The date of the next review and the owner.
Item 6 is the easiest to leave out. When someone pastes the wrong thing into the wrong tool, the first hour matters, and a person who knows exactly whom to tell and what to write down is worth more than a long clause about acceptable use.
When a check fails
Most failures are fixable without abandoning the idea. If confidentiality terms are unclear, move to a plan with clearer terms, or restrict the purpose to tasks with no client data. If data location is unsatisfactory, choose a different product or region option. If the insurer hesitates, ask what conditions they would accept, such as human review of every output. If supervision cannot be resourced, narrow the purpose until it can. The no-go decision is the right one when a check fails and the task cannot be narrowed to avoid it, and recording that decision is as useful as recording a go. The firm has still done the diligence, and the next review starts from a clear record.
Where the checks usually fail
The tool inside a tool
The firm approves one assistant, and then a transcription feature, an email add-in and a document tool each turn on their own AI. Each has its own terms. Put "AI features in existing software" on the sheet, and ask the IT supplier for a list of what has been switched on.
Personal accounts on the side
A solicitor with a free personal account pastes in a client email because the approved tool is slower. The policy has to say plainly that this is not allowed, and the approved tool has to be good enough that people do not want to. Training matters here; a one-hour team session plan for solicitors covers the content.
A sign-off that nobody reads
The form exists in a folder and no fee earner knows its content. A one-page summary of the rules, in the induction pack and on the intranet, does more than a thirty-page policy.
Supervision on paper only
If the named reviewer is a partner who signs things without reading them, the supervision check is empty. Test it by planting a wrong fact in an AI draft during training and seeing whether the reviewer finds it.
Keeping the approval current
Review each approved tool at six months, and whenever the supplier announces a change to terms, plans or data handling. Keep a one-line change log. Ask fee earners two questions at each review: what has the tool been useful for, and what has it got wrong? The second question produces the better answers. For the regulator's side in more detail, the checklist based on the SRA's guidance is the natural next read.
The Law Society's own page, Generative AI: the essentials, is the primary source for the points above and is updated, so read the current version rather than relying on a summary. If you would like a second opinion on the sign-off form before you adopt it, that is part of what my AI implementation consultation covers, though it does not replace your own regulatory and compliance advice.
Questions solicitors ask before approving an AI tool
Do solicitors have to tell clients they use AI?
The Law Society's guidance says the SRA currently has no specific guidance on generative AI use or disclosure for client care, but advises that firms and clients decide together whether and how AI is used, because clear communication prevents misunderstandings. A short line in the client care letter is the simplest way to do it. Check current guidance before settling your wording.
Who is responsible if the AI gets it wrong?
The solicitor and the firm. Professional obligations apply whether the tool was used personally or by someone under a solicitor's supervision. An AI output is treated like work from a trainee: it needs checking by someone competent before it reaches a client or a court.
Can we use a free consumer chatbot for client work?
Not for anything confidential. Free and personal plans often have different data terms from business plans, and you may have no contract that covers client data. Use a business-grade product on terms you have read, and keep personal accounts out of client matters.
Does our insurer need to know?
Ask them. The Law Society's guidance advises speaking to your insurer about whether use of an AI tool is covered under your existing policy. Get the answer in writing, because assumptions about cover are exactly what claims turn on.
Is a written AI policy compulsory?
The Law Society's guidance does not set out a standalone requirement, but its checklist tells firms to adhere to their wider policies on IT, AI, confidentiality and data governance, which is hard to do without a written AI policy. Two pages is enough if staff read it.
Further reads
- Best AI for Solicitors in the UK: Matched to Each Task — Which tools suit which solicitor tasks, once approved.
- AI Software for Solicitors: Data, Supervision and Supplier Checks — Supplier checks on data, supervision and contracts.
- Which Legal Tasks Should a Small Firm Never Hand to AI? — Tasks that stay with a person whatever the tool.
- SRA Transparency Rules: Use AI to Draft Price and Service Pages — Drafting price and service pages with AI under transparency rules.
- AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day plan that follows a go decision.
- AI Time Capture for Solicitors: Recover Hours You Never Billed — A low-risk first use for AI in a solicitors' firm.
- ChatGPT for Solicitors: Five Safe Tasks and Three to Avoid — Safe starting tasks after sign-off.
- AI Readiness Checklist for Accountants, Solicitors, Consultants — A wider readiness check before the firm adds more AI tools.
- Law Firm AI Usage: How to Audit What Your Fee Earners Really Use — A checklist to find out which AI tools fee earners really use: survey, browser and billing checks, then an approved list.
- Can AI Help Immigration Law Firms Answer an RFE? What to Check — How an immigration firm can use AI on a request for evidence: issue list, evidence matrix, draft response and the checks a lawyer must still make.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Law Society, Generative AI: the essentials (checklist, professional obligations, lifecycle considerations for SMEs) and the SRA Code of Conduct paragraphs it cites; ICO guidance on AI and data protection (archived October 2026); OpenAI Service Terms, ChatGPT business pricing and data-export help pages; Anthropic Claude pricing; Microsoft Learn Copilot privacy page, checked November 2026. Wording on regulator pages changes, so verify the current text.