Data protection and cyber security

Putting client data into AI tools safely

Guardrails for solicitors before client data goes into an AI tool: data classes, supplier checks, staff rules, human review and what to tell clients.

Ava research teamPublished 7 min read

On this page
  1. The risk is already inside the firm
  2. Classify the data before you pick the tool
  3. Supplier checks before any client data goes in
  4. Who approves a new tool
  5. Rules for staff at the keyboard
  6. Human review and accountability
  7. If client data goes into the wrong tool
  8. What to tell clients
  9. Example: summarising a bundle
  10. Frequently asked questions
  11. References

Key takeaways

  • Assume staff already use AI on their own devices; a written rule is safer than silence.
  • Classify the data first: public, anonymised, client personal data and privileged or special category information need different rules.
  • Check the supplier terms on storage, training on your inputs, retention and access before any client data goes in.
  • A named person reviews every AI output before it reaches a client or a court; the firm remains accountable for it.

The risk is already inside the firm

Ask a team whether anyone has pasted an email into an AI tool to tidy the wording, and the honest answer is usually yes. Across the UK, 55% of employees report using AI for work or education, while only around 35% of businesses report using at least one AI technology[1]. The difference is unofficial use, often on personal phones and free accounts.

AI use at work: employees versus employers, UKThe gap is the space where unofficial use of AI tools happens.
AI use at work: employees versus employers, UKComparison. Employees using AI for work or education: 55%; Businesses reporting use of at least one AI technology: 35%.55%Employees using AI for work or educationEmployees using AI for work or education: 55%35%Businesses reporting use of at least one AItechnologyBusinesses reporting use of at least one AI technology: 35%AI use at work: employees versus employers, UKComparison. Employees using AI for work or education: 55%; Businesses reporting use of at least one AI technology: 35%.55%Employees using AI for work oreducationEmployees using AI for work or education: 55%35%Businesses reporting use of at leastone AI technologyBusinesses reporting use of at least one AI technology: 35%

Source: Office for National Statistics, 2026[1]

Show the data
ItemValue
Employees using AI for work or education55%
Businesses reporting use of at least one AI technology35%

The SRA has now said plainly what worries it. Its warning notice on the misuse of AI highlights two concerns: inaccurate output such as fictitious citations, and client confidential information being entered into public AI tools. It adds that both paid-for and free tools may not provide the contractual and technical safeguards needed to keep client information confidential[2].

The Legal Ombudsman makes the same point to consumers: not all free AI tools are secure or keep what users type confidential[3]. And the SRA has been clear for some time that firms remain responsible and accountable for outputs from AI they use[4].

Classify the data before you pick the tool

Most problems come from treating all text the same. A four-level rule is easy to remember and covers nearly every request staff make.

A four-level rule for what may go into an AI tool
Data classTypical contentAllowedConditions
1. Public or firm-onlyMarketing copy, published guidance, internal templates with no client detailsAny approved toolNo client names or matter facts
2. Anonymised client materialA clause or letter with names, addresses, dates and amounts removedApproved business toolsCheck that the remaining facts cannot identify the client
3. Client personal dataNames, contact details, matter facts, correspondenceOnly tools that have passed supplier checksSigned processing terms, no training on inputs, UK or adequate storage
4. Privileged or special categoryHealth, criminal allegations, children, privileged adviceOnly where a partner has approved the specific useRecorded approval and a documented risk assessment

Put the table on one page and pin it beside every desk. When staff are unsure, the rule is to treat the text as one class higher.

Borderline cases come up every week. A precedent clause with the client's name removed is usually class 2. The same clause with an unusual property description and a completion date may still identify the client, which makes it class 3. A note of a phone call about a child arrangements dispute is class 4 even if the names are stripped, because the facts themselves are sensitive. When in doubt, ask before pasting, not after.

Supplier checks before any client data goes in

The Law Society's generative AI primer, aimed particularly at small and medium-sized firms, flags misuse or disclosure of confidential information as a key risk[5]. Turn that into questions your supplier must answer in writing.

  1. Processing terms

    Is there a signed data processing agreement that names the firm as controller and limits the supplier to your instructions?

  2. Training on inputs

    Are your prompts, files or outputs used to train or improve models? The answer for client data must be no, in the contract, not just in a settings menu.

  3. Storage and transfers

    Where is data stored and processed, and what safeguards apply to any transfer outside the UK?

  4. Retention and deletion

    How long are prompts and outputs kept, and can the firm delete them on request?

  5. Access and logging

    Who at the supplier can see your data, and can you see an audit log of who used the tool at the firm?

  6. Security evidence

    What certifications, penetration testing or independent assurance can the supplier show?

Keep the answers with your record of processing and repeat the check whenever the supplier changes its terms.

Who approves a new tool

Unofficial use thrives when there is no quick way to get a tool approved. Give one person, usually the COLP or the data protection lead, the job of saying yes or no within a set time, and publish a short register of approved tools.

  • Tool name, supplier and the account type the firm uses.
  • The data classes it is approved for, such as 'classes 1 and 2 only'.
  • Approved uses, for example drafting first versions of standard letters or summarising documents the firm already holds.
  • Uses that are not allowed, such as legal research without citation checks.
  • Date approved, date of next review and the person who approved it.

Review the register every six months. Suppliers change their terms, add features and move data between regions, so an approval given last year may no longer hold.

Rules for staff at the keyboard

  • Use only tools on the firm's approved list, signed in with a firm account, never a personal one.
  • Never paste class 3 or class 4 data into an unapproved tool, including transcription and grammar apps.
  • Strip names, addresses, dates of birth, reference numbers and amounts before using class 2 material.
  • Never ask an AI tool for case law without checking every citation against a trusted source.
  • Save any output used in a matter to the file, with a note of the tool used.
  • Report a mistaken paste immediately to the data protection lead, the same way you would report a misdirected email.

Give staff safe examples as well as prohibitions. 'Rewrite this paragraph in plain English' with a template paragraph containing no client details is fine. 'Summarise this letter from Mrs Patel's ex-husband' pasted into a free app is not. A handful of worked examples does more than a page of rules.

Human review and accountability

Across UK businesses that use AI, 84% apply at least some human checking to AI outputs and 67% apply significant checking[6]. For a law firm the bar should be higher: every output that reaches a client, another party or a court is read and approved by a named person who understands the matter.

Record who approved it. When something goes wrong, the question will be who checked the output and what they checked it against, not which tool produced it.

A reviewer should check, at minimum:

  • Every fact against the file, especially dates, names, amounts and deadlines.
  • Every legal citation against a trusted source, never against the tool itself.
  • That nothing from another client or matter has crept into the text.
  • That tone and advice suit this client, including any vulnerability noted on the file.
  • That the output says nothing the firm could not stand behind if it were read out in a complaint.

If client data goes into the wrong tool

Treat a mistaken paste like a misdirected email. The steps are the same, and speed matters, because a notifiable personal data breach must reach the ICO without undue delay and, where feasible, within 72 hours of the firm becoming aware of it[7].

  1. The person who made the mistake tells the data protection lead straight away, with the tool, the account and what was pasted.
  2. Where the tool allows it, delete the conversation and any uploaded files, and record that you did.
  3. Check the supplier's terms on retention and training to judge whether the data may persist.
  4. Decide whether the incident is notifiable and whether the client should be told, and record the reasoning.
  5. Use the case, anonymised, in the next training huddle.

What to tell clients

Clients care about being told. When the LSB Public Panel asked legal service users about AI, AI acting without the user's knowledge or consent was the single most urgent concern, named by 56%[8]. A short, honest sentence in your client care information goes a long way.

Sample wording: "We may use secure software, including AI tools, to help draft documents and summarise information. Your information stays within systems we have checked, it is not used to train public tools, and a qualified member of our team reviews any work before it is sent to you or relied on."

If a client objects, record it on the matter and switch the use off for that file.

Put the wording in the client care letter and the privacy notice, not only on the website. If the firm starts using a new kind of tool partway through a matter, such as transcription of meetings, tell the client before the first use rather than after.

Example: summarising a bundle

Example: a fictional associate wants to summarise 300 pages of disclosure in an employment matter. The documents include health information, which makes them class 4. She asks the supervising partner, who approves the firm's vetted document tool for this matter only, records the approval and the reason, and asks for the summary to be checked against the key documents before use.

The associate saves the summary to the file with a note of the tool and her review. The client care letter already explains the firm's use of such tools, so nothing further is needed. Total extra admin: about five minutes.

Pair these rules with your BYOD policy and cover them in your staff training, because most accidental pastes happen on phones.

Frequently asked questions

Can solicitors use free AI tools at all?

For public or firm-only material with no client details, many firms allow it. For client data, free consumer tools rarely offer the processing terms, retention controls and confidentiality a law firm needs.

Do we need client consent to use AI?

Not always as a legal requirement, but clients expect to be told. Explain it in your client care information and honour objections on individual matters.

Who is responsible if an AI tool makes a mistake?

The firm. The SRA is clear that firms remain accountable for outputs from AI they use, including errors from third-party tools.

Is anonymising data enough?

Only if the remaining facts cannot identify the client. In a small town, an unusual set of facts can identify someone even without a name.

References

  1. Office for National Statistics (2026). Artificial intelligence in UK businesses: 2023 to 2026 (opens in a new tab) ↑

  2. Solicitors Regulation Authority (2026). Misuse of AI - Warning notice (opens in a new tab) ↑

  3. Legal Ombudsman (2026). AI and Complaints: A Consumer Guide (opens in a new tab) ↑

  4. Solicitors Regulation Authority (2023). Risk Outlook report: The use of artificial intelligence in the legal market (opens in a new tab) ↑

  5. The Law Society (2026). Generative AI - the essentials (opens in a new tab) ↑

  6. Department for Science, Innovation and Technology (GOV.UK) (2026). Artificial Intelligence Adoption Research (opens in a new tab) ↑

  7. legislation.gov.uk (The National Archives) (2016). UK GDPR Article 33: Notification of a personal data breach to the supervisory authority (opens in a new tab) ↑

  8. Legal Services Board (2026). AI in Legal Services: Research Report (LSB Public Panel) (opens in a new tab) ↑

Book a demo

See Ava handle your calls.

Book a short demo. We’ll walk through your calls and workflow, and agree what Ava should take care of.

Book a demo