AI Data Security for Law Firms: Keep Client Data Safe

Lawyers have good reasons to use AI. Summarising a bundle, drafting a first-pass letter or turning meeting notes into a chronology can save hours. The problem is what goes into the prompt to make that work: client names, facts of a dispute, deal terms, medical records, witness statements.

For most businesses, a leak into an AI tool is a data protection problem. For a law firm, it can also be a breach of professional duty and, in some cases, a loss of privilege the client can never get back.

This guide covers what regulators in the US and UK have said, where the real risks sit, and the practical controls a firm of any size can put in place.

This article is general information, not legal advice. Check the rules that apply in your jurisdiction.

What regulators are saying

UK: the SRA's "Misuse of AI" warning notice

In August 2026 the Solicitors Regulation Authority issued a warning notice on the misuse of AI. According to the Law Society Gazette, the SRA said it is "seeing instances of confidential client information being entered into public AI tools such as ChatGPT that lack appropriate safeguards". It warned this will likely breach client confidentiality, and that legal professional privilege "may be permanently waived and unable to be recovered".

The regulator received 42 reports relating to potential AI misuse between July 2025 and July 2026, as Legal Cheek reported. Confidentiality sits alongside fake citations and supervision failures among the issues.

The notice is not an AI ban. Legal Cheek reports the SRA's position that client information should only be entered into AI systems where appropriate contractual, technical and organisational safeguards are in place, and that firms must have effective governance, systems and controls to manage the risks. Even for in-house or enterprise tools, the Gazette reports the SRA expects firms to satisfy themselves that client data stays in a secure environment, is not used to train models unless explicitly authorised, and is not kept longer than necessary.

US: ABA Formal Opinion 512

In the US, the American Bar Association issued Formal Opinion 512 on July 29, 2024, its first ethics guidance on generative AI. It points lawyers to the existing Model Rules on competence (1.1), communication (1.4), fees (1.5) and confidentiality (1.6).

On confidentiality, the opinion focuses on "self-learning" tools that can retain and reuse what you type. As an ABA Litigation Section article explains, the opinion concludes lawyers should get the client's informed consent before putting information relating to the representation into such a tool, even inside a firm's own closed system. That consent "cannot be accomplished by a boilerplate acknowledgment or notice clause in an engagement letter". The same article notes that a self-learning system can surface one client's information in another client's matter without either lawyer realising.

The message on both sides of the Atlantic is the same: AI is allowed, but the firm has to know what data goes where, and be able to prove it.

Where the risk actually sits

Most firms already have a document management system, an email gateway and some form of DLP. None of them see a prompt typed into a browser tab. That gap shows up in a few predictable ways:

• Personal accounts. A fee earner uses a free ChatGPT or Gemini account on a work laptop because it's quicker than waiting for the approved tool. This is the classic shadow AI pattern.
• Copy and paste of whole documents. Contracts, pleadings and expert reports are long, so people paste the full text, including names, addresses and financial details.
• Support staff and juniors. Paralegals, trainees and secretaries often do the summarising. The SRA notice makes clear that supervisors carry responsibility too.
• Mixed client data. Prompts built from previous matters can carry one client's information into another client's work.
• Approved tools with the wrong settings. An enterprise licence helps, but only if staff actually use it and the data handling terms match what you've told clients.

A practical control framework

You don't need a large security team to get this under control. The steps below work for a 20-person practice and scale to a few hundred.

1. Decide which tools are approved

Pick the AI tools the firm will support, and check their contractual terms on data retention, training and location. Write down which tools are approved for client data and which are for non-client work only. Our AI acceptable use policy template is a good starting point.

2. Classify what must never leave the firm

Define the data types that cannot go into any AI tool, or only into approved ones. For law firms that usually means client identities, matter details, privileged communications, special category data (health, criminal records), financial account details and anything under a court order or NDA.

3. Get client consent right

For US firms, Opinion 512 makes informed consent a real conversation, not a clause. For UK firms, check client engagement terms and any client-imposed outside counsel guidelines, which increasingly restrict AI use. Record what each client has agreed to.

4. Enforce the rules where the data leaves

A policy that relies on memory will fail on a busy afternoon. Put a control in the browser that checks prompts before they are sent, warns on borderline content, masks identifiers where the task still works without them, and blocks the rest.

5. Keep an audit trail

If a client or regulator asks "did any of our data go into an AI tool?", you need an answer. Log which tool was used, by whom, what type of data was detected and what action was taken.

6. Train and supervise

Short, specific training beats a long policy document. Show people real examples from your own practice areas, and make supervisors responsible for how their teams use AI.

Quick checklist for law firms

Control	Done?
Approved AI tools listed, with data terms reviewed	☐
Personal AI accounts addressed in policy	☐
Client and privileged data defined as restricted	☐
Client consent and outside counsel guidelines checked	☐
Prompts inspected before they reach AI tools	☐
Audit log of AI use available for clients and regulators	☐
Supervisors briefed on their responsibilities	☐
Training delivered with practice-specific examples	☐

If you're in the UK or EU, read this alongside our guide on whether using ChatGPT at work is GDPR compliant, because client data is usually personal data too.

Where Nura helps

Nura (AI Guard by Nura) is a browser extension that inspects prompts before they reach ChatGPT, Claude, Gemini, Copilot, Perplexity and DeepSeek. It detects sensitive data across 47 categories out of the box, including customer PII, financial data and internal strategy, and you can add custom categories for things like matter numbers or client names using patterns, regex or LLM classification.

For each data type, team and AI tool you choose whether to Allow, Warn, Mask or Block. So a litigation team can use an approved tool with names masked, while personal AI accounts are blocked for client data altogether. Checks run inline in under 50ms with no proxy.

Every event goes into a full audit log showing the user, AI tool, action and policy reason, which you can export or send to your SIEM. Nura supports SSO and SCIM, and runs in Chrome, Edge, Brave, Arc, Firefox and Safari on Mac.

If your firm works with an IT provider, Nura is also available through MSPs and resellers, who can manage it for you from a single console.