A buyer’s guide to legal AI security
The data privacy and security questions to ask a legal AI provider before signing the dotted line.
Legal work runs on information clients expect you to protect, from contracts, to confidential advice, to financial records and the details of disputes that may never become public.
AI can help lawyers work through that material faster. But a prompt or document may pass through an application, a cloud provider, a model provider and connected systems before an answer comes back.
Security claims can sound reassuring. “Enterprise grade”, “encrypted” and “private” each describe part of the picture, yet they don't explain what happens to a client document when a lawyer uploads it.
This guide looks at the controls that matter most when choosing legal AI, the evidence to request and the questions to take into a vendor conversation.
Start with the journey of a client document
Ask a prospective provider to walk you through an ordinary task: a lawyer uploads a confidential agreement, asks for a summary, refines it and saves the result. Which systems receive the document and the prompts? Which organisations operate those systems? What is stored, and for how long?
Then repeat the exercise for a connected workflow. If the AI can search your document management system, use a model supplied by another company or send a draft to another application, the data journey has more steps. Each one needs a clear purpose and an appropriate control.
1. What happens to our data?
Find out exactly what the provider does with prompts, uploaded files, generated answers and feedback. Ask whether any of these are used to train or improve models, whether people can review them, and whether the same terms apply to every model and feature offered in the product.
“Your data is not used to train the model” is an important answer. It is not the whole answer. Information might still be retained in logs, backups or connected applications. Buyers need to know the purpose of that retention, who can access it and when it is deleted.
Ask what happens when a user removes a document, leaves the organisation or the contract ends. A useful answer covers the live service, copies and backups, with any exceptions explained in the agreement.
2. Who can access a matter?
A secure platform must respect the boundaries inside the firm. A lawyer working on one matter should not gain access to another client’s documents because an AI search can reach across a shared knowledge store.
Check how the product handles single sign-on, multifactor authentication, user roles and matter permissions. For document management integrations, ask whether the AI inherits the permissions in the source system, how quickly permission changes take effect and whether administrators can limit access to particular repositories.
Test this with two realistic users during a pilot. Give them different access to the same source system, then check what each can retrieve through the AI. This is more revealing than a demonstration using an administrator account.
New AI workflows. Built for legal standards.
3. Where does the information go?
Ask where information is processed and stored, which cloud and model providers are involved, and whether authorised support staff can access it from other countries. Data location and access arrangements should be clear for the particular product and contract you are buying.
Where personal information is involved, have your privacy team assess any international transfers and the relevant contractual arrangements.
The ICO’s guidance on international transfers explains that access from another country can matter as well as physical storage. The buyer should also understand which organisations act as processors or sub-processors and how changes to that list are communicated.
4. How is the platform tested and monitored?
Encryption in transit and at rest is a sensible starting point. Ask how the provider controls privileged access, detects suspicious activity, tests the service and fixes vulnerabilities. Request current, relevant assurance documents and check which products and systems they cover.
AI also creates risks that a conventional security questionnaire may miss. A malicious instruction hidden inside a document or website could try to redirect an AI workflow. The NCSC’s secure AI guidance calls for security across design, development, deployment and ongoing operation.
Ask how the vendor tests AI-specific threats, limits what a workflow can do, and reviews changes to models and integrations.
No provider can promise that every attack will fail. A credible answer explains the controls, their limits and how the team responds when testing uncovers a weakness.
5. What can the AI do on our behalf?
This becomes more pressing as AI moves from answering questions to completing tasks. A workflow that reads a contract has a different risk profile from one that can edit a document, retrieve a whole matter file or send information to an external system.
Ask which actions are possible, what the default permissions allow and where a person must approve the next step. Can administrators restrict particular connectors or actions? Are there audit logs showing what the workflow accessed and did? Could a user or an untrusted document cause the system to retrieve or disclose more than intended?
Start a pilot with a narrow use case and limited access. Expand it when the team can see how the controls work in practice.
6. What happens if something goes wrong?
Every provider should have a process for investigating security incidents. Ask how quickly they will tell you about an incident affecting your information, what details they will share and how they will help you establish which users, documents and systems were involved.
Check the contractual terms for notification, cooperation, service recovery and deletion or return of data.
The ICO’s AI procurement framework highlights the need to define security measures and assistance with personal data breaches in processor agreements.
For a law firm, the practical question is whether it could establish the impact on clients and matters in time to take the necessary steps.
Beyond the security checklist
A product can protect access to data and still produce an inaccurate answer. Legal buyers should assess both. The SRA has warned about confidential information being entered into AI tools without appropriate safeguards, alongside the risk of false or incorrect legal material being relied upon.
Ask how the product presents sources, lets a lawyer check an answer and makes clear when human review is needed. Security controls should give people a safe way to use the tool; training and supervision should help them use it well.
How LexisNexis approaches security
LexisNexis says customer data is not used to train public AI models or improve AI performance for other customers in Lexis+ with Protégé integrations. Its Trust Center describes encryption in transit and at rest, layered security controls, and logging and detection systems. These are useful starting points for a buyer’s review; your security team should confirm the scope and terms that apply to the UK product and the workflows you intend to use.
For teams evaluating Lexis with Protégé, the next conversation should be specific: show us the data flow for our use case, the permissions on our connected documents, the applicable model and sub-processor arrangements, and the evidence behind the controls. Explore Lexis with Protégé.
Questions legal AI buyers should ask
Before selecting a provider, ask for answers you can share with your security, privacy and legal teams:
Will our prompts, documents or outputs be used to train or improve AI models?
Check the answer for every model, feature and third-party provider in the service, including any optional settings.
Who can see our information, and how long is it kept?
Include vendor staff, model providers, logs and backups. Ask how deletion works at the end of a matter or contract.
Will the AI respect our existing access controls?
Test matter-level permissions, user roles and connected repositories using ordinary accounts during a pilot.
Where is our information processed, and who are the sub-processors?
Request a current data-flow map, the relevant contractual terms and a process for changes to providers or locations.
What evidence supports the provider’s security claims?
Review relevant independent assurance, testing, incident processes and the scope of any certification or audit.
What actions can an AI workflow take, and what can we control?
Look at connector permissions, approval steps, activity logs and safeguards against malicious instructions in source material.
How will these controls hold up as our use grows?
Ask what changes when more teams, models, repositories and automated workflows are added, and how the provider will tell you about material changes.
Confidence comes from clear answers
The best security conversation starts with a real piece of work. Trace where its information goes, check who can access it, test what the AI can do and ask to see the evidence.
Legal teams should be able to put AI to work without guessing what happens to confidential information behind the interface. A provider that can explain the journey clearly makes it easier for buyers to assess the risk, agree sensible controls and use the technology with confidence.
Workflows that draft
Create outstanding drafts with workflows built on your documents and LexisNexis sources.
Pre-built & custom workflows
Move faster on everyday tasks and complex decisions with workflows that fit how you work.
Trusted content
Strengthen consistency with workflows that combine organisation knowledge and exclusive sources.
2026 research & thought leadership reports
In CTO we trust
Chief Technology Officers at leading law firms face a new challenge: not just deploying AI, but making it trustworthy at scale.
The Mentorship Gap
AI can help new lawyers to search, summarise, draft, and reason. But can it teach?
Integrating AI into legal workflows
Anything-goes AI use won't raise your standards. Legal teams need to restructure their workflows based on suitable AI.
Death of the rainmaker
Love them or loathe them, rainmakers are a law firm's bread and butter. But will AI-powered legal solutions see the practice take precedence over the partner?