AI & Data Privacy
Data Security in LegalTech: What to Ask Before Adopting Any AI Tool
A vendor evaluation checklist written for the buyer, not the seller. If a vendor cannot answer these quickly and in writing, that is the answer.
Most legaltech security conversations happen in the wrong order: pilot first, security review during procurement, and the awkward questions after case material has already been uploaded. This is the list to send before the pilot.
It is written to be forwarded. Send it to the vendor, ask for written answers, and treat the speed and specificity of the reply as data in itself — a vendor who has been asked before will answer in a day.
Key takeaways
- Ask for answers in writing. A statement in a sales call is not a contractual term.
- The three that eliminate most vendors: training on your content, retention after deletion, and whether a processing agreement exists at all.
- Exit matters as much as entry. Find out what happens to your data when you stop paying.
1. What happens to the content itself
- Is customer content used to train, fine-tune, or evaluate any model? State it as a contractual term, not a help-centre article.
- If the product calls a third-party model provider, what are that provider's retention and training terms, and under which tier is the account held?
- Is content ever reviewed by a human — for abuse monitoring, quality, or support — and under what controls?
- Is anything derived from our content (embeddings, indexes, caches) retained separately from the content, and for how long?
2. Encryption and key custody
- Encrypted in transit and at rest — with which algorithms?
- Who holds the encryption keys, and is key management operated separately from the storage layer? If the same party holds both, encryption at rest protects you against a stolen disk and not much else.
- Are keys rotated, and what is the process if a key is compromised?
3. Tenancy and access
- Is each customer's data logically or physically segregated?
- Which vendor staff can access customer content, under what authorisation, and is that access exceptional or routine?
- How is access to our workspace controlled on our side — roles, and whether a matter can be walled off from part of the team?
4. Location and sub-processors
- In which country or region is data stored and processed? Name the region, not the cloud provider.
- Give the full list of sub-processors that can touch customer content, and how we are notified when it changes.
- If personal data leaves the EEA, which Chapter V transfer mechanism applies?
- For Bahraini matters, does the destination appear on the Personal Data Protection Authority's adequacy list under Article 12, or is a specific authorisation in place?
5. Retention, deletion and exit
- How long is content retained by default, and is that configurable?
- When we delete a file, what is actually deleted, and how long does it persist in backups?
- On termination: can we export everything in a usable format, and how quickly is our data destroyed afterwards?
- Is a certificate of deletion available on request?
6. Incidents
- What is the contractual notification window for a personal data breach? The GDPR gives a controller 72 hours to notify a supervisory authority — a vendor who is not obliged to tell you promptly makes that deadline impossible.
- Who is notified, through which channel, and what information comes with the notice?
- Has there been a reportable incident in the last 24 months, and what changed afterwards?
7. Paper
- Is there a data processing agreement meeting Article 28 requirements, and can we see it before the pilot?
- Which independent assessments exist — and, more usefully, what was the scope? A certification covering a corporate network says nothing about the product.
- Does the contract permit us to disclose the arrangement to our client, and to a tribunal or regulator if asked?
- Does professional indemnity or cyber cover extend to this processing?
Reading the answers
Three response patterns are worth noticing. Deflection to a certification — a badge answers a different question than the one you asked. Answers that live only in a webpage — help-centre text can be edited; a contract cannot be edited unilaterally. Speed — a vendor selling to regulated buyers has answered this list many times and will send it back quickly.
One closing point: the answers determine which obligations you can actually meet. What those obligations are, across the GCC and the GDPR, is set out in why pasting case files into ChatGPT could violate client confidentiality.
This article is general information about professional practice and technology, not legal advice. Data protection and professional conduct rules differ by jurisdiction and change often — the position described here is as at August 2026. Check the rules that bind you, and your firm's own policies, before changing how you handle client material.
Keep reading
AI & Data Privacy
Why Pasting Case Files Into ChatGPT Could Violate Client Confidentiality
A general-purpose chatbot is a third party. The moment a client document goes into one, you have made a disclosure decision — whether or not you thought of it that way.
AI & Data Privacy
Attorney-Client Privilege and AI Tools: What Every Legal Professional Should Know
Most of the anxiety about AI and privilege is aimed at the wrong target. In this region the binding constraint usually is not privilege at all — and where privilege does apply, the risk is waiver by disclosure.