Vendor due diligence for legal AI: a checklist
By the ChamberOne team 5 minute read
Buying a legal AI product is a different exercise from buying a diary app. Client confidences will pass through it, and some of that text will leave the vendor's own systems and travel to a model provider. Managing partners and IT leads often have a list of questions that were written for ordinary software and miss the points that matter here. This checklist is built for legal AI. It is organised by topic, and each topic notes the kind of answer you should expect and the kind that should make you pause.
Key points
- Write down what client information you will put into the tool before you read any vendor reply.
- Ask for assurance documents with their scope and dates, not logos.
- Insist on named model providers and sub-processors, a clear data flow and the ability to switch AI features off.
- Check logs, access controls, incident notice and a working data export before signing.
Before the questions: know what you are protecting
Start with your own position. What kinds of client information will go into the tool? Are there client or regulator rules that restrict outside processing? Which matters, if any, must stay out? The answers set how strict you need to be. A tool used only for internal marketing drafts needs less scrutiny than one that reads client contracts.
Write the answers down. They become the benchmark against which every vendor reply is read, and they help when a client asks you the same questions later.
Security assurance
Ask what independent assurance the vendor holds. Common answers refer to a security certification or an independent audit report. Ask for the document itself, the date it covers, its scope and any exceptions noted. A certificate that covers the vendor's office but not the product does not help you.
Do not accept a logo on a web page as an answer. Equally, do not assume a young company without a certificate is careless. Some are in the process. If so, ask what has been done, what is planned and by when, and decide whether you can live with the gap.
- What assurance do you hold, for which systems, and for what period?
- Can we see the report under a confidentiality agreement?
- How is data encrypted in transit and at rest?
- How often is the product tested by outside specialists, and are findings fixed on a schedule?
- How do staff get access to customer data, and is that access recorded?
Model providers and sub-processors
Most legal AI products pass text to a third-party model provider. Ask who that is, which models are used, what is sent, where it is processed and what the vendor's contract with the provider says about retention and training. The same applies to hosting companies and any other sub-processors.
A good answer names the providers, describes the data flow in plain words and offers a list that is kept up to date, with notice of changes. A weak answer says 'industry-leading partners' and nothing more. Also ask whether you can choose a model, switch AI features off, or restrict them to particular matters. Control of that kind is worth more than reassurance.
Where the data lives and who can reach it
Find out where the product is hosted and where the model provider processes text, because these may be different countries. Ask how the vendor responds when a government or court demands customer data, and whether it would tell you. Ask about backups and how long deleted data survives in them.
Check these answers against your own obligations. The guidance from your regulator or bar, your data protection law and your clients' instructions may all bear on where data can go. If you are unsure, take advice before you sign.
Logs, access and review
A firm that uses AI on client work should be able to show what was done. Ask whether the product records who used an AI feature, when, on which matter and with which model. Ask how long records are kept, whether they can be exported and whether a user can switch logging off.
Ask about access control too. Can a sensitive matter be restricted to its own team? Does the AI feature respect that restriction, or can it read everything? Does the product support two-step sign-in and a role for read-only reviewers?
Incidents, exit and the contract
Ask how and when the vendor will tell you about a security incident, who your contact will be, and what help you can expect. Look at the contract for notice periods, liability limits and what happens to your data if the vendor is sold or closes.
Exit rights deserve their own attention. Can you export all your data, in a usable format, at any time and not only on request at the end? How soon after termination is your data deleted, and will the vendor confirm it in writing? The ability to leave keeps every other promise honest.
Ask also about the practical side of leaving: how long an export takes for a firm of your size, whether documents come out with their folder structure and metadata, and whether the vendor will help at a reasonable rate if the export needs support. Try a test export during the trial, so that you learn what you would receive before you need it.
Reading the answers, and keeping the record
Score each topic as satisfied, partly satisfied or open, and note who will follow up. Treat an unwillingness to answer in writing as an answer. Be wary of replies that restate your question as a statement without any detail.
Test the claims you can. Ask for a trial with non-client data, try the export function, and look for the AI features in the settings. Keep the questionnaire, the replies and your notes in one place. Review them each year, and whenever the vendor changes a model provider or a hosting arrangement.
Where a claim matters to your decision, ask for it to be written into the contract or an order form, because marketing pages change without notice and contracts do not.
Involve the people who will use the tool. A fee earner can tell you in ten minutes whether the access settings match how your teams actually work, and an IT lead can tell you whether the sign-in options fit your setup.
- Score each topic as satisfied, partly satisfied or open
- Test the claims you can, such as the export function
- Ask for important claims to be written into the contract
- Involve the people who will use the tool
- Treat an unwillingness to answer in writing as no answer
- Rely on marketing pages that can change
What ChamberOne publishes on security and AI
- Lets you choose the model for each AI feature, GPT-5 or Claude, or switch AI features off for the firm.
- Records every read and write in an audit trail kept for 7 years by default, including every AI call.
- Lets you mark a matter or client as Restricted so only its team, the Owners and people you name can see it.
- Offers two-step verification with an authenticator app, which can be required for the whole firm.
What it does not do
- When AI features are on, the text needed for a request is sent to the AI provider (OpenAI or Anthropic), outside the country where ChamberOne is hosted.
- SAML single sign-on is on the roadmap and is not available today.
- Tamper-evident hashing of the audit trail is on the roadmap and is not in the product today.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
Sri Lanka's Anti-Corruption Act 2023: a guide for practitioners
What the Anti-Corruption Act No. 9 of 2023 provides, how CIABOC exercises its powers, and what advisers must understand about offences, declarations, and a proposed 2026 amendment.
6 minute readClient confidentiality and AI tools: a checklist before you paste anything in
The questions to answer about retention, training use, access, hosting and sub-processors before client information goes into any AI tool.
5 minute readShadow AI in law firms: staff are already using it, so what next?
Why bans fail, how to find out what is being used, and how to offer an approved alternative with clear rules.
5 minute read