Three types of legal AI tools
A legal AI tool differs not by its name but by the data it answers from and by where it processes your requests. By these two features, the market splits into three types.
1. Public chatbot: ChatGPT and similar general services. It writes text fast but answers from the model’s general knowledge, and it stores prompts with the vendor under the vendor’s rules.
2. Legal research service: legal AI on top of a database of law and case law, for example Lexis+ AI or Westlaw AI-Assisted Research in the USA. It gives links to sources, but the firm’s documents go to the vendor’s servers.
3. Private model: an open language model on the firm’s server or in a cloud in Russia, connected to the firm’s templates, positions and archive. Prompts do not leave the firm’s perimeter, and answers are built on your documents.
According to the Clio Legal Trends report 2025, 79% of lawyers use AI, and 53% say their firm has no AI policy or they do not know of one. So the lawyer, not the firm, more often chooses the type of tool, and chooses by convenience, not by the data processing terms.
What to check in the service terms
Before you upload the first client document, read the service terms and the privacy policy. The answers to five questions decide whether you can use it for client matters.
1. Where prompts and files are physically stored, and for how long.
2. Whether the vendor uses your prompts to train the model, and whether you can turn this off.
3. Who at the vendor sees the conversations, and when the vendor discloses them to third parties and courts.
4. Whether there is a data processing agreement for personal data.
5. Whether a cross-border transfer is needed: if the server is abroad, since 1 March 2023 the operator must notify Roskomnadzor before the transfer, under article 12 part 3 of 152-FZ.
OpenAI CEO Sam Altman himself answered the third question on 25 July 2025: conversations with ChatGPT have no legal protection like a conversation with a lawyer or doctor, and the company would have to disclose them if a court demands it.
In February 2026 the court in the Heppner case relied on the policy of a public service under which prompts and answers are collected and may be disclosed to third parties. The court did not accept that the defendant’s conversations with such a service were protected by attorney-client privilege.
Why AI invents case law
A language model does not look up a decision in a database; it predicts plausible text. If the case is not in its data, it may write citation details that look real, although no such ruling exists. This is called a hallucination.
In May 2024 Stanford RegLab and HAI tested legal services on case law queries: Lexis+ AI and Ask Practical Law AI erred in more than 17% of answers, and Westlaw AI-Assisted Research in more than 34%. In an earlier study, general chatbots erred on 58–82% of legal questions.
The lesson for practice: access to a legal database lowers the share of errors but does not remove them. Only a check built into the process works, where every citation is compared with the text of the ruling before the document goes to court or to the client.
The lawyer who signed the document answers for each citation in it, whoever found it. So when you choose an AI tool, look not at how confidently it answers but at how it shows the source.
When you need your own model
You need your own model when the work involves client confidentiality, personal data or counterparties’ trade secrets. For most practices this is an ordinary working day, not an exception.
1. Documents contain personal data, and the firm does not want a cross-border transfer that requires notifying Roskomnadzor.
2. Clients require that their documents are not passed to third parties without consent.
3. You need answers from the firm’s archive: positions, templates, opinions, not from the model’s general knowledge.
4. Work must not depend on whether a foreign service is available from Russia tomorrow.
Your own model is installed on the firm’s server or in a cloud in Russia. Personal data is processed under 152-FZ, the AI has a separate account with rights only to its own process, and every action is logged.
A model inside the perimeter can be replaced without redoing the rules and connections, so the firm is not tied to one vendor.
A 12-point checklist
Go through the list before you pay for a subscription or sign a rollout contract. Count any point without a clear answer as a no.
1. Server: in Russia or abroad, at your firm or at the vendor.
2. A data processing agreement for personal data.
3. Model training on your prompts: whether it is on and how to turn it off.
4. How long prompts are stored and how they are deleted.
5. Disclosure to third parties: when the vendor passes on conversations.
6. A source for every citation in an answer.
7. A flag on a citation that is not in the database, instead of a confident answer.
8. Work with the firm’s archive, not only with the general legal database.
9. Roles and rights: which lawyers see which matters.
10. Action log: who asked what, and when.
11. Replacing the model without redoing templates and connections.
12. A test on your closed matters before payment.
A detailed review of each criterion, questions for the vendor and the test procedure are on our page on choosing AI for lawyers.