Questions to ask an AI vendor before a law firm signs

ABA Formal Opinion 512 treats confidentiality as a fact-specific risk analysis under Rule 1.6. Before a lawyer inputs information relating to a representation, the firm has to understand the vendor's terms: who can access what is typed, whether the tool trains on it, and what happens when the engagement ends.

California attorneys should also read the California State Bar Practical Guidance on the Use of Generative Artificial Intelligence in the Practice of Law (November 2023) next to Opinion 512. The Practical Guidance emphasizes confidentiality themes for generative tools in California practice. It is State Bar ethics guidance—not a Bar endorsement of any vendor, course, or questionnaire. A national checklist is a starting point; California counsel still adapts it to the firm's matters and engagement letters.

Before a pilot has client files

Run diligence before the pilot, not after the first client memo is pasted in. A pilot that already contains matter data is an operational decision under Rule 1.6 themes, not a harmless experiment. Put answers in a written questionnaire attached to procurement. If the vendor will only answer in a slide deck, treat that as incomplete for confidentiality purposes.

These questions also feed the "approved tool" bucket in the firm AI use policy. Do not approve a product in the policy until training, access, retention, breach notice, indemnity, deletion, and cross-matter mixing are answered in writing. See what a firm AI policy should include.

Contract vs dashboard toggle

A dashboard toggle that "turns off training" is not the same as a contract the firm can enforce. Settings can reset when seats are added, when a workspace is cloned, or when the product ships a new default. Prefer language in the order form, DPA, or MSA that states whether prompts, outputs, and uploads are used to train models; who may change that setting; and how the firm can verify it. Ask the same question about subprocessors. If the only assurance is a help-center article, write that gap into the risk file before client information goes in.

Printable questionnaire: ask before you sign

Copy these seven questions into your diligence memo. Require written answers. Keep the packet with the contract file and with the firm AI policy's approved-tool appendix.

1. Does the vendor train models on our prompts, outputs, or uploaded files? Can that setting be turned off, in the contract, not only in a dashboard?
2. Who at the vendor, and which subprocessors, can access a prompt that includes client information?
3. Where is the data stored, and for how long is it retained after we delete a matter or end the subscription?
4. Will the vendor notify us of a breach or of a legal demand for production of our inputs?
5. What does the indemnity actually cover if client information is disclosed, or if the tool's output is wrong and we relied on it?
6. Can we export or delete a matter's inputs, and do we get a confirmation that they are gone?
7. Does the tool's output get mixed with another client's matter, inside the firm or outside it?

Those questions belong in a written questionnaire, not in a sales demo. After answers arrive, map them into the policy: which tools are approved, which paste destinations are banned, and who may authorize an exception. Vendor-neutral diligence is the point—this page does not rank products or claim superiority over any CLE or technology provider.

Informed consent and Rule 1.6 themes

Opinion 512 frames confidentiality under Rule 1.6 before a lawyer inputs information relating to a representation. Informed consent, when it is required, is not a one-line boilerplate in an engagement letter. The Practical Guidance's confidentiality themes point the same direction for California practices: understand where matter information goes, who can see it, and what remains after the matter ends.

Vendor terms are part of that analysis. So is the supervision structure in the firm AI use policy. If associates can open a consumer chatbot with client facts because "the firm has not decided yet," the firm has already made a decision—just not a written one. For CLE that maps Opinion 512 and the Practical Guidance, see California MCLE that covers ABA Formal Opinion 512.

FAQ

Questions to ask an AI vendor before a law firm signs

Before a law firm signs an AI contract, ask whether the vendor trains on your inputs, who can access prompts, where data is stored, what indemnity covers, and whether a matter's inputs can be deleted. Put answers in the contract, not only in a dashboard toggle.

Do I need vendor diligence questions before a pilot that has client files?

Yes. Ask before a pilot already has client files in it. ABA Formal Opinion 512 treats confidentiality as a fact-specific Rule 1.6 analysis. The California State Bar Practical Guidance (November 2023) likewise emphasizes confidentiality themes for generative tools. A sales demo is not a diligence file.

Is a dashboard toggle enough to turn off training on firm prompts?

No. Prefer contract language the firm can enforce—not only a UI setting that can change, reset, or differ by workspace. Ask whether training-off is contractual, how it is audited, and what happens if a subprocessor still trains.

How do informed consent and Rule 1.6 themes relate to AI vendor terms?

Opinion 512 frames confidentiality under Rule 1.6 before a lawyer inputs information relating to a representation. Informed consent, when required, is not a one-line boilerplate. Vendor terms on access, training, retention, and deletion are part of that risk analysis—alongside the firm AI policy.

Related: firm AI use policy, hallucinated citation already filed, CA MCLE & Opinion 512.

The on-demand California MCLE course is $249. It includes the five modules, the 111-page reading, and the 35-question assessment. California MCLE accreditation is pending. The certificate is issued when the provider number is assigned. The course does not guarantee compliance, prevent sanctions, or satisfy a court, insurer, or client.

Enroll Now — $249

Course detail.