Does Your AI Tool Train on Client Data? The Vendor Question Small Firms Skip

You added an AI tool to your practice. Maybe it drafts documents. Maybe it transcribes client meetings. Did you ask what happens to the client data you feed it?

Most firms didn't. And that question just got teeth.

Why This Question Suddenly Matters

The ABA's Formal Opinion 512 set the rule: before you input confidential client information into a self-learning generative AI tool, you need informed consent. Not a line buried in your engagement letter. Actual, explained consent about the actual risk.

Most firms never gave clients that explanation. They signed up for a tool, clicked accept, and moved on.

California is closing that gap. A proposed March 2026 amendment to Rule 1.6 would expand the definition of "reveal" to explicitly cover AI use. Vendor data handling stops being a best practice you can skip. It becomes a direct confidentiality-rule question. New York has already moved in the same direction: its Formal Opinion 2025-6 applies the same scrutiny to AI meeting-recording and transcription tools — the kind firms often adopt through Zoom or Teams without treating it as an "AI decision" at all.

Here's the scale of the problem. The 8am 2026 Legal Industry Report, which surveyed more than 1,300 legal professionals, found that 43% of firms have no formal AI policy and no plans to write one. Only 9% have a written policy that's actually enforced. LawNext's coverage of the same report found firm-level adoption lagging far behind what individual lawyers are doing on their own.

Put those together: most firms that adopted AI tools over the past two years never asked the vendor a single question about where client data goes. That's not a someday problem. It's a live exposure sitting in your Q4 renewal stack right now.

What "Training on Your Data" Actually Means

Not every AI tool is the same risk. Opinion 512 draws a line between self-learning tools, which use your inputs to improve their models, and non-learning tools, which don't retain or reuse what you feed them. That distinction decides whether you have a consent problem at all.

Confidentiality doesn't leak in one obvious place. It leaks in three:

• Prompts. Whatever you type into a chat tool, someone — or something — can potentially see.
• Uploaded documents. Contracts, pleadings, and client files pushed into a drafting or research tool.
• Retained logs. Session histories and transcripts many tools keep by default, sometimes indefinitely.

And here's the trap most firms fall into: vendors advertise "opt-out available" as if it's the same as "off." It isn't. If training is on by default and opting out takes a support ticket you never file, your data is training the model right now.

The Five Questions to Put to Every AI Vendor

Before you renew, or before you sign anything new, get these answered in writing:

1. Does our data train your model, and can we opt out in writing? Verbal assurances from a sales rep don't count.
2. Where is our data stored, for how long, and who can access it? Get specifics — not "industry-standard security."
3. What happens to our data if we cancel? Ask for a deletion timeline and proof it happened.
4. Is there a signed data processing agreement, and does it name subprocessors? If the vendor hands your data to a third party, you need to know who.
5. What's the incident-notification commitment if there's a breach? Find out how fast you'd hear about it, and from whom.

If a vendor can't answer these in writing, that's your answer.

Turning Answers Into an Actual Policy

Getting answers is step one. Doing something with them is the part firms skip.

Build a one-page questionnaire. Use the five questions above. Send it to every AI vendor you use today, not just new ones you're evaluating.

Fix your engagement letter. Generic AI boilerplate doesn't satisfy Opinion 512's informed-consent standard. Name the tools you use and describe the actual risk in plain language.

Assign ownership. Pick one person — managing partner, office administrator, whoever — who owns this. "Everyone's job" means no one checks it after the first month.

The Bottom Line

This isn't a someday issue. It's a compliance issue today, and it's cheap to fix.

Here's a two-week starting plan:

• Week 1: List every AI tool your firm uses. Send the five-question questionnaire to each vendor.
• Week 2: Update one clause in your engagement letter to reflect what you actually learned. Assign one person to own vendor reviews going forward.

No new hires. No new software spend. Just a documented answer on file instead of an assumption.

If you're not sure where your firm stands, that's exactly what a free audit is for. We'll walk through your current AI tools, flag the governance gaps, and hand you a plan — no obligation, no sales pitch.