Mark Cuban posted a suggestion for doctors in private practice: sit down with the patient and configure their AI assistant. Custom instructions, a set of saved prompts, and a recurring task that asks the patient the same questions every day at 5pm and sends the results to the doctor. The doctor emails over new questions, and the patient’s AI picks them up the next morning.
His summary line was, “Help the patient use AI to help you help the patient.”
Swap in a client for the patient and it works the same way.
What it looks like in a law practice
Personal injury: the daily symptom, treatment, and missed-work log that every client agrees to keep and almost nobody keeps.
Family law: expenses, exchange times, and incidents recorded the day they happen instead of reconstructed under oath a year later.
Employment: contemporaneous notes about what was said, by whom, and when.
Business and estate clients: a quarterly nudge about renewals, entity housekeeping, and the life events that should trigger a call to you.
The value isn’t the AI writing anything. It’s that somebody asks the client the right questions on a schedule, which is the part that always breaks down.
The confidentiality problem
Whatever your client types into a consumer chatbot becomes a written record, and courts so far treat AI companies as third parties rather than as part of the privileged circle.
In United States v. Heppner (S.D.N.Y., February 2026), Judge Rakoff held that a defendant’s chatbot documents weren’t protected by attorney-client privilege or the work product doctrine. Part of the reasoning: counsel hadn’t directed him to use the tool, so neither he nor the AI was acting as an agent of counsel. A week earlier, a Michigan federal court in Warner v. Gilbarco reached the opposite conclusion on work product. Different courts, opposite results, one week apart.
Why the problem might be the opportunity
That “by or at the behest of counsel” language cuts both ways. A client who builds an AI journal on their own is in Heppner’s position. A client whose lawyer designed the setup, wrote the questions, and directed the work in anticipation of litigation is arguing from better ground.
I don’t want to oversell this. No court has approved an arrangement like it, privilege and work product are separate doctrines with separate waiver rules, and this area is moving month to month.
But it suggests a sensible design rule. Keep the client’s AI limited to facts — dates, symptoms, expenses, events. Keep out speculation about the case, and keep out anything you’ve told them. Deciding what belongs on each side of that line is lawyer work, and it’s the part a client can’t do alone.
What works today
ChatGPT runs scheduled tasks on its paid plans, and the other tools either have their own version or are adding one. The piece Cuban describes where the AI emails everything to the doctor is the shakiest part — these tools mostly notify the user, and pushing output to a third party takes extra plumbing. Having the client forward the day’s entry works now and needs no setup.
;-)
Ernie
P.S. This is the kind of thing we take apart together in the Inner Circle — including the parts nobody has figured out yet
→
https://ernietheattorney.net