How can you use AI without putting client information into ChatGPT?
Two honest answers, and only one of them scales. Where the confidentiality line actually sits, and why the paste going in is the exposure rather than the answer coming back.
There are two honest answers, and only one of them scales. You can strip the client out of the question before you ask it, which works for a narrow set of tasks and fails quietly on the rest. Or you can stop moving the file to the tool and put the tool inside the system that already holds the file, which is what most firms end up doing once the first approach costs them an afternoon.
Either way, the thing to be clear about is where the exposure actually happens. It is not the answer coming back. It is the paste going in.
Rule 1.6 does not care what the tool is called.
Information relating to the representation is confidential, and you owe reasonable efforts to prevent its disclosure. Putting it into a third party's system is a disclosure, and the question is whether the efforts you made around it were reasonable. That turns on what the vendor's terms say happens to the input, whether the firm decided this deliberately or an associate decided it alone at 11pm, and whether anyone could reconstruct who the client is from what you typed. ABA Formal Opinion 512 walks through the same ground for generative tools specifically, and several state bars have added their own. None of them ban the tools. They ask whether you knew what you were handing over.
What is fine to type into a general chatbot.
Usually fine
Doctrinal research with no facts attached. Rewriting your own boilerplate. Explaining a statute to yourself. Drafting a clause from scratch that has no party in it. A question about how a court rule works. In each of these the confidential part never leaves your office, because it was never in the question.
Riskier than it looks
The anonymized fact pattern. Remove the names and a wrongful death arising from a specific crash on a specific date in a county of forty thousand people is still one identifiable family. Unusual facts re-identify themselves, and the more unusual they are, the more likely they were the reason you asked.
Not without a decision
Uploading a document from the file. Pasting a deposition excerpt, a contract, a medical page, an email thread with the client. These are the tasks worth doing, which is exactly why they should run somewhere the firm has already reviewed the terms rather than in whichever tab is open.
Delgado · agent
Nothing pastedDo not move the file. Move the tool.
The copy-and-paste step is the part that creates the problem, and it exists only because the assistant lives somewhere your case file does not. CaseFuze runs the agent against the matter you already have open, so the parties, the dates, the documents and the ledger are in context without anything being typed into a browser tab. The models are served under HIPAA-compliant terms through Amazon Bedrock, they retain nothing, and none of your material is used for training.
- One decision, made once. The firm reviews the terms of the platform it already uses instead of every attorney reviewing a tool alone.
- Permissions carry across. The agent cannot show anyone a matter they could not already open themselves.
- No wrong option in the menu. Every model offered is already under the same terms, so there is nothing to pick by mistake.
Six things to settle before somebody settles them for you.
Find out what is already being used
Ask the question without consequences attached and you will get an honest answer, which is usually that several people are already using something on personal accounts. That is the actual starting position, not the policy you have not written yet.
Name one approved tool
A short list of one is easier to follow than a long list of rules. People reach for whatever is available, so make the approved thing the most available thing and the problem mostly solves itself.
Ban personal accounts for client work
Not because the technology differs, but because the terms do, the firm cannot see it, and nothing is logged. A consumer account is a system the firm has no relationship with holding material the firm is responsible for.
Write down what needs no permission
Research with no facts, your own boilerplate, formatting. Telling people what they may do without asking is what stops the policy being ignored, and it is the half most firms leave out.
Decide the engagement letter question once
Whether you tell clients how you use technology on their matter, in what words, and from which date. Deciding it in advance is far easier than deciding it in the middle of a conversation with a client who read something worrying.
Say who owns the output
Whoever signs the document is responsible for every word in it, including the words a model wrote. Make that explicit rather than assumed, because the failure that reaches a court is never the tool. It is a draft that went out without anyone reading it as their own work.
The safest place for the file is the one it is already in.
Ask about a matter and the agent already knows it. Upload a document and the models read it, answer, and keep nothing. Your choice of provider, all under the same compliant terms, inside the system holding your matters, your ledger and your documents. In the $49 plan.
Keeping client information out of AI, answered.
Is a paid ChatGPT account enough to make it safe?
Does redacting names solve it?
What if the matter is already public record?
My staff are using AI on personal accounts. How bad is that?
What about medical records specifically?
Does an assistant built into my software solve the whole problem?
Keep the file where it already is.
An agent inside the system holding your matters, on models that retain nothing. $49 per user per month.