In plenty of small offices around Western New York, an AI chatbot now sits open in a browser tab. The insurance agent asks it to tidy up a claim summary. The bookkeeper pastes in a messy list of transactions. The front desk at a dental practice drafts a patient letter, and the property manager has it reword a lease reminder. It saves time, and on a busy Friday afternoon nobody wants to give that up.

The trouble is what goes in with the request. A claim summary has a policyholder’s name and address. A list of transactions has account numbers. A patient letter has a diagnosis. Once that text is pasted and sent, it sits on a server run by a company outside the business, and the person it describes never agreed to that. Sooner or later, an owner starts to wonder whether something has gone wrong.

There’s no single yes or no answer, and anyone who offers one is guessing. But New York law, the federal health privacy rules and OpenAI’s own published policies give an owner enough to make a sensible call. Here’s what each of them says, and what a small office can do before the next paste, and after one that has already happened.

What happens to client data once it’s in ChatGPT

Start with the account, because it shapes everything else. On a personal ChatGPT account, OpenAI says it may use your content to train its models, depending on your settings. The switch is called Improve the model for everyone, under Settings and then Data controls. On ChatGPT Business, Enterprise and Edu, OpenAI says it doesn’t use inputs or outputs for training by default.

Deleting the conversation doesn’t undo the paste right away, either. OpenAI says a deleted chat is removed from your account immediately and scheduled for permanent deletion from its systems within 30 days, with exceptions for material it must keep for security or legal reasons. Files follow their own rules: OpenAI’s help pages say deleting a chat doesn’t delete files saved to its separate Library, so an attached spreadsheet can outlive the conversation.

None of this means OpenAI is careless with what it receives. It means the text is now held by a third party under that third party’s terms, and the business that collected it no longer decides on its own how long it stays there. That loss of control is what turns a convenience into a legal question.

What New York’s breach law actually says

New York’s breach notification statute is General Business Law section 899-aa, which the SHIELD Act expanded. It protects private information: personal details combined with items such as a Social Security number, a driver’s license number, an account or card number with the code that unlocks it, medical information or health insurance details. A pasted claim file or client spreadsheet can easily hold several of them.

Under the statute, a breach is unauthorized access to or acquisition of computerized data that compromises that information. The law lists factors to weigh, including signs that the information is in the possession of an unauthorized person or has been copied. It also says good faith access by an employee for business purposes isn’t a breach, as long as the information isn’t used or subject to unauthorized disclosure. The statute says nothing about AI tools, so how a paste fits is a judgment call.

If it does count, the clock is short. Affected New York residents must be told in the most expedient time possible and no later than 30 days after discovery. There’s an exception for an inadvertent disclosure by people authorized to see the data, where the business reasonably decides misuse is unlikely, but that decision has to be documented in writing and kept for at least five years. If more than 500 New Yorkers are affected, it goes to the Attorney General within ten days.

The SHIELD Act’s other half: reasonable safeguards

The same law has a second section that matters more day to day. Section 899-bb requires any person or business that owns or licenses computerized private information of a New York resident to develop, implement and maintain reasonable safeguards to protect its security, confidentiality and integrity. That reaches businesses outside the state too, as long as they hold New Yorkers’ data.

Among the safeguards the statute lists is choosing service providers capable of maintaining appropriate safeguards and requiring those safeguards by contract. A free chatbot account opened by an employee on a lunch break is hard to square with that. Small businesses, defined by thresholds such as fewer than 50 employees, get some room: their program only has to fit their size, their activities and the sensitivity of the information they collect.

For a five-person office, that’s good news. Nobody expects a security department. What the law does expect is a decision, written down and followed, about which tools may touch client information and how. For readers who want the wider view, a guide on whether putting client data in ChatGPT is a breach compares the US, the UK, Ireland and Australia and adds a first-hour checklist.

Medical offices: HIPAA and the business associate agreement

Health care practices carry an extra layer. Under the HIPAA Privacy Rule, a covered entity may disclose protected health information to a business associate only with satisfactory assurance that the associate will safeguard it, and that assurance has to be documented in a written contract or agreement. A dental office, physical therapy clinic or pharmacy that pastes patient details into an AI tool without one is on shaky ground.

OpenAI is unusually direct about which products it will sign that agreement for. Its help center says only ChatGPT Enterprise or Edu customers with a sales-managed account are currently eligible for a BAA for ChatGPT, that its API has a separate route, and that it doesn’t offer a BAA for ChatGPT Business. A personal account isn’t on the list at all.

New York adds one more step for health data. A covered entity that reports a breach to the federal Department of Health and Human Services must also notify the New York Attorney General within five business days of that notice. For a small practice, the simplest course is to keep patient details out of general chatbots entirely unless a signed agreement covers the specific product in use.

What to do before the next paste, and after one that already happened

If something has already gone in, act while memory is fresh. Write down what the text contained, whose information it was and which account it went into. Export the conversation through the account’s own data export before deleting anything, so there’s a record. Then call whoever handles your legal questions, and check your insurance policy for a clause on reporting incidents. Deciding alone that it was nothing is the riskiest choice.

For the future, a short written rule beats a long policy nobody reads. Client and patient details stay out of personal AI accounts. Work goes through business accounts the owner has approved. Names, account numbers and dates of birth are removed or replaced before any text goes in, because the AI rarely needs them to rewrite a letter or summarize a file.

That last step is the one people skip when the phones are ringing, so some offices hand it to software. One option is Nonimo, a desktop app for Mac and Windows that catches names, ID numbers, bank details and other personal information and replaces them with placeholders on the computer itself, before the text reaches ChatGPT, Copilot or Gemini. The letter still gets written; the real client just never appears in the prompt.