Your Front Desk Is Already Using AI and Nobody Signed Anything

Nobody in your practice decided to adopt AI. It arrived the way the office microwave arrived, one person at a time, because it was useful and free and nobody said not to.

Ask your front desk this week whether anyone has used a chatbot to help word a reply to a patient. Ask gently, because the answer is almost certainly yes, and the person who did it was trying to do a good job. They had a difficult message to answer, they were busy, and a tool on their phone wrote a better version in nine seconds.

Here is the part that is hard to explain and impossible to unlearn: that was a disclosure of protected information to an outside company, and whether it was lawful has nothing to do with how careful the person was.

A rewritten message is a disclosure

The rules do not ask whether a vendor's security is good. They ask whether you had an agreement with that vendor before you handed over the information. Pasting a patient's message into a consumer chatbot is a disclosure to the company that operates it, and without an executed agreement it is impermissible on its face, no matter how strong the encryption behind it happens to be.

That framing matters, because the instinct in every practice is to argue about the security. It is the wrong argument. A vendor can run the most secure data center in North America and the disclosure is still unauthorized if nobody signed.

Which accounts carry a signed agreement

This is the detail that gets reported wrong constantly, usually as a flat claim that a given chatbot "is not compliant." The truth is narrower and more useful: it depends entirely on which plan you are on.

As of autumn 2026, OpenAI will execute a healthcare agreement for sales-managed Enterprise, for its healthcare and clinician products, and for the API. It will not execute one for the free tier, Go, Plus, Pro, or the self-serve Business plan. Same model, same company, same interface. The paperwork is the difference, and the paperwork is attached to the subscription rather than to the technology.

Which means the staff member who upgraded to a personal Plus account to be more productive has not solved anything. Paying for a consumer plan buys faster responses, not permission.

Removing a name is not de-identification

The reasonable-sounding middle path is to strip the identifying details and use the tool anyway. That path is real, and the standard for it is stricter than people expect.

Properly de-identified information is no longer protected information, and no agreement is required. But the permitted methods are specific, and deleting the patient's name is not one of them on its own. Dates, ages above a threshold, geographic detail finer than a state, phone numbers, email addresses, record numbers and several other categories all have to go, or a qualified expert has to certify that the re-identification risk is very small.

"Can you make this message to a 41-year-old in Garden City sound warmer" has not been de-identified. It has been lightly disguised.

Where the exposure actually sits

Three places, in the order we find them.

The first is drafting patient replies, which is the most common and the easiest to fix. The second is dictation, where somebody speaks a clinical summary into a transcription app on a personal phone and the audio goes to a company nobody has ever evaluated. The third is the quiet one: a staff member uploading a spreadsheet to get help cleaning it up. A lead list with phone numbers and inquiry notes is one file, one drag, and several hundred disclosures at once.

That third line is not softness. Without it you will never learn where the exposure is, because the people who know are the people with the most reason to stay quiet.

Questions practice owners ask us

Can staff use AI for anything at all?

Yes, and they should. Advertisement copy, a social caption, a job posting, a letter to a referring physician with no patient detail in it, a policy draft, a spreadsheet formula. The constraint is the patient information, not the tool.

Does upgrading to a paid plan fix it?

Only if the plan is one the vendor will sign for, which generally means a sales-managed or healthcare-specific plan rather than the one you can buy with a credit card in a minute.

What about the AI features already inside our existing software?

Ask the vendor two questions in writing: which subcontractor processes the data, and is that subcontractor covered by the agreement we already have with you. A feature that appeared in a product update does not inherit paperwork that was signed before it existed.

If a staff member does this, who is liable?

The practice. That is the whole reason this is an owner's problem rather than a training problem.

The uncomfortable conclusion

Most practices reading this have already had an unauthorized disclosure, probably several, almost certainly without harm and almost certainly without anybody noticing. The useful response is not an investigation. It is a written policy, an approved tool with its paperwork in order, and the understanding that the next person to arrive at your front desk will also have a chatbot on their phone.

The adjacent problem, and the more expensive one, is what happens when AI listens to the consultation itself. That is the next post.