If this sounds like your kind of team
We’d love to hear from you. Whether you’re ready to talk or just want to see how your current setup measures up, there’s an easy way to start.
Most of the medical practices we support in Dallas are already using AI somewhere. A front-desk person drafting appointment reminders, an office manager summarizing a vendor contract, a physician cleaning up a referral letter. That’s fine, and often a good use of time and resources. The trouble starts when the same tool that wrote the reminder gets handed a patient chart, because the rules for those two jobs are completely different.
Nobody is going to ban AI from healthcare, and we wouldn’t suggest it. The owners we worry about fall into two groups. The ones who avoid AI entirely because compliance scares them, and the ones whose staff started using it months ago without anyone deciding where patient data was allowed to go and what was allowed to go in it. The second group is bigger than most owners think, and it usually includes them doing these actions as well.
The Department of Health and Human Services does not certify or endorse any technology as HIPAA compliant. A vendor can put a security badge on its website, and that badge tells you nothing about whether your practice is compliant. Compliance depends on your organization, your agreement with the vendor, how the tool is configured, and what information flows through it.
So the first question is not “is this tool compliant.” The first question is whether you’re putting Protected Health Information (PHI) into it at all. PHI is defined by 18 identifiers, and the list is broad. Names, addresses, any date tied to a person, phone numbers, email addresses, record numbers, photos, and more. If any of those sit alongside anything about someone’s health, it’s considered PHI.
If the answer is no, your risk changes completely. Policies, marketing, job descriptions, meeting agendas, generic patient education, internal procedures, none of that needs patient data, and all of it is fair game in whatever AI tool your team likes.
When a cloud provider creates, receives, stores, or transmits electronic PHI on your behalf, HHS generally treats that provider as a business associate. That means you need a Business Associate Agreement before the data goes in. Without one, sending PHI to the tool is a violation on its own, however carefully you use it.
The standard AI plans don’t come with one. Free ChatGPT, the personal paid tiers, the small-business team plans from any major vendor. The BAA-covered options exist, and the picture has changed a lot in 2026. Anthropic signs one for its API and for sales-assisted Claude Enterprise plans, and an administrator still has to switch on the HIPAA configuration before it applies. OpenAI signs one for its API and for sales-managed ChatGPT Enterprise, plus a ChatGPT for Healthcare product. Google will sign one too. What none of them cover is the plan you’d sign up for with a credit card on a weekday or the ones your employees are using.
A BAA is not a permission slip. Signing one doesn’t end your responsibility. You still have to look at how the tool is used, what data is exposed, who can reach it, how long it’s kept, and how it’s backed up. HHS treats that risk analysis as the foundation of the Security Rule, and it’s something we’d expect your compliance advisor to walk through with you.
This is where practices get caught. Someone vets the AI platform, signs the agreement, and forgets about everything around it. The voice recorder. The transcription service. The automation tool that moves the transcript into the CRM. The browser extension. The shared drive. The connector between two systems that looked harmless when someone set it up.
Every hop where PHI is created, received, stored, or moved needs the same scrutiny. A compliant AI connected to a non-compliant scheduling app breaks the chain, and it looks perfectly fine from the outside. Before anything gets connected to a system holding patient data, ask the vendor in writing whether they’ll sign a BAA. If they won’t, the tool never touches PHI.
One more thing on this. Don’t send an AI browsing agent into a patient portal or an EHR login to pull data for you. Those agents read everything they can reach, and hidden instructions on a web page can redirect them. Keep them well away from anything near patient records.
Here is the situation we are running into most often. A practice has a perfectly reasonable official process, and a billing coordinator pastes a denied claim into her personal ChatGPT account to reword the appeal. Or a nurse drops a visit transcript into a free summarizer because it saves twenty minutes at the end of a long day.
That’s shadow AI, and it isn’t an IT problem so much as a people and process problem. The fix is a short written policy that names the approved tools, says what can go into them, and says what never can. Then the team hears it more than once and there is tracking and policies in place to prevent it.
Ask yourself two things. Could you name every AI tool your staff used last week and if patient data was input into it or not. If patient’s information turned up in one of them, would you find out from your own systems or from other source where information was leaked?
Properly de-identified information is no longer PHI, and that opens up a lot of safe uses. But HHS recognizes two methods for getting there. Safe Harbor, which means removing all 18 identifiers, and Expert Determination, which means a qualified expert documents that the risk of re-identification is very small. Stripping out a name and leaving the date of birth, zip code, and record number does not qualify.
If your team wants AI to work on patient files, tools like RedactLocal and PrivacyScrubber remove the identifiers from a document on your own computer before it reaches any AI, and you restore the real names afterward. Your EHR likely has AI features built in as well, and those are usually already covered under your agreement with the EHR vendor. That’s the safest place to start.
“Does the vendor train on our data” and “does the vendor keep our data” are separate questions. A tool can decline to train on your prompts and still retain them for thirty days where a support engineer could read them. Ask both, ask whether an administrator can control the settings, and ask what happens when you delete something.
The one people forget is accuracy. Even with the data handling done right, a generative AI tool can produce something wrong or incomplete, and in coding, billing, medication, or patient-care contexts that’s a different kind of exposure entirely. Human review of anything that touches a patient decision is not optional.
We manage the Microsoft 365 environment, the endpoints, the backups, and the security stack for our medical clients, and we monitor the account compromise that turns a small AI mistake into a breach notification. Security awareness training is part of what every managed user gets from us, and the AI policy conversation is a part of that. When a client asks “can we connect this tool to our Microsoft 365,” Graham’s answer usually is. Yes, and here are the risks you’re going to run into, so let’s look at them before you turn it on.
What we don’t do is complete your compliance attestation forms or sign off on your legal exposure. That’s a job for your HIPAA counsel or compliance professional, and we’re glad to work alongside them. Everything above is what we’ve learned in the past few years with AI along with protecting practices in Dallas for more than a decade, shared as education rather than legal advice. Please have your own advisor verify your setup before you rely on it.
Reach out to us!
If you’re a practice in the Dallas area and you’d like a second set of eyes on the tools your team is already using, give us a call. We’ve usually seen the same tool at another office down the road, and we’ll tell you plainly whether it belongs anywhere near your patient data.
We’d love to hear from you. Whether you’re ready to talk or just want to see how your current setup measures up, there’s an easy way to start.