Editor’s note: Dr. Heather Signorelli, DO, reviewed this piece as operational guidance. It is not medical, legal, or compliance advice, and NatRevMD does not endorse any specific AI vendor. Verify any workflow against your own HIPAA and payer obligations.
Using AI with patient data in a medical practice is not dangerous because the technology is reckless. It is dangerous because it is easy, and easy is exactly what causes a well-meaning staffer to paste something they should not into a box that keeps it forever. The good news is that you do not need a compliance department to use AI safely. You need four rules that everyone on your team understands. And your workflow, not hope, should enforce them.
Here are the four. They are short on purpose. A rule nobody remembers protects no one.
Rule 1: No PHI in a consumer AI tool without a signed BAA
This is the rule that all the others serve. Under HIPAA, you may share protected health information with a vendor only if that vendor has signed a Business Associate Agreement. That agreement means the vendor accepts legal responsibility for the data. The free and consumer tiers of the major AI tools generally do not come with a BAA, which makes them the wrong container for any patient data, full stop.
That does not mean these tools are off-limits. It means patient identifiers are off-limits inside them. And if you want to route PHI through a paid or enterprise tier that offers a BAA, get the current agreement in writing. Then have someone qualified read it before you rely on it. Terms change, vendors rename tiers, and last quarter’s blog post is not a compliance decision.
Rule 2: De-identify patient data before you prompt, every time
Since most billing and front-desk value does not require a specific patient, the practical enforcement of Rule 1 is a habit: scrub the identifiers before anything touches the AI. Names, dates of birth, medical record numbers, member IDs, addresses, phone numbers, emails, full dates of service, account and claim numbers. Replace them with neutral stand-ins like “the patient,” “Payer A,” or “CPT 99214.”
The AI does not need to know who the patient is to explain a denial code, draft an appeal skeleton, or write a benefits script. It needs the structure of the problem. So you feed it the structure, get the useful output, and then fill the real details back in by hand inside your own secure systems. The PHI never leaves the building. Make this the first step of every AI standard operating procedure, not an afterthought, and Rule 1 enforces itself.
A specific warning on screenshots. It is easy to paste an image to save typing and forget that its corner shows a patient banner. Crop and check every image, or simply do not upload images of clinical or claims screens at all.
Rule 3: Treat every output as a draft, never a final answer
AI is a fluent, confident, tireless drafter, and it will sometimes be fluently, confidently wrong. It will invent a timely-filing limit, cite a policy number that does not exist, or state a coverage rule that sounds authoritative and is made up. If your team acts on those inventions, AI stops being a time-saver and becomes a source of expensive mistakes.
The rule is simple: AI drafts, humans decide. Before anyone acts on a claim, especially a payer rule, confirm it in the payer’s portal or your verified reference. A person should read and personalize each letter before it goes out. And someone credentialed signs off on any code the AI discusses. Used as a drafting and thinking tool with a human check, AI is a genuine asset. Used as an oracle, it is a liability.
Rule 4: Put the rules in the workflow, not just in people’s heads
The first three rules fail the moment they depend on someone remembering them under pressure at the end of a long day. Willpower is not a compliance program. So the fourth rule is about design: build the guardrails into how work actually gets done.
That means posting a one-page written AI-use policy where staff work. Name one approved tool the team may use, rather than allowing a free-for-all. Bake a de-identification step into the standard operating procedure, so scrubbing identifiers is just how the task starts. Set a clear default of “when in doubt, do not paste it.” And revisit the policy whenever a vendor changes its terms. When the rules live in the workflow, a rushed staffer follows them automatically, because the process does not offer an unsafe path.
The mindset that keeps AI and patient data safe
Under all four rules is one instinct: err toward caution. The cost of de-identifying an extra field or double-checking one more payer rule is a few seconds. The cost of a breach or a bad appeal built on an AI hallucination is a very different kind of day. Small practices do not have a compliance team to absorb those mistakes. That is why the rules have to be simple enough to actually follow.
Get these four right and your practice can use AI confidently, because your design closes off the risky paths and leaves the useful ones wide open. That is the whole goal: real help on the repetitive work, with the patient’s data never in a place it should not be.
We built our AI Kit around these exact four rules, with a ready-to-adopt use policy and de-identified prompt templates so the safe path is also the easy path. If you want the guardrails and the tools in one place, start there.
Protecting patient data while using AI is non-negotiable in how we run AI in medical billing.
Get the AI Kit: https://eligibility.natrevmd.com/natrevmd-ai-kit-tool
