HIPAA and AI tools: what a small practice checks before connecting anything
No AI tool is HIPAA compliant on its own. Compliance is a property of the contract you signed, the data you allowed through, where that data lives, and what the tool is permitted to decide. Four checks cover it, and a practice can run them in an afternoon before the first demo.
Before any AI tool touches patient data, a small practice checks four things: the vendor has signed a Business Associate Agreement, the tool receives only the minimum data its job needs, the practice knows where that data lives and for how long, and the tool is not permitted to make a clinical or coverage decision. Consumer chatbots fail the first check: OpenAI does not sign a BAA for consumer or Team ChatGPT (HIPAA Journal), and Anthropic's BAA covers only its API and HIPAA-activated Enterprise plans (Anthropic).
Compliance is the setup, not the software
Vendors say "HIPAA compliant" the way they say "AI-powered," and neither phrase is a property of a product. HIPAA obligations attach to the covered entity, which is the practice, and to its business associates, which are the vendors that handle protected health information on its behalf. A tool becomes part of a compliant setup when the contract, the data flow, and the operating rules are right. The same tool in a different setup is a breach waiting for a complaint.
That is good news for a small practice, because it means the four checks below are within the practice's control, and none of them requires an IT department.
Check 1: the Business Associate Agreement
Any vendor that will access, process, or store PHI on the practice's behalf signs a BAA before it does. Not after the pilot. Before. The BAA is what makes the vendor legally responsible for the safeguards and the breach notification. Without one, disclosing PHI to the vendor is itself the violation, however secure the vendor's servers happen to be.
This is where the consumer tools fall out. HIPAA Journal states that OpenAI will not enter into a BAA for the Free, Plus, Team, or Enterprise versions of ChatGPT, and that its healthcare product, introduced in January 2026, is built for hospitals and regulated environments. Anthropic's BAA covers its first-party API and Enterprise plans where the Primary Owner has activated HIPAA, and does not cover Free, Pro, Max, or Team. Activation on Enterprise is a one-way step the Primary Owner takes after reviewing the agreement and the implementation guide, and Anthropic's Claude Code and Cowork are excluded from coverage. The cloud platforms have their own BAA programs, and a vendor built on one of them still has to sign its own BAA with you.
Ask every vendor the same three questions: will you sign a BAA, which of your products does it cover, and which does it not.
Check 2: minimum necessary data
A signed BAA is permission to handle PHI, not an invitation to handle all of it. HIPAA's minimum-necessary standard means the tool gets the data its job requires and nothing more. ai.dentist, written by a practicing dentist about the same problem in dental software, gives the test in one example: a scheduling tool needs appointment times and durations, not clinical notes or treatment details.
Apply it to each tool. A receptionist gets the schedule and contact details. A scribe gets the audio of the visit and writes one note. A billing tool gets codes and payer data. A vendor that asks for full chart access to do one of those jobs has not thought about the boundary, and the practice should not be the one to discover the consequences.
Check 3: where the data lives
Three questions, each with a short answer the vendor should have ready. Where is the data stored, and in what country? How long is it kept, and can the practice set that? Is it used to train the vendor's models? The last one matters because "no training on your content by default" is a product setting on many plans, including Claude's Team and Enterprise tiers per Anthropic's pricing page, and it is not the same thing as a BAA. A tool can promise not to train on your notes and still not be a business associate.
There is also a fourth answer, which is that the data does not leave the building. Some workflows can be built to run on the practice's own machine, with nothing sent to a cloud model at all. That is not always possible, and it is always worth asking about, because it removes most of the questions above.
Check 4: what the tool may not decide
The last check is not in the statute and it is the one that protects patients. Write down the decisions the tool is not permitted to make: diagnosis, triage severity, medication, coverage and denials, and anything else where the practice would want a clinician's name on the outcome. Then check what the tool does when its input is uncertain. The right answer is that it stops and shows the gap. The wrong answer is that it fills the gap with something plausible.
Run the four checks before the first demo.
If a tool fails any of them, the demo is a waste of an afternoon. If it passes all four, the ten questions will tell you whether it is the right first project for the practice.
The pasted-note problem
The most common HIPAA problem with AI in a small practice is not a vendor at all. It is a nurse or a front-desk employee pasting a note, a message, or a lab result into a free chatbot to reword it. That is a disclosure of PHI to a company that has not signed a BAA, and the practice is responsible for it. The fix has two parts: a written policy that names which tools are permitted and which are not, and a permitted tool that does the same job, so the policy is not asking people to work slower.
How we built for a clinic
Our patient presentation automation for an orthodontic clinic is the fourth answer to Check 3. The system runs on the clinic workstation and assembles presentations from intake PDFs and clinical photos. Patient inputs, extracted fields, images, and working files stay local; per-run files are wiped after delivery. Missing inputs surface as warnings rather than invented values, and the operator confirms identity and decides whether the output is ready. Nothing goes to a cloud model, so no BAA question arises for that workflow. The record makes no claim about adoption or labor saved.
The one-page policy
Every practice can write this in an hour, and it is most of the compliance work for AI tools.
- Permitted tools, by name, each with a signed BAA on file.
- Prohibited tools, by name, including the consumer chatbots, and the sentence that says PHI never goes into them.
- What each permitted tool receives, in one line per tool.
- Where each tool's data lives, how long, and whether it trains models.
- What no tool decides: diagnosis, triage, medication, coverage.
- Who reviews exceptions, by name, and how fast.
Hand that page to every vendor before the demo and to every employee on the first day. It is not a legal opinion, and the practice's counsel should see it. It is the difference between a policy and a hope.
What to hold on to
- Compliance is the setup. The contract, the data flow, where the data lives, and what the tool may decide. Not a product label.
- BAA first, or nothing. Consumer ChatGPT tiers have none. Anthropic covers its API and HIPAA-activated Enterprise, not Free, Pro, Max, or Team.
- Minimum necessary data. A scheduler gets the schedule. A scribe gets one visit. Full chart access to do one job is a warning sign.
- "No training on your data" is not a BAA. Ask where it is stored, how long, and whether it can run locally instead.
- Write down what no tool decides. Diagnosis, triage, medication, coverage. Then check what the tool does when it is unsure.
Questions owners ask us
Is ChatGPT HIPAA compliant for a medical practice?
Not the Free, Plus, Team, or Enterprise versions. OpenAI does not sign a Business Associate Agreement for them, so entering patient information is an impermissible disclosure (HIPAA Journal). OpenAI introduced a healthcare product in January 2026 aimed at hospitals and regulated environments, and its API can be covered under a BAA on application.
Does Anthropic sign a BAA for Claude?
Yes, for two surfaces: the first-party Claude API on a HIPAA-ready organization, and Claude Enterprise plans where the Primary Owner has activated HIPAA compliance and accepted the BAA (Anthropic). It does not cover the Free, Pro, Max, or Team plans, and Claude Code and Cowork are excluded from coverage (HIPAA-ready Enterprise plans).
What is minimum necessary access for an AI tool?
The tool receives only the data its job requires. A scheduling tool gets appointment times and contact details, not clinical notes. A scribe gets the audio of one visit. A billing tool gets codes and payer data. A vendor that asks for full chart access to perform one task has not designed for the boundary.
Is "no training on your data" the same as HIPAA compliance?
No. It is a product setting that keeps your content out of the vendor's model training, and it is worth having. A Business Associate Agreement is a contract that makes the vendor legally responsible for safeguarding PHI and notifying breaches. A tool can have the first without the second.
Can an AI workflow run without sending patient data to the cloud?
Some can. Our patient presentation automation runs entirely on the clinic workstation: inputs, extracted fields, images, and working files stay local and are wiped after delivery. Not every workflow can be built that way, and it is always worth asking, because a local-first design removes most of the vendor questions.
What should a small practice’s AI policy say?
Which tools are permitted, each with a BAA on file; which are prohibited, including consumer chatbots; what each permitted tool receives; where its data lives and whether it trains models; what no tool decides; and who reviews exceptions. One page, reviewed by counsel, given to every vendor before a demo and every employee on day one.
Sources
Vendor BAA positions are quoted from the vendors and from HIPAA Journal; the minimum-necessary example from a practicing dentist; the local-first build from our record. This is not legal advice. Checked September 5, 2026.
- HIPAA Journal, "Is ChatGPT HIPAA Compliant?". January 13, 2026.
- Anthropic, "Business Associate Agreements (BAA) for Commercial Customers" and "HIPAA-ready Enterprise plans". Coverage by product and the activation steps.
- Anthropic, Claude pricing. The no-training-by-default setting on Team and Enterprise.
- ai.dentist, "Open Dental API Integration: HIPAA-Compliant AI Tool Setup". March 26, 2026, by Dr. Jordan Thomas. The minimum-necessary example.
- Patient presentation automation, our local-first implementation record.