AI is arriving in Montana healthcare the same way it arrived everywhere else: a transcription tool here, a scheduling assistant there, a staff member quietly pasting notes into a chatbot. Most of the compliance risk lives in how these tools get deployed: as if HIPAA had an exemption for new technology, which it does not. Four risks account for most of the trouble: PHI flowing into tools that never signed a business associate agreement, your patient data training someone else’s model, AI systems missing from your security risk analysis, and building for today’s rule while the rule itself is changing.
Here is each one, and what to do about it before an auditor asks.
There Is No AI Exemption in HIPAA
Every rule that governs your EHR governs an AI tool the moment it touches protected health information: the Privacy Rule, the Security Rule, breach notification, the minimum necessary standard, all of it. An AI scribe that hears a patient encounter is handling PHI. A chatbot summarizing a referral letter is handling PHI. AI can even create new PHI, because a model’s output about a patient is just as protected as its input.
That sounds obvious written down. In practice, AI tools arrive through the side door: a free trial here, a feature switched on inside software you already own there. Each one is a compliance decision, whether or not anyone treated it as one.
Risk 1: PHI in a Tool That Never Signed a BAA
Any AI vendor that accesses PHI on your behalf is a business associate, and HIPAA requires a signed business associate agreement before patient data flows. Not after the pilot. Before.
Two traps catch healthcare organizations here. The first is the consumer tier: the free or personal versions of popular AI assistants are generally not HIPAA-appropriate in their default configurations, and no BAA is available for them. The enterprise tiers of some tools can be, but that has to be verified with the vendor, in writing. The practical rule for staff: if the vendor will not sign a BAA for the tier you are using, no patient information goes into it, named or not.
The second trap is taking “HIPAA compliant” at face value. The phrase is marketing language; no certifying body stands behind it. Due diligence means reading the BAA’s terms on how PHI is processed and stored, what happens to data after processing, and asking for independent evidence like a SOC 2 Type II report.
Risk 2: Your Patients, Someone Else’s Training Data
The question that matters most in any AI contract is one sentence long: will our data be used to train or improve models that serve other customers?
If the answer is yes and patients have not authorized it, that is a serious violation waiting to be noticed. If the answer is unclear, treat it as yes until the vendor puts it in writing. The right answer looks like a contractual commitment that your PHI is not used for model training, is segregated from other customers, and is deleted or returned on termination.
Risk 3: AI Is Missing From Your Security Risk Analysis
The security risk analysis is the most commonly cited deficiency in HIPAA enforcement actions, and the arrival of AI gives that old problem a new edge. An SRA that was accurate in January is incomplete the day an AI tool starts touching PHI.
Fold AI into the analysis you already run: which systems touch PHI, what data they can reach (the minimum necessary standard applies to AI access too), where output goes, and what happens if the tool is wrong or breached. For clinical tools, that review should also cover accuracy and bias, because a model that performs badly for part of your patient population is a patient-safety problem and a liability problem at once.
Montana saw what dependence on a single digital intermediary looks like when the Change Healthcare attack disrupted providers across the state. AI vendors are becoming the same kind of dependency; treat them with the same seriousness in your risk analysis and your contingency planning.
Risk 4: Building for Today’s Rule While the Rule Changes
In January 2025, HHS proposed the first major update to the HIPAA Security Rule since 2013: mandatory multifactor authentication, encryption at rest and in transit as a requirement rather than an addressable option, technology asset inventories, network segmentation, and tighter recovery expectations.
As of this writing, that proposal has not been finalized, and the current Security Rule still governs. But “addressable” has never meant optional, and the direction of travel is unambiguous. An organization adopting AI today should build to the proposed standard now: MFA and encryption everywhere, plus an asset inventory that includes every AI tool in use. Doing it now costs a project. Doing it after a final rule lands costs a scramble, and doing it after an audit finding costs far more.
What “Ready” Looks Like Before the Audit
The checklist is shorter than the anxiety suggests:
- Inventory every AI tool in use, including the unofficial ones. You cannot assess what you have not found.
- A signed BAA for every vendor touching PHI, with training-data use addressed explicitly.
- An updated security risk analysis that names AI systems and their data access.
- A written AI policy for staff, covering which tools are approved and what never goes into unapproved ones.
- MFA and encryption treated as required, whatever the current rule technically allows.
For Montana clinics and hospitals, this is the audit-readiness work we handle every week as part of healthcare IT support.
Know Where You Stand Before the Examiner Does
Compliance gaps tend to surface during an audit, which is the most expensive possible way to learn about them. The free 30-Minute Compliance Reality Check gives you the earlier option: a plain-language look at where your controls actually stand, including how AI fits in, before anyone with a checklist arrives.
Frequently Asked Questions
Can our clinic use ChatGPT or other AI assistants with patient information?
Only through a version of the tool where the vendor will sign a business associate agreement, configured so PHI is protected, and approved in your policy. Consumer and free tiers generally do not qualify. Removing names is not enough on its own, because combinations of details can still identify a patient.
Do AI vendors really count as business associates?
Yes, whenever they create, receive, maintain, or transmit PHI on your behalf. That triggers the BAA requirement before any patient data is shared, the same as a billing company or a cloud EHR vendor.
Has the new HIPAA Security Rule taken effect?
Not yet. The January 2025 proposal, which would mandate MFA and encryption among other changes, remains unfinalized as of late 2026, and the current Security Rule still applies. Preparing for the proposed requirements now is the practical move, since each one is already standard security practice.
What is the most common HIPAA finding we should worry about?
An inadequate or outdated security risk analysis. It is the deficiency regulators cite most often, and adding AI tools without updating the analysis is the fastest way to earn it.
Where should a small Montana practice start?
With the inventory. List every AI tool staff actually use, official or not, then work the checklist above in order. If you want an expert pass over the whole picture, that is exactly what the Compliance Reality Check is for.
