Healthcare administrators are in an uncomfortable position right now. AI is everywhere, adoption is accelerating faster than any technology in recent memory, and the consequences of getting it wrong are far more serious than in most industries. A bad AI rollout in retail costs you money. A bad AI rollout in healthcare can cost a patient, cost you a seven-figure OCR penalty, and land your organization on the front page.
What makes this harder is that the biggest AI risk in healthcare today isn't coming from the systems you officially deployed. It's coming from the tools your staff is already using without telling you.
The adoption wave
In 2023, 38% of U.S. physicians reported using AI in their practice. By 2024, that number had jumped to 66%, a near-doubling in a single year. 75% of U.S. health systems now use at least one AI application, up from 59% just twelve months earlier.
AI-powered clinical note-taking now has 68% adoption, growing 62% year over year. Physicians currently spend one to two hours on EHR documentation for every hour of direct patient care. Any tool that cuts that ratio is going to get used, approved or not.
The global healthcare AI market is now at $37 billion, growing at roughly 40% annually and projected to exceed $100 billion by 2030. The FDA had cleared or approved approximately 1,250 AI and machine learning-enabled medical devices by May 2025. Adoption at this speed doesn't wait for governance to catch up.
The problem is hiding in your own building
A December 2025 survey found that 17% of healthcare workers admitted to using unauthorized AI tools at work. That's the number who admitted it. IT teams typically detect fewer than 20% of the AI tools that employees are actually using, and audits routinely uncover around 70 unauthorized AI applications per health system.
The industry has started calling this "shadow AI." It's a more dangerous version of the old shadow IT problem. Shadow IT usually meant someone using a personal Dropbox for work files. Shadow AI means a nurse pasting a patient's medical history into ChatGPT to draft a care summary, or a billing coordinator running patient records through a free AI tool to check codes. Clinicians are burned out and documentation-buried. When they find something that works, they use it.
Under HIPAA, intent doesn't matter. Every time Protected Health Information (PHI) enters a system that hasn't signed a Business Associate Agreement (BAA) with your organization, that's an unauthorized disclosure, regardless of whether anything leaks. The disclosure is the violation, full stop.
Shadow AI incidents add an average cost of $670,000 per breach, before the OCR investigation that often follows.
What regulators are doing right now
HIPAA enforcement used to focus primarily on whether organizations had performed a risk analysis. Many organizations cleared that bar without doing much else. That's no longer the standard.
OCR resolved 21 HIPAA enforcement cases in 2025, with 76% including penalties for risk analysis failures. The enforcement focus has moved. OCR now expects organizations to demonstrate not just that they identified compliance risks, but that they documented what they did about them. Finding a gap and leaving it open is now itself an enforceable failure.
OCR imposed $4.18 million in HIPAA penalties across 13 enforcement actions in 2025, nearly double the prior year. Tier 4 penalties for willful neglect, the category that covers "we knew and didn't act," now reach $73,011 per violation with an annual cap of $2.19 million. For a health system that discovers it has been routing PHI through an unapproved AI vendor for a year, the math compounds fast.
In January 2025, HHS proposed the first major update to the HIPAA Security Rule in over 20 years, driven explicitly by AI and ransomware risks. The proposed changes eliminate the distinction between required and addressable safeguards, meaning controls you previously could defer are now mandatory. The final rule is expected this summer.
Then there's Change Healthcare. The 2024 ransomware attack eventually compromised 192.7 million records, nearly two-thirds of the U.S. population. But the breach's real lesson for healthcare administrators wasn't about data exposure. The attack halted eligibility checks, prior authorizations, and claims submissions nationwide for weeks. Providers lost cash flow. Patients couldn't fill prescriptions. It demonstrated a category of third-party AI and technology vendor risk that most organizations hadn't priced into their thinking: when the vendor goes down, care delivery can go with it.
The Emergency Care Research Institute named AI-enabled health technologies the top technology hazard in healthcare for 2025. Not a hazard on the list. The top hazard.
Getting compliant without building a bureaucracy
Start with vendor contracts. Any AI vendor whose system will touch PHI must sign a Business Associate Agreement before your organization transmits a single record. No BAA, no vendor. The agreement also needs to specifically address AI-related data handling: whether the vendor can use your data to train their models, how long they retain PHI after a session, and what their incident response looks like. Most standard BAA templates predate generative AI and don't cover these questions. Ask explicitly, and get the answers in writing.
The shadow AI problem can't be solved by prohibition alone. If you ban all unapproved AI without giving staff a sanctioned alternative for the documentation and coding workflows they're trying to fix, you won't eliminate shadow AI. You'll push it underground, making it harder to see. Censinet recommends maintaining an approved AI catalog, a running list of vetted tools staff can use for specific tasks. That's the practical counter to unauthorized adoption: give people a better option, not just a policy.
On the technical side, verify that your patient data won't end up in a vendor's training dataset. Ask about retrieval-augmented generation (RAG) data controls, prompt guardrails, and data retention after a session closes. Vendors who can't answer these clearly are probably not configured to handle PHI safely, regardless of what their marketing says.
And document your remediation, not just your risk analysis. OCR's enforcement shift is specific: you can no longer show you identified risks and call it done. You need to show what you did about what you found, who owns it, and when it gets fixed. Risk analysis that gathers dust is now a liability, not a compliance checkbox.
The actual decision
Healthcare administrators didn't get into this field to become AI compliance specialists. That's the reality of the job in 2026, regardless.
Gartner projects that 60% of healthcare organizations will face delays in digital transformation due to noncompliance. The organizations that get the governance infrastructure right first will be the ones that can actually deliver on what AI promises: a lighter documentation burden, faster authorizations, and better diagnostics. The ones that move fast without the infrastructure won't avoid those benefits. They'll reach them later, after the breach, or the penalty, or the operational crisis that finally forces the investment.
Most organizations are still figuring this out. That's an opening if you're willing to take it.
At Tristella Advisors, we help healthcare organizations build AI governance frameworks that are actually implementable, not 80-page policy documents nobody reads. If you're trying to get your arms around what AI is running in your organization and whether it's compliant, the conversation starts with an honest inventory.
Reach out at tristellaadvisors.com.
Sources: