AI adoption is moving faster than the legal guardrails in physician practices
Physician groups are using AI tools to draft visit notes, manage scheduling, and process prior authorizations, but every workflow shift carries privacy and liability consequences once the software can touch protected health information. HIPAA applies to those systems in the same way it applies to every other technology that handles patient data, and the practical exposure often begins long before a lawsuit or regulatory inquiry appears. A contract that looks routine on paper can become the wrong place to discover who owns the risk when data persist after termination, deletion is incomplete, or a vendor’s promises are narrower than the practice assumed.
Tatiana Melnik, J.D., a Tampa-based health care attorney who represents physician practices, framed the issue around a basic operational question: what exactly happens to patient data when the relationship with an AI vendor ends? Her warning is straightforward. If a vendor keeps data after the contract expires and the damages cap is built around fees paid during the prior year, a practice can be left with the bulk of the liability when a problem surfaces later.
Contract language around data retention now carries more weight
Before the spread of generative AI, allowing a vendor to de-identify patient data could have looked like a low-risk business decision. Re-identification is far easier now, and that changes how practices should read the contract. Clauses that sound administrative, such as “usage data,” can create real ambiguity if an AI prompt is swept into a category originally meant to describe software clicks used for troubleshooting. “Confidential information” deserves the same scrutiny, because broad or vague wording can either limit the practice’s control or fail to define what the vendor may keep, reuse, or delete.
Deletion rights deserve special attention. If a vendor says it can delete data but cannot actually do so in a verifiable way, the agreement may offer only the appearance of control. A practice also needs to know whether the vendor may retain data after termination, because the legal consequences of retaining access are not theoretical. When the damages cap is tied to a short payment window and the issue surfaces years later, the cap can fall to zero, leaving the covered entity exposed under HIPAA while the vendor’s financial responsibility evaporates.
An empty damages cap is not a shield; it is a warning sign.
AI scribes create consent problems and a new malpractice argument
AI scribes are already changing the rhythm of the exam room, but they are also creating a fresh consent problem. California patients have filed consumer class actions against practices and hospital systems using AI scribes, arguing that they never agreed to that use of their information. The legal exposure does not stop at privacy claims. Melnik expects malpractice carriers to begin pushing back in cases where clinicians rely on AI-generated notes without sufficient review, especially if a physician signs off on documentation that was not carefully checked.
The likely carrier argument is familiar to anyone who has watched health care disputes migrate from charting to liability coverage: if the mistake came from technology rather than from a clinical judgment failure, the insurer may try to recast the event as a technology problem rather than a malpractice claim. Electronic health records provide a useful analogy. Audit trails make it easier for plaintiff attorneys to ask who edited a note, when it was edited, and whether the record reflects the visit or a later reconstruction of it. AI scribes add another layer to that same scrutiny because the note may begin as machine-generated text and then move through a clinician’s hands without enough visible accountability.
Consent and disclosure should be built into existing privacy workflows
A practice does not need to invent an entirely new compliance structure to disclose AI use. The notice of privacy practices already exists, and it can carry AI language if the practice decides to add it. Patients should receive affirmative consent, and the practice should have a process ready for those who decline. In some settings, the workaround may be simple: turn the tool off for that encounter. In others, a practice may decide that a refusal means the patient cannot be seen there, just as some organizations handled resistance when electronic records first became standard.
Staff training matters because the disclosure process cannot be improvised in the middle of a busy clinic day. Front-desk teams, medical assistants, and clinicians need the same script and the same thresholds for escalation. Patients who arrive after asking a chatbot about symptoms should be managed through the same clinical process used for people who have already been searching the internet for answers. A chatbot may shape expectations, but it does not replace history-taking, triage, or a clinician’s review of risk.
Human review remains the dividing line between convenience and liability
AI can accelerate documentation and administrative work, but it cannot own the clinical record. If a physician does not review an AI scribe’s note, the practice creates a vulnerable record that can be attacked from two directions: privacy consent and documentation reliability. The more a workflow depends on automation without visible oversight, the easier it becomes for a plaintiff or payer to argue that the practice accepted the output without exercising the standard review that patient care requires.
Melnik’s advice to physician groups is to inventory existing and new tools, understand what patient permissions already exist, and read the fine print as if a dispute will eventually test every clause. That approach is not about slowing adoption for its own sake. It is about recognizing that AI has entered the same legal environment as every other system handling protected health information, only with more speed, more data movement, and more uncertainty about retention. Practices that assume the contract resolves those questions may discover too late that liability stayed behind after the vendor relationship ended.
Technology is coming into the exam room whether practices are ready or not. The deciding factor is whether the legal and operational safeguards arrive first, or after the first complaint, audit trail, or denied claim forces the issue.