Governing GenAI in Healthcare Operations: A Compliance-First Playbook

Ambient note-takers (microphone-equipped systems that capture clinician-patient conversations and produce structured notes) are the highest-volume GenAI deployment in US healthcare today and the most regulated. The workflow handles raw PHI in audio and transcript form, produces structured PHI in note form, writes to the EHR, and increasingly drives downstream coding and billing.

Required controls. Business Associate Agreement that explicitly permits the audio processing use case, names the subprocessors, and binds the vendor to HIPAA Security Rule expectations (encryption at rest and in transit per 45 CFR 164.312, access controls, audit logging). Patient notification at the point of care with a meaningful opt-out, consistent with HHS guidance on AI in healthcare and state-level expectations (California, Texas, and New York have all signalled specific positions on consent). Human review of the generated note before sign-off, with the clinician's signature carrying the same authority as on a self-authored note. Provenance metadata captured for each AI-generated section so that downstream coding, billing, and quality reporting can be reconciled to the underlying audio. Quarterly accuracy audit by sampled note review, with the sample stratified by specialty and patient population.

The trap to avoid. Treating the ambient system as a transcription tool when it is in fact a clinical decision support input. The note that goes into the EHR becomes the source of truth for downstream care, billing, and quality measurement, and an inaccurate AI-generated section can propagate for years. The Areebi healthcare solution binds the ambient workflow to the inventory and the policy engine so that every note carries provenance and every model version is auditable.

GenAI in claims and coding sits between revenue-cycle operations and clinical care. The most common deployment is a coding assistant that proposes ICD-10 and CPT codes from clinical documentation, with a human coder reviewing and accepting or modifying each proposal. Adjacent deployments include prior authorisation summarisation for utilisation management nurses and denial appeal drafting.

Required controls. BAA covering the specific claims data flow, with attention to the secondary uses the vendor may make of de-identified or aggregated outputs. Human-in-the-loop on every coded claim before submission, with the coder bearing the same accountability as on a manually coded claim. Audit log per claim showing which codes were AI-proposed and which were human-amended, retained for the seven-year HIPAA records-retention window. Anti-upcoding controls including a periodic statistical comparison of AI-influenced coding distributions against the organisation's historical baseline. Patient communication path for any denial or coverage decision that was materially shaped by AI output, even when the formal decision was human-made.

The trap to avoid. Coding assistants drift toward higher-value codes over time as the model is fine-tuned on the organisation's historical data, which is itself a feedback loop. Without an upcoding-distribution check, this can produce a False Claims Act exposure that is detectable only by external audit. The HIPAA compliance hub covers the relevant controls in depth.

GenAI-mediated patient communication covers chatbot triage, appointment reminders, post-visit summaries, and increasingly the patient-portal message draft assistant. Each of these touches PHI, each shapes patient understanding and behaviour, and each is subject to the HIPAA Privacy Rule expectations on disclosure and minimum necessary use.

Required controls. BAA with the model vendor covering patient-directed communication, including specific provisions on retention and training. Reading-level and language-access requirements consistent with Section 1557 of the Affordable Care Act (Office for Civil Rights enforcement on language access has intensified through 2025). Clinician sign-off on any communication that contains clinical instruction or that responds to a clinical question. Patient-facing disclosure that the message was AI-assisted, where state law (notably California AB 3030 and similar bills) requires it. Audit log of every patient-directed message with the model version and prompt template that produced it.

The trap to avoid. Patient-portal message drafting that suggests language a clinician then rubber-stamps creates an automation-bias risk where the clinician accepts AI-suggested clinical content they would not have authored independently. AMA AI guidance and the state medical board positions emerging through 2025 both flag this as a discipline-of-medicine issue. The control is a clinician-review interface that surfaces the AI-suggested text alongside the underlying chart context, so that the clinician can verify before signing.

GenAI in research and quality improvement touches PHI under different rules than clinical care. HIPAA permits research use of PHI under several pathways (waiver of authorisation by an IRB, limited data set with a Data Use Agreement, de-identification under the Safe Harbor or Expert Determination method), and each pathway has its own implications for GenAI vendors.

Required controls. Pathway-specific BAA or DUA aligned to the research design. Re-identification risk assessment for any GenAI output that combines de-identified records (the combination of fields a generative model can produce may exceed the de-identification threshold even when each individual field is below it). IRB review of any GenAI use that goes beyond a documented quality improvement scope into human-subjects research. Audit log of every research query, the model version, and the data set version. Output review for any GenAI-produced finding that will be cited in a publication or a regulatory filing.

The trap to avoid. Treating GenAI as a black-box tool in research workflows. The reproducibility expectations of clinical research require that the model version, the data set version, and the prompt strategy be documented to the same standard as a traditional statistical method. The NIST AI RMF MAP function deep dive covers the capability-documentation discipline that operationalises this in MAP 3.

GenAI in administrative operations covers HR, scheduling, supply chain, marketing, and the wider non-clinical surface. The dominant risk pattern is incidental PHI exposure (a marketing dataset that contains patient identifiers, a scheduling system whose chatbot leaks visit reasons) rather than primary PHI processing.

Required controls. Inventory of every administrative system that touches the EHR or any system containing PHI, with the data-flow documented and the GenAI components labelled. Data classification policy that prevents PHI from being pasted into non-BAA-covered tools, enforced by the AI control plane at the prompt boundary. Vendor risk assessment for every administrative GenAI vendor, including the model subprocessors. Training programme for administrative staff on what counts as PHI under HIPAA, refreshed annually. Incident response playbook covering the specific case of PHI leaking into a non-BAA-covered model.

The trap to avoid. Administrative GenAI is often deployed under SaaS procurement processes that do not surface the BAA requirement. By the time the gap is detected, employees have routinely been pasting patient information into public chatbots for months. The shadow AI primer covers the broader pattern, and the cost of one shadow-AI breach walks through the financial exposure.