AI in Healthcare
Healthcare AI: Constrained but High-Value
Healthcare is one of the highest-stakes domains for AI deployment. The combination of life-safety implications, strict regulation (HIPAA, FDA for clinical decision support), and complex liability structures means AI adoption moves more carefully here than in other industries.
The high-value use cases are real — but so are the constraints, and effective healthcare AI deployment requires understanding both.
High-Value Use Cases
Clinical Documentation
Healthcare providers spend significant time on documentation — clinical notes, discharge summaries, referral letters, prior authorization requests. AI can substantially reduce this burden.
Ambient documentation (AI scribe) systems transcribe patient-clinician conversations and generate structured clinical notes. Clinicians review and edit rather than compose from scratch. Vendor-reported time savings range from 10–20 minutes per encounter in controlled pilots; a large-scale 2026 study across 400+ clinicians found a more modest 13-minute daily reduction in total EHR documentation time.1 Results vary significantly by specialty, workflow integration, and documentation complexity.
Structured note generation: given clinical inputs, AI drafts notes in required formats (SOAP, H&P, DAP). Clinicians verify accuracy.
Critical constraint: the clinician must review and attest to AI-generated clinical documentation. AI does not make the clinical judgment; it drafts documentation of a judgment the clinician has already made.
Medical Literature Review and Synthesis
Reviewing large bodies of literature for evidence-based updates, clinical research, or specific clinical questions is time-consuming synthesis work.
AI tools:
- Summarize key findings across a set of papers
- Extract specific data points (study design, sample size, outcomes) from abstracts
- Surface relevant references given a clinical question
Limitation: AI summarizes what's in the papers provided. Clinical interpretation of evidence quality and applicability to specific patient populations requires domain expertise.
Context on current capability: frontier LLMs now score above the average passing threshold on the US Medical Licensing Examination (USMLE), and several exceed the average score of first-year residents.2 This reflects genuine strength on standardized test-style clinical questions — not clinical reasoning in complex, ambiguous, or rare presentations. Benchmark performance and clinical competence are measuring different things.
Patient Communication
Drafting patient-facing communications — care instructions, appointment reminders, educational materials — at appropriate reading levels is a task AI handles consistently. High-volume, high-consistency work.
Health literacy: AI can be prompted to write at specific reading levels. This is reliably more accurate than asking humans to self-calibrate their writing complexity.
Translation: drafting communications in multiple languages with consistent accuracy.
Administrative Workflows
Prior authorization requests, insurance correspondence, referral letters, and compliance documentation are structured writing tasks that AI accelerates. The clinician or administrator provides the clinical facts; AI drafts the required document format.
Healthcare-Specific Constraints
HIPAA and data privacy: patient health information (PHI) has strict handling requirements. Before using any AI tool with patient data, verify it meets HIPAA Business Associate Agreement requirements. Consumer-facing AI tools are generally not appropriate for PHI. The finalized HIPAA Security Rule (May 2026) eliminates the "addressable vs. required" safeguard distinction and requires annual risk assessments that explicitly enumerate all AI systems touching PHI.3 New mandatory controls include encryption of all ePHI at rest and in transit, multi-factor authentication, vulnerability scanning, asset inventory, network segmentation, and 72-hour breach incident reporting. Healthcare organizations using ambient scribes, clinical decision support, or any AI tool processing patient data must include those systems in formal HIPAA risk analysis documentation — not just in vendor contracts.
AI training data: The finalized HIPAA Security Rule extends PHI protections to ePHI used in AI training data.4 Healthcare organizations cannot use patient records to fine-tune or train AI models through third-party providers without a HIPAA Business Associate Agreement and appropriate PHI handling controls — even if the model output will never contain identifiable patient information.
FDA clinical decision support: AI tools that analyze patient-specific data to provide clinical recommendations may be regulated as Software as a Medical Device (SaMD). This is a real regulatory risk for clinical deployment that requires counsel.
Liability: who is liable when AI-assisted documentation contributes to a clinical error is unsettled law. Involve institutional risk management and legal counsel in clinical AI deployment.
Hallucination risk: AI hallucinations in clinical contexts are patient safety issues. A study in NEJM AI (Soroush et al., 2024) found LLM accuracy on medical coding tasks below 50% across all tested models — GPT-4 reached 45.9% on ICD-9-CM and 33.9% on ICD-10-CM codes. A study testing six LLMs against clinical vignettes containing deliberately planted misinformation (Omar et al., 2025) found a 64.1% hallucination rate under adversarial conditions specifically designed to elicit errors. Both studies represent challenging edge cases; structured tasks like ambient clinical documentation — where AI transcribes and organizes spoken content — report much better reliability and consistent time savings of 10–20 minutes per encounter. Any AI-generated clinical content requires review by a qualified clinician before use.
Healthcare AI adoption is most sustainable when it starts with administrative and documentation workflows where a human reviews all AI output before it affects patient care. Build organizational AI literacy and governance structures there before expanding toward clinical decision support.