AI in Healthcare

The most-deployed healthcare AI in 2026 is not a diagnostic. It is a stenographer. Ambient documentation tools — Microsoft DAX Copilot (Nuance), Abridge, Suki, and the Epic-integrated scribes — are now used by tens of thousands of physicians daily across the Mayo Clinic, Kaiser Permanente, Stanford Health, Sutter, Atrium, and most of the major US academic medical centers. Permanente Medical Group disclosed in 2024 that it had deployed Abridge to over 10,000 physicians across The Permanente Medical Group region. Mass General Brigham, UPMC, and Emory have publicly named Abridge as their ambient scribe of record. This is the part of healthcare AI that is unambiguously working: it reduces documentation time, it bills cleanly through existing E/M workflow, and the physician signs the note.

The second deployed category is imaging triage. Aidoc, Viz.ai, and RapidAI have FDA-cleared algorithms running in production at over a thousand US hospitals collectively, flagging suspected large-vessel occlusion strokes, pulmonary embolism, and intracranial hemorrhage on CT. These are narrow, FDA-cleared, time-to-treatment-shortening tools that route critical findings to the on-call specialist's phone. Viz.ai's LVO algorithm received the first New Technology Add-on Payment from CMS for AI software (NTAP, effective FY 2021), which created the first real reimbursement precedent for AI in radiology.

The third category that ships: clinical search and evidence summarization. OpenEvidence, founded by Daniel Nadler, is used by a substantial share of US physicians by its own disclosures and Mayo Clinic partnership announcements, providing point-of-care medical literature retrieval grounded in NEJM and JAMA partnerships. UpToDate from Wolters Kluwer is integrating LLM-based summarization on top of its editorially curated base.

Where the sector stalls: autonomous clinical decision-making. There is no FDA-cleared large language model authorized to diagnose without a physician sign-off. Hippocratic AI's “safety-focused” agents are deployed for outbound non-diagnostic patient outreach — appointment reminders, post-discharge check-ins, chronic care follow-up — explicitly avoiding the diagnostic decision surface. Drug discovery AI companies have generated candidates that have reached clinical trials (Insilico's INS018_055 in Phase II for IPF, Recursion's REC-994 in Phase II for cerebral cavernous malformations), but as of the current cycle no drug whose discovery was AI-led has reached FDA approval. The hype line and the deployment line are far apart, and a senior operator must hold them separate.

Autonomous diagnosis. No FDA-cleared LLM is authorized to deliver a final diagnosis without a clinician in the loop. The FDA SaMD pathway treats autonomous diagnostic LLMs as Class II or III devices requiring clinical evidence the field has not yet produced at scale.

End-to-end drug discovery. No drug whose discovery was AI-led has received FDA approval. Multiple candidates are in Phase I and II (Insilico INS018_055, Recursion REC-994, BenevolentAI and Exscientia candidates), with mixed Phase II readouts.

Hallucination-free open-domain clinical Q&A. Even with retrieval-augmented generation, LLM medical Q&A still produces confidently wrong citations and dosage errors at non-zero rates. Every shipping product in this space gates output behind a clinician edit step or restricts scope to non-diagnostic outreach.

Autonomous mental health therapy. No FDA-cleared autonomous psychotherapy LLM exists. Woebot Health and others operate as wellness, not medical devices, and the regulatory line is being actively redrawn.

Real-world data interoperability at scale. Health systems still struggle to extract structured data across Epic, Oracle Health (Cerner), and MEDITECH. TEFCA is rolling out but does not yet deliver the consistent training-data substrate AI vendors need.

Bias-audited deployment at scale. Published audits (Obermeyer et al. 2019 on a commercial risk algorithm) continue to find racial and socioeconomic bias in deployed algorithms. The ONC HTI-1 transparency rule is the first federal requirement that decision-support algorithms in certified EHRs disclose their training data and intended use.

HIPAA governs PHI in all clinical AI. Any vendor processing patient data must execute a Business Associate Agreement. AWS, Azure, and GCP all offer HIPAA-eligible services; OpenAI offers HIPAA-eligible deployments through Azure OpenAI and via direct BAA arrangements disclosed in 2024.

FDA SaMD pathway.The FDA's CDRH maintains a public list of over 1,000 AI/ML-enabled medical devices that have received clearance, the overwhelming majority being radiology and cardiology imaging. The Predetermined Change Control Plan guidance (draft 2023, finalized 2024) is the mechanism by which adaptive models can be updated post-clearance without a new 510(k). No standalone LLM has received SaMD clearance as a diagnostic device as of this writing.

ONC HTI-1. Finalized December 2023. Requires certified EHR developers to disclose information about predictive Decision Support Interventions — source data, intended use, fairness considerations. The first federal transparency requirement for AI inside the EHR.

CMS reimbursement. NTAPs for AI software (Viz.ai LVO as the precedent) and the FY 2024 IPPS rule established the inpatient pathway. Outpatient AI reimbursement remains case-by-case via CPT category III codes.

FTC enforcement.The FTC's 2023 settlement with BetterHelp ($7.8M) and 2023 order against GoodRx ($1.5M) over health-data sharing with advertisers set the floor for health-adjacent AI on consumer data practices.

GDPR / EU AI Act. The EU AI Act (in force 2024, full applicability through 2026–2027) classifies most clinical decision support as high-risk AI with conformity assessment, data governance, and post-market monitoring obligations. EU-deployed healthcare AI now needs both MDR/IVDR and AI Act compliance.

For a provider CIO/CMIO: Pick one ambient scribe (Abridge, DAX Copilot, or Suki) and pilot in two specialties — ideally one primary care and one specialty with high after-hours documentation load (oncology, cardiology). Measure documentation time, after-hours EHR access, and patient throughput. Require HTI-1 disclosure, BAA, SOC 2 Type II, and a sample bias audit if the tool touches clinical decision content. Do not sign enterprise-wide before single-specialty data clears.

For a payer: Evaluate AI for prior authorization automation cautiously. The UnitedHealth nH Predict controversy and the Cigna PXDX class action are the regulatory backdrop. Any auto-denial pipeline carries real litigation surface. Focus AI investment instead on member outreach (Hippocratic-style non-diagnostic agents), call center augmentation, and claims adjudication where the human remains the decision-maker.

For a pharma operator: Run a discovery partnership with one of the named AI platforms (Isomorphic Labs, Recursion, Insilico, Iambic) on a narrow target where you control the chemistry validation. Do not outsource the full pipeline. AlphaFold is table stakes; the differentiated capability is in the proprietary chemistry and assay loop, not in the structure prediction itself.

For all three: Stand up an internal AI governance committee with clinical, legal, IT security, and compliance representation before deployment scales. The 2025 enforcement cycle will surface real penalties for vendors and deployers that ignored HTI-1, the AI Act, or HIPAA BAA discipline. The cheap move is to be the operator that already documented its governance when the audit arrives.

• FDA's 'Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices' public list (fda.gov) — over 1,000 cleared devices, overwhelmingly imaging and cardiology.
• ONC HTI-1 Final Rule, Federal Register, December 13, 2023, requiring DSI transparency in certified EHR technology.
• Permanente Medical Group public statements (2024) on Abridge deployment; Abridge press releases naming Kaiser, Mass General Brigham, Emory, UPMC, Christus, Yale New Haven.
• Tempus AI Form S-1 and subsequent 10-Q filings with the SEC following its June 2024 IPO (NASDAQ: TEM).
• Recursion Pharmaceuticals SEC filings (NASDAQ: RXRX) and the November 2024 closing of the Exscientia merger.
• Microsoft's 2022 closing of the Nuance acquisition ($19.7B, disclosed in Microsoft 10-K filings).
• Isomorphic Labs press releases (January 2024) announcing strategic research collaborations with Novartis and Eli Lilly.
• JAMA Network Open, Ayers et al. and UCSD AI-drafted patient message studies (2023–2024).
• CMS FY 2021 IPPS Final Rule establishing the first NTAP for AI software (Viz.ai LVO).
• FTC settlements with BetterHelp (2023, $7.8M) and GoodRx (2023, $1.5M) — Federal Register and FTC press releases.

• FDA Center for Devices and Radiological Health, 'Artificial Intelligence and Machine Learning in Software as a Medical Device' guidance documents.
• ONC HTI-1 Final Rule, Federal Register, December 2023.
• Obermeyer, Powers, Vogeli, Mullainathan, 'Dissecting racial bias in an algorithm used to manage the health of populations,' Science 2019.
• Coalition for Health AI (CHAI) 'Blueprint for Trustworthy AI Implementation Guidance and Assurance for Healthcare,' 2023 and subsequent revisions.
• AMA 'Augmented Intelligence in Health Care' policy and physician-survey reports (2023–2024).
• The Permanente Medical Group and Kaiser Permanente published reports on ambient scribe deployment outcomes.
• DeepMind / Isomorphic Labs AlphaFold 3 paper in Nature, May 2024.
• EU AI Act, Regulation (EU) 2024/1689, Official Journal of the European Union.