Agentic AI for Healthcare

Agentic AI is reshaping healthcare from a system built on human-operated software into one where autonomous agents handle clinical documentation, orchestrate care coordination, accelerate drug discovery, and manage the vast administrative machinery that consumes over 30% of U.S. healthcare spending. Unlike earlier waves of healthcare AI that focused on narrow tasks—reading a radiology scan, flagging a lab anomaly—agentic systems operate in sustained loops: observing patient data streams, reasoning about clinical context, taking action across EHR systems and payer portals, and learning from outcomes. The result is a fundamental shift from AI as a decision-support overlay to AI as an autonomous participant in the care delivery chain.

Ambient Clinical Intelligence: From Documentation to Autonomous Scribes

The most widely deployed agentic AI systems in healthcare today are ambient clinical documentation agents. Microsoft's Nuance DAX Copilot, now deeply integrated with the Microsoft Azure health cloud, operates as an always-on agent during patient encounters—listening to physician-patient conversations, extracting structured clinical data, generating SOAP notes, populating EHR fields, and even drafting referral letters and prior authorization requests. By early 2026, DAX Copilot is deployed across over 200 health systems including UW Health, Atrium Health, and the University of Michigan. Physicians using the system report saving 40–90 minutes per day, with documentation accuracy rates exceeding 90% on first draft.

Abridge, which raised $212.5 million in Series C funding in late 2024, has become the leading independent ambient documentation platform, partnering with Epic Systems for native EHR integration. Deployed at UPMC, UCI Health, Emory Healthcare, and dozens of community health systems, Abridge's agent doesn't just transcribe—it structures conversations into clinically coded notes, identifies gaps in care plans, and flags discrepancies between what was discussed and what was ordered. Nabla and DeepScribe compete in this space with differentiated approaches to specialty-specific documentation.

What makes these systems genuinely agentic rather than mere transcription tools is their operational loop. DAX Copilot and Abridge agents persist across the patient encounter lifecycle: they pre-populate visit summaries from prior records, listen and document during the visit, generate after-visit summaries for patients, and update care plans—all without the physician explicitly commanding each step. This is the ambient intelligence paradigm: the agent is always working in the background, reducing the cognitive burden that drives physician burnout.

Revenue Cycle and Administrative Automation

Healthcare's administrative complexity—prior authorizations, claims processing, benefits verification, denial management—is perhaps the single largest opportunity for agentic AI. The American Medical Association estimates that prior authorization alone costs the U.S. healthcare system $35 billion annually in administrative overhead.

Infinitus Health deploys AI voice agents that autonomously call insurance companies to verify patient benefits and obtain prior authorizations. The agent navigates phone trees, waits on hold, communicates with payer representatives, and extracts structured coverage information—tasks that previously required dedicated staff spending 45–60 minutes per call. Infinitus reports handling millions of calls and reducing average call resolution time by over 70%.

Akasa, backed by $60 million in funding from BOND Capital and others, applies large language models to the full revenue cycle: coding, billing, claim submission, denial prediction, and appeal generation. Their agentic system doesn't just flag issues—it autonomously resolves claim denials by analyzing denial codes, pulling supporting clinical documentation, drafting appeal letters, and resubmitting claims. Waystar's AI platform similarly automates claims lifecycle management at scale across thousands of provider organizations.

Cedar and Notable Health focus on the patient-facing side of administrative automation, deploying agents that handle intake forms, insurance verification, payment plan setup, and post-visit billing inquiries without human staff involvement. These aren't chatbots waiting for questions—they proactively reach out to patients before appointments to collect information and resolve coverage issues.

Drug Discovery and Biomedical Research Agents

Drug discovery represents the highest-stakes application of agentic AI in healthcare. The traditional drug development pipeline takes 10–15 years and costs $2.6 billion per approved drug. Agentic AI systems are compressing both timelines and costs by autonomously navigating the hypothesis-generation, compound design, and preclinical optimization phases.

Isomorphic Labs, the Alphabet/DeepMind spinoff, signed partnerships with Eli Lilly and Novartis worth up to $3 billion to apply AI-driven molecular design to therapeutic targets. Building on the AlphaFold breakthrough in protein structure prediction, Isomorphic's agents can autonomously propose novel molecular structures, predict binding affinities, assess toxicity profiles, and iterate on designs—compressing what was once months of medicinal chemistry into days.

Recursion Pharmaceuticals has built what they call a "Recursion OS"—an integrated platform where AI agents operate across the full discovery pipeline. Their partnership with NVIDIA combines Recursion's massive biological dataset (tens of petabytes of cellular imagery) with NVIDIA's BioNeMo platform for training biological foundation models. Recursion's agents autonomously design experiments, analyze high-throughput screening results, and identify drug candidates for rare diseases and oncology targets.

Insilico Medicine's AI-discovered drug INS018_055 for idiopathic pulmonary fibrosis advanced to Phase II clinical trials—one of the first AI-originated compounds to reach this milestone. Their Pharma.AI platform uses multi-agent systems where separate agents handle target identification, molecule generation, and clinical trial prediction, coordinating through shared knowledge graphs. Xaira Therapeutics launched in 2024 with $1 billion in funding to build a fully AI-native drug discovery platform, with former Google DeepMind researchers leading the effort.

Clinical Decision Support and Care Coordination Agents

Google's Med-Gemini models, extensions of the Gemini multimodal architecture optimized for medical reasoning, have demonstrated expert-level performance on clinical benchmarks including USMLE, medical visual question answering, and long-context clinical reasoning tasks. While not yet deployed as autonomous clinical agents due to regulatory constraints, Med-Gemini powers agentic research assistants at HCA Healthcare and Mayo Clinic that autonomously search medical literature, synthesize patient histories, and generate differential diagnosis workups for physician review.

Epic Systems, the dominant U.S. EHR vendor, has integrated agentic AI capabilities directly into its platform through partnerships with Microsoft and its own in-house development. Epic's AI agents can autonomously triage patient messages in the physician inbox, draft responses to routine clinical questions, and flag messages requiring urgent physician attention—reducing inbox burden by an estimated 30–40% at deployed sites.

Hippocratic AI, which raised $120 million in its Series A, is building safety-focused AI agents specifically for non-diagnostic patient interactions: post-discharge follow-up calls, chronic disease check-ins, medication adherence monitoring, and pre-operative preparation. Their agents conduct extended voice conversations with patients, following clinical protocols while adapting to patient responses, and escalate to human clinicians when they detect concerning symptoms or emotional distress. The company's approach of targeting non-diagnostic use cases is a deliberate strategy to navigate the regulatory landscape while delivering immediate value.

The Regulatory and Safety Landscape

The FDA had authorized over 1,000 AI/ML-enabled medical devices by early 2026, but the vast majority are narrow, deterministic systems—primarily in radiology and cardiology. Autonomous agentic LLM-based systems occupy a regulatory gray zone. The FDA's evolving framework distinguishes between Clinical Decision Support (CDS) that merely informs clinicians and autonomous systems that take action, with the latter facing significantly higher regulatory bars.

HIPAA compliance adds another layer of complexity. When an agentic AI system processes patient conversations, navigates EHR data, and communicates with payer systems, every data flow must be covered by Business Associate Agreements. Cloud-based inference—where patient data leaves the hospital network for processing by Anthropic, OpenAI, or Google models—requires careful architectural decisions about data residency, encryption, and audit logging.

The EU AI Act, which took effect in phases through 2025–2026, classifies most healthcare AI as "high-risk," imposing requirements for transparency, human oversight, robustness testing, and conformity assessments. For agentic systems that operate autonomously over extended periods, meeting these requirements demands comprehensive logging of every decision the agent makes—a technical challenge that has driven the emergence of specialized AI observability platforms for healthcare.