AI Agents for Healthcare: HIPAA-Compliant Workflow Automation

Every consequential action in the agent trace is logged to immutable storage (we use AWS QLDB or PostgreSQL with append-only audit tables) keyed to the clinician's NPI on each approval step. When a state medical board reviews a clinical decision the AI participated in, we can produce the full agent trace: which model version drafted, what evidence it retrieved, which clinician approved each step, and at what timestamp. This satisfies the attribution requirement that Texas, Florida, and California medical boards have all published guidance on for AI-augmented clinical workflows.

Most don't, because they keep a clinician in the loop on consequential decisions: this is the FDA's 'augmented' category under the Clinical Decision Support guidance, exempt from premarket SaMD clearance. If you want fully autonomous clinical decision-making (no human review), you're in regulated SaMD territory and clearance becomes part of the engagement (we've supported one client through a 510(k) pathway for an AI triage tool, which adds 12-18 months and $400-800K to the program).

Yes for read access via FHIR R4 (most BearPlex healthcare agents use this path): Epic exposes FHIR endpoints to authenticated apps without requiring App Orchard listing. App Orchard listing becomes important for write access at scale, distribution to other Epic customers, and access to certain non-FHIR APIs. For a single-customer deployment writing back via FHIR DocumentReference (e.g., ambient scribe inserting SOAP notes), Orchard listing is typically not required, but the customer's Epic account team must approve the integration. We've shipped 4 healthcare agents via the non-Orchard path; 2 have since pursued Orchard listing for cross-organization deployment.

Three layers because no single defense is sufficient in clinical context. (1) RAG with citation tracking: every clinical claim must reference a source document (we typically use Anthropic's Citations API when on Claude, custom citation extraction on other models). (2) Constitutional AI-style guardrails specific to the clinical context: refuses to fabricate dosages, lab values, or diagnostic codes; flags uncertainty explicitly. (3) Mandatory clinician review on any output that affects patient care, with the agent's reasoning trace surfaced to the reviewer. Pure prompt-engineering defenses against hallucination aren't sufficient when malpractice liability is at stake: the Mata v. Avianca (S.D.N.Y. 2023) precedent in legal contexts is being actively cited in physician-AI cases.

We sequence them: HIPAA BAA chain validation in weeks 1-2 (covers the basics: covered entity, business associate, subcontractor BAA flow), HITRUST CSF v11.4 evidence collection in weeks 3-6 (the AI Risk Management track requires specific documentation around model selection, prompt evaluation, and inference logging), SOC 2 Type II evidence collection runs continuously throughout the engagement and produces the audit-ready evidence package by week 12. We work with the customer's existing GRC tooling (Vanta, Drata, Tugboat) where present, and ship Markdown evidence dossiers when no GRC platform exists. Roughly 25% of total engagement effort goes to compliance documentation in heavily-regulated deployments.

Ambient scribe ROI is faster but smaller per deployment: 2.4-3.1 hours per physician per day reclaimed (our measurements), translating to 1-2 additional patient visits per day per physician at typical RVU economics. For a 50-physician medical group, that's $1.8-3.6M annual additional revenue against engagement cost of $200-350K. Prior auth automation ROI is slower to realize but larger per organization: cycle time from 14 days to <24 hours reduces denial-related write-offs by 8-15% (per our 3 deployments) and frees 0.5-1.5 FTE per 1,000 monthly PAs. For a 250-bed hospital running ~3,000 monthly PAs, savings land at $400-900K annually against engagement cost of $300-600K. Most clients sequence: ambient scribe first to build physician trust and quick win, prior auth second once governance scaffolding is in place.

Yes. For organizations that can't allow cloud LLM inference (typically VHA, certain Catholic Health systems, and specific state Medicaid integrators), we deploy fine-tuned Llama 3.3 (or Qwen 2.5 for multilingual environments) on the client's on-premise NVIDIA H100 cluster, with ambient scribe / RAG agents running entirely within the facility network. Performance is competitive with frontier models for narrow clinical tasks (we measure within 4-7% on internal SOAP-note quality benchmarks vs Claude Sonnet on the same prompts); engineering effort is meaningfully higher (typically +30-40% engagement scope for the on-prem variant, primarily for inference infrastructure setup and model serving optimization with vLLM or TensorRT-LLM).