CASE STUDY / HEALTHCARE & LIFE SCIENCES
Giving 2,400 physicians their evenings back with ambient AI
HIPAA-compliant ambient clinical documentation integrated with Epic — 38% less after-hours charting and adoption by 87% of eligible physicians in eight months.
38%
reduction in after-hours documentation time
87%
adoption among eligible physicians in eight months
3.6 min
average physician edit time per drafted note
+19pt
improvement in documentation-burden satisfaction scores
THE CLIENT
Context
The client is an integrated delivery network: 14 hospitals, 300+ ambulatory clinics, 2,400 employed physicians on a single Epic instance. Its own burnout survey put documentation burden as the top driver of physician attrition, with clinicians averaging 1.9 hours of EHR time after clinic hours per day — 'pajama time' the CMO had committed to the board to cut in half.
A vendor ambient-scribe pilot the previous year had stalled: transcription quality was acceptable, but notes arrived as unstructured blobs that physicians spent nearly as long restructuring as they had spent typing, and the tool lived outside Epic, forcing constant window-switching.
THE CHALLENGE
What was at stake
The bar was precise: draft notes had to arrive inside the Epic workflow, structured to each specialty's documentation patterns, accurate enough that editing took minutes rather than re-writing — and the entire pipeline had to satisfy the health system's privacy office, from BAA coverage to audit trails on every transcript and inference.
Specialty variation was the hidden difficulty. An orthopedic follow-up, a primary-care visit with six chronic conditions, and a behavioral-health session need fundamentally different note structures, terminology handling, and sensitivity rules. A one-size-fits-all summarizer would fail exactly the way the previous pilot had.
Engagement at a glance
- Client
- A 14-hospital integrated health system
- Region
- United States
- Duration
- 11 months
- Team
- 10-person team: AI engineers, FHIR integration, clinical informaticist, compliance lead
Services applied
THE SOLUTION
What we built
We built an ambient documentation pipeline deployed entirely inside the health system's Azure tenancy under its BAA. Consented visit audio streams to a medical-tuned speech-to-text layer, then through a multi-stage LLM pipeline: clinical entity extraction, specialty-specific note assembly against templates co-designed with each department, and a verification pass that flags low-confidence content for physician attention rather than hiding uncertainty.
Draft notes land directly in Epic via FHIR and Epic's native integration points as pended documentation — physicians review, edit, and sign inside the workflow they already know. Every draft carries provenance: tap any sentence and see the transcript segment it came from, which proved decisive for physician trust and for the privacy office's audit requirements.
Rollout ran specialty by specialty, starting with primary care and orthopedics. Each specialty got a two-week template co-design phase, a 20-physician pilot with side-by-side quality scoring against clinician-authored notes, and go/no-go review by the medical-informatics committee before expansion.
// ARCHITECTURE
All PHI processing stays inside the client's Azure Health-grade tenancy: streaming transcription, Azure OpenAI models under BAA, and a LangGraph-orchestrated note-assembly pipeline with per-specialty configuration. FHIR R4 and HL7v2 interfaces handle Epic integration; consent state is checked at the point of recording. Immutable audit logs capture every transcript, inference, and edit.
A quality-monitoring service samples signed notes against drafts to measure edit distance and section-level accuracy per specialty, feeding the monthly clinical-safety review.
Core stack
- Azure OpenAI (BAA)
- FHIR R4
- HL7v2
- Epic integration
- LangGraph
- Whisper (medical-tuned)
- Azure Health Data Services
- PostgreSQL
- Terraform
HOW IT WAS DELIVERED
Implementation approach
Value delivered in phases with go/no-go evidence at each gate — never a big-bang bet.
- 01
Compliance architecture (months 1–2)
Designed the PHI boundary, consent flow, and audit model with the privacy and security offices; signed off before any audio was processed.
- 02
Primary-care pilot (months 3–5)
Co-designed templates with 20 pilot physicians; measured note quality against clinician-authored ground truth until edit time fell under four minutes per note.
- 03
Specialty expansion (months 5–9)
Rolled out to orthopedics, cardiology, and five further specialties, each with template co-design and committee go/no-go gates.
- 04
Scale and monitoring (months 9–11)
Opened enrollment system-wide with self-service onboarding, drift monitoring, and monthly clinical-safety reporting.
“Our physicians don't describe it as AI — they describe it as getting home for dinner. The provenance feature was the turning point: doctors could verify every sentence, so they actually trusted the drafts. That's why adoption is at 87% and climbing.”
Dr. Elena Vasquez
Chief Medical Information Officer — 14-hospital integrated health system
KEEP READING
More case studies
Want results like these?
Bring us the metric you need to move. Our architects will map how a comparable engagement would work in your environment — systems, timeline, and expected impact.