Deploying AI in Healthcare

What It Covers

Nikhil Buduma, CEO of Ambiance Healthcare, discusses deploying AI in clinical settings with a16z's Julie Yu. Ambiance reaches 75% daily clinician adoption at major academic medical centers, with one health system projecting $30M in net new margin from the platform across documentation, coding, and revenue cycle use cases.

Key Questions Answered

• •Clinical AI adoption threshold: Vendors claiming AI capabilities routinely achieve only 20-30% clinician adoption at health systems, making 75%+ daily utilization the true benchmark for enterprise viability. Health system CEOs should demand utilization data before purchasing, since low adoption renders any downstream ROI claims meaningless regardless of how compelling the vendor pitch sounds.
• •AI product clock speed vs. AI model clock speed: Building healthcare AI requires separating foundation model iteration cycles from product deployment cycles. Ambiance maintains a dedicated infrastructure layer that pulls and normalizes data from any EHR instance, reducing the incremental cost of launching each new use case and enabling the company to scale from 2 to 24 products without rebuilding integrations from scratch each time.
• •Decision traces as the critical data asset: Most EHRs use mutable data structures that destroy clinical decision traces — the sequential record of how a clinician reasoned toward a diagnosis or code. Companies building durable healthcare AI must architect systems that capture immutable decision traces from the start, since this data is what enables domain-specific post-training that generalist foundation models cannot replicate.
• •Revenue cycle as the primary financial ROI driver: Early AI scribe adoption was justified primarily by clinician retention and satisfaction, not hard financials. The shift to measurable ROI now comes from linking ambient documentation directly to coding accuracy, CDI query prevention, denial reduction, and net cash collected — requiring vendors to download EHR data warehouses and build CFO-grade attribution analytics to prove the margin impact.
• •Pre-visit and post-visit agents as the next expansion layer: The near-term product frontier beyond ambient documentation involves deploying agents that gather patient context before appointments and conduct structured follow-up after visits — confirming medication pickup, lab completion, and pre-procedure preparation. This virtual care team model expands clinician panel capacity without increasing physician workload or requiring additional headcount.