Healthcare AI · Multi-Agent Orchestration
Production multi-agent system for patient acquisition and care operations.
Autonomous interactions
500K+
Resolution rate
96%
Inbound deflection
70%
Bot-to-human ratio
1:24
Production error rate
<1%
01 · CONTEXT
Who. What scale. What was at stake.
A regulated digital pharmacy operating at approximately $15M/month in revenue needed to absorb higher inbound patient volume without scaling headcount linearly. Patient acquisition, clinical triage, lifecycle communication, and renewal were separate workflows with separate systems. No shared context traveled between them. Every handoff was a seam. Every seam was a place the patient could fall out.
02 · CONSTRAINT
The architectural problem.
The system had to operate autonomously at scale while remaining PHI-safe, auditable, and clinically accountable. Most off-the-shelf AI vendors treated PHI as a configuration option rather than a system property. The second constraint: human-in-the-loop at every clinical decision. Automation handled volume. Humans owned accountability. Those two had to coexist.
03 · DECISION POINTS
Three decisions that shaped everything downstream.
Decision 01
Infrastructure boundary
Dual-cloud: AWS as system of record, Azure as orchestration layer. PHI never left the controlled boundary. Mixing them would have created a PHI surface that could not be governed cleanly.
Decision 02
Escalation logic
Deterministic escalation rules over LLM judgment. Every clinical handoff followed audit-logged decision logic. The system proposed. Humans confirmed. LLMs do not own accountability. The architecture reflected that.
Decision 03
Identity vs content
Engines operated on identity tokens (harbor_id, patient_id, session_id), not clinical content. Decisions happened without the operational layer reading PHI. Raw clinical data stayed sealed.
04 · SYSTEM
What was built.
Multi-agent orchestration across SMS, voice, chat, telephony, and contact center. Single event bus. All channels, one audit trail.
Dual-cloud PHI architecture. AWS system of record. Azure orchestration layer. PHI de-identified at ingress before touching the operational layer.
Deterministic escalation logic. Clinical handoffs routed through audit-logged decision rules. Human-in-the-loop at every clinical action.
Acquisition engine: autonomous outreach at 20K interactions/day, 38% engagement rate.
Support engine: triage, retention, and renewal running 24/7. 500K+ interactions in the first five months.
Observability layer: resolution rate, error rate, escalation rate, and cost-per-interaction tracked in real time.
05 · OUTCOMES
Five months of production operation. All metrics from live system telemetry.
500K+ autonomous patient interactions in five months of production
96% resolution rate. The remaining 4% escalated cleanly with full context preserved.
70% reduction in inbound human-handled volume
20K outbound interactions per day at 38% engagement rate
1:24 bot-to-human ratio at full production load
Sub-1% production error rate across all agent types
06 · DISCUSS FURTHER
The architecture above is public. What follows is a conversation.
Implementation-level diagrams, vendor selection rationale, team composition, exact cost and savings figures, agent specification, PHI firewall detail, escalation rule logic, and lessons learned are shared in a 30-minute conversation with executives evaluating similar work.