Health systems moving from pilot AI agents to production face twin challenges: maintaining clinical trust and managing the cost of each AI query, according to leaders at Advocate Health and ECU Health.
Advocate Health, a multi-state system running two of the largest Epic EHR instances globally, has deployed autonomous AI agents for inpatient pharmacy workflows and infusion chart preparation using Epic’s no-code Agent Factory platform. The health system enforces a strict human-in-the-loop rule for every agent action, meaning the AI cannot order tests, make diagnoses, or execute clinical tasks without clinician approval.
“I don’t allow it to order, I don’t allow it to diagnose,” said Andy Crowder, Advocate’s chief data and AI officer. Real-time exception logging lets clinicians flag agent outputs they chose not to follow, and the system disables underperforming agents automatically. Advocate is also planning AI literacy training across its 162,000 employees.
ECU Health, a smaller rural academic network in North Carolina, deployed two Epic-built agent prototypes focused on patient transfers and discharges. Being an early Agent Factory pilot partner gave the organization its first experience with custom AI despite a lean IT team.
Crowder warned that token costs — the per-query fees charged by large language model providers — are a real constraint health systems must architect for. Without active monitoring, agent-driven token consumption can outpace budgets quickly. Both organizations stressed that data governance and workflow design must precede any agent deployment, and that organizations chasing AI out of fear of missing out risk building tools they cannot sustain.
