Healthcare organizations are treating agentic AI as another software rollout, writing policies and forming committees that assume standard IT guardrails will keep operations safe. But agentic AI behaves differently than any previous digital health tool, and clinging to the old playbook creates massive unseen risks, according to Julia Zarb, founder of Blue x Blue, a company building AI infrastructure with governance embedded. Speaking at the eHealth26 conference in Halifax, Nova Scotia, she warned that the widely accepted "human-in-the-loop" safety net is actually a liability trap.
The human-in-the-loop fallacy
Executives are often reassured that a human will review AI outputs before finalizing a clinical decision. In practice, those reviewers lack the time and system explainability to catch microscopic errors under high stress. "What we're doing is transferring the liability that comes with that review," Zarb said. "We don't have what's called explainability to go back in and see how that got to it, who knew what, when, and where, and why?"
Invisible errors that multiply
AI models fail in tiny, logical steps that easily bypass human reviewers. Agentic AI processes information by carrying forward what Zarb calls "micro fissures" - small gap-fillers that look perfectly logical to the system. "What makes agentic AI good is also its internal flaw," she explained. "The way that it processes, it carries forward little micro fissures and flaws." Standard IT guardrails fail because AI naturally jumps boundaries. When multiple AI agents interact, a single logical error propagates until it results in a much larger failure. "If 100 agents speak to 100 other agents and those mistakes are brought forward, these aren't mistakes that are just errors, these are mistakes that make sense [to AI at the time] because of the way the AI is built."
Why this matters for healthcare leaders
Relying on "human-in-the-loop" as a safety default shifts legal and operational liability onto clinicians who lack the tools to audit AI decisions. The push to deploy AI for Healthcare must be matched with governance structures designed for agentic systems. Healthcare IT leaders need to insist on explainability, traceability, and limits on agent-to-agent communication before an invisible error cascades into patient harm.
Your membership also unlocks:
