Agentic AI systems are already operating at national scale in European health systems, handling scheduling, triage, and consent without human intervention at each step, speakers at the Royal College of Radiologists' 2nd Annual Global AI Conference reported. The rapid deployment of these systems is rewriting how health services build software and assign clinical responsibility, with some organizations now generating 100% of their code through agents.
National-scale deployments are live
Dr. Rowland Illing, chief medical officer at Amazon Web Services and a former interventional radiologist, told the audience that agentic AI "is already running at national scale." In Finland, multiple counties use BeeHealthy, a citizen-facing platform that manages scheduling, identification, and consent without a human at any step. In Portugal, the UpHill system has triaged more than 40,000 patients through the national health hotline, with nurses reviewing the agent's decisions to catch errors.
"Every agent should be paired with a human," said Dr. Hatim Abdulhussein, chief executive officer of Health Innovation Kent Surrey Sussex, part of the NHS Health Innovation Network. His team built twelve dimensions of risk alongside an autonomy framework, tracking how far an agent can reach, how long it persists, and what data it can touch. This approach aims to clarify who answers when an agent acts without a human checking every decision.
How the technology stack works
Illing described the architecture as a stack. Point solutions for single workflow steps-scheduling AI, image acceleration, dictation-sit at the base. Multimodal foundation models that combine imaging with electronic medical record data and pathology sit above. Agents and multi-agent systems then coordinate on top, typically through a supervisor agent that delegates to sub-agents and receives their output. A semantic layer, a shared machine-readable definition of an institution's terminology and data models, is essential. Without it, agents interpret the same clinical concept differently across departments. This is why deployments so far cluster inside single organizations rather than spanning networks.
The architecture reflects a growing focus on AI Agents & Automation within clinical workflows. Illing also described a five-level scale for coding autonomy, from level zero, where a human writes every line, to level four, fully autonomous engineering. Most organizations sit around level one today. Synava, the company behind the European radiology platform medavis, moved from zero to 100% agent-written code in a year, deployed with 94% accuracy and signed off by humans.
Where agentic systems fail
Panagiotis Kouvaros, PhD, principal research scientist at Safe Intelligence, explained that as agents become more capable, they also open more paths to failure. For image-based AI, his lab uses formal verification-a mathematical proof that a model's output stays consistent across a defined range of input variations. However, that approach breaks down for agents that retrieve external data and maintain context over long tasks. Instead, his team uses specification-driven validation, building datasets from clinical protocols, decision rules, and ontologies to construct edge cases.
He described a triage agent that correctly flags a textbook heart attack description as high urgency. Change the wording to reflect natural patient vocabulary, and the same case can slide into a low-urgency muscular complaint. "A human clinician catches that instantly. An agent trained on narrower language patterns might not," Kouvaros said. He added that validation will always lag behind AI development: "AI will always be one step ahead of validation. Validation catches up. AI moves again. I don't think it will run ahead indefinitely."
Why this matters for healthcare professionals
Agentic systems are not coming-they are already live in national health services, and they carry risks that standard performance metrics miss. The validation methods that work for static image models do not apply to agents that plan and act over time. For clinicians and health system leaders, the immediate task is to pair every agent with a human, define clear boundaries on what data an agent can touch, and build validation routines that reflect real-world language variation, not just textbook cases. The shift from writing code to reviewing agent-generated code is already underway in some organizations, which changes what technical skills health IT teams need to develop and maintain. AI for Healthcare training and governance will need to keep pace with a technology that, as Kouvaros said, stays one step ahead.
Your membership also unlocks:
