Every hospital with a radiology department has a version of the same problem. Scans arrive. A worklist system routes them to radiologists. The routing logic was set up years ago, updated occasionally, and runs the same way regardless of what is actually happening on the floor.
The consequences are well documented. Radiologists tend to pick the easier, faster scans from the queue, a behavior the industry calls cherry-picking. It is not malicious: it is human nature under pressure, and it is partly driven by how radiologists are paid, with compensation tied to volume rather than complexity. The complex, time-consuming cases sit longer. Urgent ones that happen to look routine get read late.
Research across 62 hospitals and 2.2 million studies put numbers on the damage: nearly 18 minutes of added delay for expedited cases, and costs running between $2.1 million and $4.2 million across hospital networks. A separate study found that AI-assisted worklist prioritization for pulmonary embolism cases, a blood clot in the lungs that can be fatal if missed, reduced median detection time from several days to one hour. The missed rate for that condition dropped from 44.8 percent to 2.6 percent.
AWS, working with Radiology Partners, has now published a full technical blueprint for replacing static worklists with a network of AI agents. The system is built on Amazon Bedrock AgentCore, which became generally available in October 2025. The agents work in parallel: one pulls scan metadata, one retrieves patient history, one checks radiologist availability and current workload, one applies contractual requirements like response-time guarantees. A coordinating agent weighs all of this and makes the assignment, with a written rationale attached.
The detail that matters most for anyone evaluating this type of system is the memory architecture. Short-term memory keeps context alive within a session, so if an assigned radiologist rejects a case, the system does not start from scratch. Long-term memory stores patterns across thousands of assignments. Episodic memory captures specific events like deadline breaches and rejection reasons, then surfaces them as lessons the system actually uses in future decisions. This is the part that rule-based systems cannot replicate. When a rule engine produces a bad assignment, it will produce the same bad assignment tomorrow unless a human changes the rule.
The AI also handles a second workflow: real-time scan triage. When imaging models detect findings like an acute pulmonary embolism or intracranial hemorrhage with high confidence, the system automatically bumps that case to the top of the queue before any radiologist even opens the worklist. This is already validated in clinical research: AI reprioritization cut turnaround time for confirmed pulmonary embolism cases from about 60 minutes to under 48 minutes in a University of Texas Southwestern study.
Burnout is a real constraint here. A recent study found radiologists spend nearly 44 percent of their day on tasks that are not actually reading scans. Intelligent case distribution, the kind that prevents one radiologist from receiving a steady stream of the most demanding studies, directly addresses the fatigue accumulation that drives both burnout and errors. Over 70 percent of radiologists report burnout symptoms, and the shortage of trained radiologists is already straining imaging departments globally.
For operators outside healthcare, the architecture itself is instructive. A network of specialized AI agents, each handling one part of a complex decision, with memory that learns from outcomes, applies to any operation where task assignment currently runs on static rules: logistics dispatch, field service routing, claims triage, procurement approvals. The radiology case just happens to be one where the stakes make the problem impossible to ignore.
This is not a product launch in the usual sense. It is a published blueprint built on AWS infrastructure, meant to be replicated. Healthcare organizations interested in piloting it are directed to their AWS account representative. The open question is not whether the technology works, the evidence on AI worklist prioritization is reasonably strong. The question is how fast hospital procurement and compliance processes move to adopt it.