The First Multi-Agent AI Pilot for Clinical Trial Matching: Architecture, Results, and Compliance

First Multi-Agent AI Pilot for Clinical Trial Matching Achieves 30% Faster Enrollment As of May 26, 2026, a consortium of ten academic medical centers has completed the first documented pilot of multi-agent AI for clinical trial matching—a leap forward for patient-centric clinical research. The system, deployed on AWS Bedrock using Anthropic’s Claude 5 Haiku and Meta’s Llama 5, cut patient enrollment time by 30% and reduced screening errors by 22%, according to a technical report released this week by the consortium. For healthcare operations leaders evaluating agentic AI in clinical research, this pilot provides a rare, vendor-neutral blueprint that balances innovation with the strict regulatory demands of HIPAA and GxP environments. Inside the Pilot: 10 Medical Centers Unite for AI-Driven Trial Matching The consortium—comprising large academic medical centers from across the United Sta

tes—formed in early 2026 to tackle a persistent problem: clinical trial enrollment remains slow, costly, and error-prone. Manual patient screening against hundreds of complex eligibility criteria across multiple trials often takes weeks and misses qualified candidates. The pilot aimed to test whether multiple specialized AI agents, each handling a discrete part of the screening and matching workflow, could outperform traditional single-model or manual approaches. The pilot ran for eight weeks using de-identified retrospective data from over 15,000 patient records and 200 active oncology and cardiology trials. The results, detailed in the consortium’s preprint, exceeded operational expectations. Enrollment speed—measured as the median time from initial patient identification to signed informed consent—dropped from a historical average of 12 days to 8.4 days, a 30% improvement. Screening e

rrors, defined as instances where a patient was incorrectly deemed ineligible or a mismatched trial criteria was applied, fell by 22% compared with prior manual review benchmarks. Agent Roles and Orchestration on AWS Bedrock The architecture leverages Amazon Bedrock AgentCore, generally available as of 2026, to orchestrate a team of four specialized agents. Each agent has a clear scope, and they communicate through a shared, secured message bus within a HIPAA-eligible Bedrock environment. The blueprint, visualized below, is designed to be cloud- and model-agnostic, though the pilot uses Claude 5 Haiku and Llama 5 for their specific strengths. Text-based architecture: Coordinator Agent – routes tasks, maintains context, and ensures the overall workflow adheres to clinical protocol. Patient Records Agent (Claude 5 Haiku) – ingests and normalizes structured and unstructured EHR data (labs,

diagnoses, medications) into a unified patient profile. Trial Eligibility Agent (Llama 5) – compares the patient profile against trial inclusion/exclusion criteria, handling complex nested logic. Compliance Audit Agent – monitors every data exchange for HIPAA compliance, logs all decisions for audit trails, and enforces data residency policies. Agents operate sequentially but also in parallel where possible. For example, while the Patient Records Agent extracts data, the Trial Eligibility Agent can pre-fetch trial criteria. The Coordinator then synthesizes a ranked list of matching trials, complete with confidence scores and a human-readable explanation. AWS Bedrock AgentCore handles state management, error handling, and multi-turn conversations, reducing the engineering burden for healthcare IT teams. Why Claude 5 Haiku and Llama 5? Model Selection and Specialized Tasks The consortium e

xplicitly chose two distinct foundation models to balance speed, accuracy, and compliance. Claude 5 Haiku, Anthropic’s fastest model at the time of the pilot, excels at low-latency text processing and was used for the Patient Records Agent. Its training includes safeguards for handling sensitive data, making it suitable for initial screening while adhering to HIPAA’s minimum necessary standard. The model runs entirely within the customer’s Bedrock environment, with no data retention by the model provider. Llama 5, Meta’s latest open-weight model as of May 2026, powered the Trial Eligibility Agent. The centers fine-tuned Llama 5 on a curated dataset of clinical trial protocols and annotated patient cases, giving it deep domain reasoning over trial criteria. Because Llama 5 can be deployed in a private VPC, all trial-matching logic stays within the institution’s controlled cloud perimeter,

satisfying strict data residency requirements. The dual-model design illustrates a vendor-neutral pattern: use a fast, managed model for heavy-lifting data preprocessing and a customizable model for specialized reasoning—both orchestrated through a common interface like Bedrock. The 30% Faster Enro