Presentation (CTI)

OBJECTIVES: AI agents are increasingly being used for individual HEOR tasks, but whether they can be chained across the model lifecycle without eroding decision credibility is unclear. We evaluated whether validator, explainer, and report writer agents could form a governed, human-in-the-loop chain to validate, interpret, and document a completed cost-effectiveness model in advanced breast cancer.
METHODS: The artefact was a partitioned survival model in HER2-negative germline BRCA-mutated advanced breast cancer (NHS and PSS perspective). The validator agent ran a checklist with static and dynamic checks, returning severity-classified findings; failing models returned for correction and re-validation until QC passed. The validated model was loaded into the explainer agent, which answered the report writer's input-form questions from model contents, and the report writer drafted a full technical report from these answers. Both were then re-loaded into the explainer for new reviewers to interrogate and check consistency. Senior health economists adjudicated every checkpoint; validator performance was tested against a human-only baseline of 22 seeded errors.
RESULTS: The validator detected 21 of 22 seeded errors (95%) which seniro health economist conifrmed,, including a state-utility mapping fault and a transition-probability normalisation error; the missed error surfaced only on expert review and was found to be non-critical. Validation completed in under three hours per cycle (a greater than 90% time reduction). The explainer answered 88% of input-form fields without manual entry, cut input-preparation effort for writing agent by approximately 65%, and reduced reviewer onboarding from 20 to 6 hours. First-draft report readiness moved from six to eight weeks to under one working day, with inter-agent error propagation the principal residual risk.
CONCLUSIONS: A validation-gated chain of specialised agents, bounded by expert adjudication, accelerated the oncology model lifecycle without displacing methodological judgement, positioning multi-agent orchestration as auditable decision support with a human-cleared validation gate as the binding constraint between agents.