Presentation (CTI)

OBJECTIVES: HTA dossier preparation requires medical writers to synthesize evidence from multiple source documents, a rigorous and resource-intensive process. Generative AI offers a way to accelerate and scale this drafting, but large language models can introduce citation errors and unsupported claims - making built-in quality controls essential for reliable use. We developed an agentic AI workflow for HTA dossier generation that revises each section until requirements are met, ensuring overall dossier quality.
METHODS: The agentic AI workflow runs within an evidence repository platform on an asynchronous, state-machine-driven pipeline. Variable-based prompts are mapped to each dossier section. Each section runs through a multi-step agentic loop: (1) Draft Generation - the LLM creates text and tables from prompts and uploaded sources, using prior sections for whole-dossier consistency; (2) Citation Fact-Checking - validates inline citations against sources, returning corrective suggestions; (3) Refinement - applies corrections, enforces citation formatting, and resolves instruction violations; (4) Quality Gate - pass/fail compliance check determines whether re-iterations are required.
RESULTS: The system was deployed across 3 HTA dossier templates containing 15 to 25 sections, tested using 6 historical dossiers. Output was reviewed by a senior medical writer, confirming quality according to success criteria and reduction of human workload in dossier drafting. Most sections required citation corrections after initial generation, confirming model tendency to extrapolate beyond cited evidence. On the first refinement cycle, 30-40% of sections passed the quality gate; approximately 90% achieved compliance by the second iteration.
CONCLUSIONS: An agentic AI workflow with iterative fact-checking and quality-gate verification enables automated HTA dossier generation, where each section self-corrects through successive refinement cycles and verification outputs directly drive targeted corrections - minimizing hallucinated references and unsupported claims while supporting scalable and auditable integration into HTA medical writing. Such workflows are recommended as a reliable first-pass drafting layer that complements expert fact-checking and clinical interpretation.