Presentation (CTI)

OBJECTIVES: Rapid, high-quality HTA dossiers are central to timely patient access to innovative therapies. We compared writer engagement and preference between an iterative agentic AI assistant and a conventional retrieval-augmented generation (RAG) tool during dossier drafting.
METHODS: In a randomised, crossover pilot (interim n = 5 writers, target n = 38) at a specialised European medical-writing consultancy, each professional drafted live dossier sections with both systems. Every new session was randomised 1:1, ensuring all writers used both pipelines. Engagement metrics were aggregated at writer level: for each writer we calculated (i) the ratio of text copied from agentic versus RAG outputs and (ii) the proportion of forced-choice prompts favouring the agentic answer; writer-level values were then averaged across the cohort. Generation latency was compared with a Wilcoxon signed-rank test. The agentic assistant executed up to ten reasoning cycles per query, whereas the RAG tool used a single-pass retrieve-then-generate workflow. All analyses are exploratory and will be repeated with the full sample. AI assistance has been disclosed per ISPOR policy.
RESULTS: Across 62 drafting sessions, agentic outputs were copied more often (mean writer-level copy ratio 1.7; 95% CI 1.2-2.3; p = 0.01) and preferred in forced choices 73% of the time (95% CI 64-81; p < 0.001). Writers exchanged a mean 3.2 versus 2.4 messages per session (difference +0.8; 95% CI 0.1-1.5; p = 0.04). Median generation time decreased by 1.2 s (95% CI -2.0 to -0.4; p = 0.006, Wilcoxon test).
CONCLUSIONS: Interim data suggest that an agentic AI assistant improves writer engagement and is more frequently preferred than a conventional RAG tool, while slightly reducing response time. Should these trends persist in the full cohort, agentic systems could streamline HTA documentation and accelerate evidence delivery for reimbursement decision makers.