Presentation (CTI)

OBJECTIVES: After a model submission, health technology assessment (HTA) bodies issue clarification questions that manufacturers must answer within tight windows, requiring rapid, accurate explanation of model logic. We evaluated whether a model explainer agent, a conversational AI agent that explains an economic model in natural language, could accelerate clarification-response drafting without compromising accuracy or methodological ownership.
METHODS: The agent ingested an Excel cost-effectiveness model ofrare disease: a Markov cohort model with 5 health states, quarterly cycles, a lifetime horizon and a payer perspective (36 worksheets, ~150,000 cells). It was parsed into a structured, queryable representation, with documentation indexed via retrieval-augmented generation (RAG). A panel of 30 technical clarification questions related to model, simulated from submissions to NICE and Canada's Drug Agency (CDA-AMC), spanned structure, parameter traceability, costs, utilities, discounting, extrapolation and scenario requests. For each, it located the worksheet, cell range and driving formula, explained the rationale and drafted a reviewer-ready response. Senior modeler adjudicated outputs against a manual baseline. The agent explained but never modified the model and allow modeler to own the output and query further for clarifications.
RESULTS: The agent correctly localised the relevant component for 30 of 30 questions (100%) on first pass, with accurate performance on logic embedded in even VBA-automated probabilistic analysis. Median time to a first draft fell from about 2 hours to 8 minutes per question and the full package from 2-3 weeks to about 1.5 days. Every adjudicated reference and assumption traced to a real cell or source passage; none were fabricated. Failure modes included residual verbosity, two scenario-tab mislabeling, and an inability to adjudicate methodological defensibility, which stayed a senior-economist decision.
CONCLUSIONS: A governed, human-in-the-loop model explainer agent accelerated clarification-response drafting across two HTA jurisdictions while preserving traceability and expert ownership, functioning as decision support rather than a replacement for methodological judgement.