Presentation (CTI)

OBJECTIVES: QC of Excel cost-effectiveness models underpins HTA credibility, but it is manual, slow, and usually left to the end, when errors are costliest to fix. We tested whether an AI validator agent, embedded as a continuous QC layer during model development rather than a final gate, could shorten build-and-validate cycles before HTA submission.
METHODS: We applied the agent through the full development of a submission-grade Markov cost-effectiveness model in early or locally advanced triple-negative breast cancer(6 health states, monthly cycles, lifetime horizon, ~185,000 cells, 14 VBA modules). Agent planned each check, mapping intent to sheets and formulas; an executor agent ran static checks (reading values and formulas) and dynamic checks (changing inputs, recalculating), returning severity-classified results with recommended fixes. The agent ran at four development milestones with selective re-runs. To test detection, 24 graded errors were seeded across transition logic, discounting, utility mapping, and cohort accounting.
RESULTS: The agent detected 23 of 24 seeded errors (96%), including dynamic-logic errors: transition probabilities exceeding one under a subgroup toggle, and a hazard ratio applied beyond its waning cycle. Across the four versions it surfaced 41 issues (12 critical), falling to zero critical errors by pre-sign-off. Each full validation pass took about 3-4 hours, against 3-4 days for an equivalent manual pass (around 80% less effort) multiple rounds; one rework cycle was avoided and sign-off advanced by about 2-3 weeks. Four flagged items were expert-reclassified as deliberate design choices. The agent verified computation but not face validity, which stayed with the modeller.
CONCLUSIONS: Embedding an AI validator agent throughout development, rather than at the end, removed every critical error before sign-off, compressed validation rounds, and de-risked HTA submission, while methodological judgement and face validity remained human.