Presentation (CTI)

OBJECTIVES: Artificial intelligence (AI) proof-of-concepts (PoCs) are increasingly used in Health Economics and Outcomes Research (HEOR) to demonstrate feasibility and innovation. However, successful PoCs often fail to translate into reliable, HTA-ready products, creating a false sense of validation. This study examined why AI PoCs do not necessarily demonstrate real-world suitability for HEOR and identified the methodological, governance, and operational gaps that must be addressed to progress from PoC to deployable product.
METHODS: A qualitative analysis was conducted across multiple AI-enabled HEOR initiatives spanning literature review support, economic modelling, real-world evidence generation, and technical reporting. PoC designs were compared with production-ready requirements across five dimensions: (1) representativeness of test data, (2) robustness to edge cases and data drift, (3) transparency and auditability of outputs, (4) integration with HEOR quality management systems, and (5) accountability and ownership post-deployment. Failure modes were synthesised into a PoC-to-product maturity framework tailored to HEOR and HTA contexts.
RESULTS: PoCs were typically optimised to demonstrate technical feasibility under constrained, low-risk conditions, often relying on curated datasets, limited scenarios, and implicit expert supervision. These conditions masked key failure modes encountered during scale-up, including loss of traceability, inconsistent performance across disease areas, sensitivity to evolving evidence, and unclear accountability for errors influencing pricing or reimbursement decisions. Successful transition to product required substantial redesign beyond the PoC stage, including explicit validation thresholds, reproducibility controls, governance checkpoints, and post-deployment monitoring aligned with HTA expectations.
CONCLUSIONS: In HEOR, AI PoCs demonstrate possibility, not readiness. Treating PoCs as evidence of validity risks premature deployment of tools that fail under HTA scrutiny. Progressing from PoC to product requires reframing success criteria around governance, auditability, reproducibility, and decision impact, positioning AI development as a regulated analytical lifecycle rather than a one-time technical achievement.