Presentation (CTI)

OBJECTIVES: As AI adoption increases in HEOR, the terms AI and automation are often used interchangeably despite referring to distinct technological paradigms. This study aims to propose a practical framework to distinguish deterministic automation from learning-based AI, drawing on lessons from AI-mature industries and translating them to common HEOR workflows.
METHODS: We conducted a cross-industry benchmarking exercise comparing robotic process automation in finance, autopilot systems in aviation, programmable logic control in manufacturing, warehouse and routing automation in logistics, and rule-based engines in insurance with applications of machine learning and generative AI. These tasks were mapped onto similar HEOR activities like systematic literature review (SLR) screening, data curation, economic model development, and real-world evidence generation. Each technology class was evaluated on adaptability to unstructured data, transparency of decision logic, and the magnitude and reversibility of error propagation.
RESULTS: Across industries, deterministic automation is most effective for stable, codifiable workflows: trade reconciliation in finance, standard flight phases, assembly-line control, parcel sorting, and eligibility checks in insurance. By analogy, rule-based tools in HEOR are superior for SLR de-duplication and structured trial data extraction, where deterministic behaviour and reproducible audit trails are required for HTA submissions. Learning-based AI is reserved for noisy, weakly structured problems: fraud detection and market anomaly spotting in finance, predictive maintenance in aviation, computer-vision quality control in manufacturing, demand and congestion forecasting in logistics, and claims fraud detection in insurance. Parallel HEOR use cases include mining heterogeneous real-world data, imputing missing outcomes, and extrapolating long-term survival from short-term surrogates.
CONCLUSIONS: Effective HEOR modernization requires demarcated strategy rather than a one-size-fits-all "AI adoption" plan. Failure to distinguish between AI and automation risks deploying opaque, probabilistic models where transparent, rule-based logic is ethically and scientifically feasible.