Presentation (CTI)

OBJECTIVES: The EU HTA Regulation has increased the volume of analytical work needed for HTA submissions while compressing timelines. Many organizations are turning to generative AI to address capacity constraints. While AI can scale production, statistical and methodological judgment may become diluted or displaced. We examine what it takes to incorporate such judgment and what this means for organizational AI strategy.
METHODS: A network meta-analysis workflow is used as an illustrative case. We map key stages of the workflow and distinguish between judgments embedded during design, such as defining the estimation target, evidence selection, and network construction, and judgment applied later through output review or validation.
RESULTS: When expert judgment is incorporated in workflow design, AI augments the expert by executing transparent, pre-specified tasks. When judgment is implicit in set defaults or AI-based automated decisions on inclusion of evidence and methodology, the expert role is reduced to a downstream validator. Early decisions about target estimands, eligible evidence, and network structure aligned with accepted HTA methodologies are the foundation needed before modeling begins. If expert decision-making is not a mandatory step in AI-assisted analytical workflows, validation may require a full reconstruction rather than a statistical review of outputs, reintroducing the bottleneck AI was intended to remove.
CONCLUSIONS: Scaling analytical workflows for HTA and HEOR using AI is a workflow design challenge, not simply an automation problem. Because HTA and HEOR rely on methodological judgment without the same formal workflow safeguards as regulatory analyses, statisticians and methodological experts need to shape design, not only validate outputs. This has implications for AI strategy: capability development, so analytical teams can engage with AI pipeline design; procurement, so external tools separate judgment from execution; and timing, so statistical input happens during system design instead of after deployment.