Presentation (CTI)

OBJECTIVES: Health technology assessment (HTA) agencies increasingly use artificial intelligence (AI) to inform evidence inputs that can shape resource-allocation recommendations. However, no AI-specific route exists to challenge AI-derived outputs before they enter committee deliberation. This produces an operational governance gap at the assessment phase and assessment-appraisal interface. We characterise this gap and assess whether existing governance instruments operationalise contestability for AI-derived inputs in HTA.
METHODS: A narrative review of peer-reviewed literature, grey literature, and governance documents (January 2018-March 2026) was conducted across five bibliographic databases (PubMed, Scopus, Web of Science, Cochrane Library, CINAHL) and HTA, regulatory, and standards sources. Twenty-nine governance instruments, spanning cross-sectoral, health-sector, and HTA-specific frameworks including three from adjacent fields (privacy law, public-sector automated administration, health-data governance), were purposively mapped against seven governance dimensions using a three-level coding scheme (implementable, articulated, absent), with coding rationales recorded.
RESULTS: The review identified three systemic governance deficits (regulatory ambiguity, procedural deficit, architectural fragmentation) and five diagnostic findings specifying what any solution must address. No instrument covered all seven governance dimensions. Contestability was the only dimension operationalised exclusively outside the AI×HTA intersection (privacy law, public-sector automated administration, and HTA procedural governance for non-AI inputs). Within the intersection, contestability was articulated as a principle but never operationalised. The gap is therefore intersectional: AI-specific governance does not engage with HTA’s procedural mechanisms, while HTA-procedural governance is largely silent on AI.
CONCLUSIONS: Contestability is the precondition that makes the adequacy of the other six governance dimensions verifiable from outside the agency, and is therefore the procedural backbone for responsible AI integration in HTA. Agencies should establish a pre-deliberative route to challenge AI-derived inputs within the assessment workflow, before they acquire deliberative standing. We propose a preliminary Responsible AI Integration Framework (RAIF), anchored in Accountability for Reasonableness, with requirements proportionate to risk. The framework is preliminary, awaiting empirical testing.