Participatory Strategic Foresight with Generative AI: Field Learnings from a Research-Through-Design Study for Health System Planning

Health systems are complex adaptive systems whose structural uncertainty limits prediction-based planning, and generative AI is increasingly proposed to enrich the foresight used to navigate them. Whether cheaper, more fluent scenario production actually improves collective strategic judgement, however, remains largely untested in real organisations. This article reports a research-through-design study of an AI-assisted participatory foresight process run for the planned Girona Health Campus (Catalonia): ten domain workshops with about 250 professionals, 170 AI-assisted scenario drafts, and an organisation-wide SmartDelphi validation. Reconstructing the process abductively from its documentation, three behaviours recur. Generation increased the number of scenarios but not the range of futures they covered, pulling repeatedly toward technological resolution: 80% of the analysed scenarios referenced technology and 61% were legible as techno-optimistic accounts. The decisive work therefore migrated downstream to synthesis, where situated meaning was preserved, diluted, or lost. The characteristic failure mode was not poor output but over-trust in fluent output, which participants countered only when the design required them to contest and restate it. Validation itself functioned as organisational diagnostics, exposing a consistent desirability–feasibility gap that varied with the kind of change each future demanded. We consolidate these findings into five transferable design principles and argue that, once the capacity to imagine futures becomes abundant, methodological attention must shift from producing more scenarios to managing that abundance while safeguarding quality.