Evaluating an AI Act Regulatory Sandbox Framework Against Pilot Cases: Towards a Methodologically Grounded Three-Round Process Framework

This research-in-progress paper advances out tentative AI Regulatory Sandbox Framework (AIRSF) in two directions. Empirically, it evaluates 46 pilot cases and with five workshops of Nordic national/regional AI ecosystems. Conceptually, it argues that the AI Act's dual objective of innovation and compliance can only be operationalised in a business-friendly manner if grounded in participatory, situation-adaptive method development with dispute resolution procedures on controversial issues. Earlier we have identified four methodological steps: stakeholder mapping, conflict surfacing, iterative KPI-building, institutionalised conflict resolution, and show their concrete design consequences for a three-round sandbox related sub-processes: (1) a revised use case and risk category definition procedures with company-paced entry; (2) multi-authority evaluation with agile guidance and co-ordination; (3) post-market surveillance options with iterative re-entry. Preliminary findings surface three structural challenges: entry friction for determining use case purpose,  borderline high-risk cases, capabilities altering the AI System design and purpose while in sandbox, and a post-market visibility gap in Round 3.