JMIR Publications

ABSTRACT

Background:

Health systems worldwide are introducing artificial intelligence–enabled clinical decision support systems (AI-CDSS), even if their clinical, professional, and organizational effects haven’t been fully established. Most empirical evidence concerns isolated, post-deployment pilots, while the pre-implementation phase, in which strategic decisions must be made under pervasive uncertainty, hasn’t been fully explored so far. Cross-level accounts spanning the micro, meso, and macro levels, at the national scale, and differentiated by type of uncertainty generated by the introduction of such systems, remain scarce.

Objective:

This study examines the main dimensions of AI-related uncertainty experienced by prospective users and managers during the pre-implementation phase of a large-scale national AI-CDSS for primary care, and how national policymakers respond to them, addressing, displacing, or leaving them unresolved.

Methods:

We conducted a qualitative-driven, multi-level case study of the Italian national AI-CDSS program for primary care, spanning the micro, meso, and macro levels of the health system. Methods comprised semi-structured interviews with general practitioner (GP) key opinion leaders (n=21), a survey of health care organizations’ top management (n=24), and documentary analysis of national procurement specifications complemented by a focus group with policymakers. Uncertainty was analyzed using an integrated framework distinguishing substantive, strategic, and institutional uncertainty.

Results:

Uncertainty was not uniform across the system. At the micro level, GPs faced predominantly substantive uncertainty, spanning epistemic, professional, relational, and ethical dimensions. At the meso level, managers showed strong convergence on general commitments but persistent dispersion on operational questions. At the macro level, institutional uncertainty dominated, with a focus on the regulatory dimension. Institutional responses were correspondingly asymmetric: epistemic uncertainty was directly addressed through medical device certification and mandatory training; professional and ethical uncertainty was transferred to individual clinicians or converted into governance responsibilities; strategic uncertainty received a procedural response through phased rollout and iterative change management; and regulatory uncertainty remained structurally open.

Conclusions:

Managing AI-related uncertainty is best understood through what we term the matched-response principle: matching each institutional response to the specific form uncertainty takes at a given level of the system. On this basis, we argue that national programs should engage the data protection authority as a governance partner, pursuing parallel development of technology and regulation; should formalize the meso level as an active governance layer before scale-up; and should reframe professional engagement creating structured arenas in which to negotiate shared meaning around professional identity, the clinical relationship, and accountability. Although the Italian case is shaped by distinctive regulatory and contractual features, the level-specific distribution of uncertainty and the asymmetry of institutional responses are likely to recur in any health system that commits to clinical AI tools.