New research from the University of Eastern Finland and Aalto University finds that generative AI projects in public administration often continue not because they work well, but because teams use a set of compelling rationales that make stopping them difficult. The study, based on 1.5 years of ethnographic fieldwork in Finland, examined how developers of an AI decision-support tool justified ongoing investment and experimentation even when the tool underperformed. Researchers say these justificatory frames — which stress efficiency, cost savings, and the promise of innovation — help sustain momentum by framing setbacks as learning opportunities rather than failures.
The specific AI tool studied was built to help claims specialists navigate vast and complex guidance documents more efficiently. When its accuracy and consistency faltered, the project team leaned into familiar AI promises like improving fairness, boosting employee well-being, and reducing costs. These tool-oriented framings helped defend the project’s value and made it appealing across different organizational groups, even when output quality lagged expectations.
In addition to tool-centric arguments, process- and ideology-oriented rationales played a key role. Project leaders emphasized speed, courage in innovation, and experimental mindsets to normalize setbacks and frame risks as necessary parts of pioneering work. This approach built a protective narrative infrastructure that made critical evaluation less likely and positioned continued testing phases as essential for eventual success.
The study also highlights the role of boundary work — efforts to build alliances with managers and consultants while distancing traditional frontline workers — in sustaining generative AI projects. By reframing problems as matters of organizational change or user adaptation rather than tool limitations, teams sustained investment and momentum. This complicates how public sector AI development should be assessed, as narrative and institutional pressures can keep projects moving even when technical performance remains weak.
