Umeå universitets logga

Generative AI tools such as GitHub Copilot are being rapidly embedded into enterprise workflows, yet the literature on their organizational and operational value remains underdeveloped. Existing research focuses primarily on individual developer productivity in controlled settings, while the mechanisms through which AI coding assistants reshape decision-making, workflows, and cost structures in real organizational settings have received limited attention. The financial assessability of these tools has been treated as a presupposition rather than an empirical question.  

This study addresses this gap through a qualitative case study of GitHub Copilot’s implementation within the Operations function of Vattenfall BA Markets, an energy trading and portfolio optimization environment where data accuracy and process reliability carry direct financial consequences. The study poses one main research question on behavioral and operational implications, and one diagnostic sub-question on the assessability of financial implications. Empirical material was gathered through semi-structured interviews with ten participants distributed across four blocks corresponding to organizational levels: developers, agile coaches, operations manager, and BA Markets executive management. Data was analyzed thematically and interpreted through a sequential four-theory framework integrating Real Options Theory, Bounded Rationality, Task-Technology Fit, and the Information Systems Success Model.  

Five findings emerge. GitHub Copilot has produced substantial individual productivity gains, accompanied by a structural shift in developer work towards upstream specification and downstream verification. These gains are unevenly distributed because the relevant skills were never formally taught. Faster coding has not translated into faster end-to-end delivery; a bottleneck has formed at the quality assurance and code review stage, and adoption travelled below the visibility of the team-coordination layer. Investment and governance decisions have been made through satisficing under low-information conditions and framed in option-theoretic terms. Finally, the financial implications of the implementation are not currently assessable: cost is structurally invisible inside an enterprise contract; benefits are quarantined at the individual level, and downstream quality constraints cap further value capture.

The study contributes a replicable analytical lens for evaluating Generative AI investments in regulated, data-intensive operational environments. It also reframes the financial assessability of such tools as a measurement problem rather than a presupposition.