From cloud to desktop: a 3-step methodology for secure AI adoption in government management

Abstract

Artificial intelligence (AI) is transforming all sectors of society, with exponential growth in capacity and applications. However, its integration into the public administration sector faces three key challenges that, depending on the context, can manifest with varying intensity: data sensitivity, the availability of specialized infrastructure, and the need for technological sovereignty. In contrast to the dominant trend of cloud-based solutions, this article addresses the existing gap in practical methodologies for resource-constrained environments where highly confidential information needs to be processed. A three-step iterative methodology is proposed and validated: (1) intensive exploratory analysis using publicly available information in open environments to extract functional requirements; (2) development of automation and early warning tools using locally executed technologies (regardless of the language or framework); and (3) deployment in controlled environments with continuous feedback that informs internal processes and tool design. Through implementation in cases involving the monitoring of economic and social programs, it is demonstrated that it is possible to balance algorithmic innovation with confidentiality requirements, reducing analysis times and improving the quality of decision-making. The second case, although validated with synthetic data, demonstrates the viability of the approach for detecting time lags between planning instruments. This article contributes to the literature on digital government and knowledge management by offering a replicable, low-cost, and technologically neutral model for the secure adoption of AI.