Public Sector : Government Agency accelerates Data Processes with Large Language Models

A large public specialist authority faced a typical challenge of the digital age: Ever-increasing volumes of data from research, monitoring, and administration needed to be processed, documented, and presented in an understandable way. The database was heterogeneous, the analysis processes complex – and many work steps were still carried out manually. Particularly time-consuming was the annual review of tens of thousands of project descriptions for relevance in a central research database. In the area of monitoring, automated analyses and visualizations were also lacking, making it difficult to identify important relationships between influencing factors and measured values without significant effort. The goal: Artificial intelligence and Large Language Models (LLMs) were to help systematically unlock efficiency potential, significantly shorten processing times, and increase decision-making certainty.

In a broad ideation phase, potential AI application areas were identified, evaluated, and prioritized according to benefit and feasibility. Prioritized initiatives were then further developed in cross-functional workshops and implemented in agile pilot projects – with a clear focus on user-friendly web applications for employees. For text classification, applications ranging from supervised learning methods up to fine-tuning of Large Language Models were used. The models were systematically benchmarked and the best ones were deployed in a proper application. For the analysis of monitoring data, we relied on exploratory methods with benchmarking of supervised and unsupervised approaches. Accompanying the project through knowledge transfer on state-of-the-art algorithms and joint reviews ensured acceptance and sustainable anchoring of the solution. 

With finely tuned Large Language Models, the authority achieved a classification accuracy of over 97 percent in automated text classification and reduced laborious review processes from months to just a few hours. This keeps the database up to date and effectively avoids duplicate work. In monitoring, quantified external influencing factors, interactive analyses, and intelligent recommender systems enable new insights into relationships between measured values and influencing factors. The scalable web applications are implemented in a near-production environment and empower employees to independently develop further AI use cases – a key step towards a more efficient, data-driven administration.