Context
In 2019, Luxembourg’s prime minister outlined a strategic vision for artificial intelligence (AI) as a key tool for government innovation. This initiative aligns with the ‘Outlier Detection’ project, which aims to streamline the budget review process for municipalities. The Department of Municipal Finances (DMF) currently faces a challenge: manually reviewing budgets and financial reports from 203 municipalities. This time-consuming task, requiring eight staff members for four months, limits their ability to work on other important projects.
The critical nature of these budgets, which establish the legal foundation for municipal spending, necessitates a timely review process. However, the sheer volume of data makes a thorough manual examination difficult. The Outlier Detection project seeks to address this challenge by leveraging AI for data-driven decision-making. This approach aims to improve efficiency and free up valuable staff resources within the DMF.
Objectives
The main objective is to ensure the legality of budgets and financial reports, freeing up time for more strategic projects. This is achieved by identifying the most likely coding errors, allowing staff to focus on pre-selected coding that have frequently been found to be incorrect. The Outlier Detection tool detects three key types of coding errors: descriptions that do not match accounting codes, incorrect expense allocations, and discrepancies between monthly data and annual budgets.
Implementation
The transition to a computerised realm was marked by the Outlier Detection project, assisted by machine learning algorithms. These algorithms cultivate an understanding of the typical vocabulary used for each budget component, constructing a multidimensional map of words and their vectorial relationships. Words statistically deviating significantly from the anticipated vocabulary for a given component are flagged as potential encoding errors. Findings are seamlessly integrated back into the algorithm, facilitating supervised learning and enhancing predictive capabilities.
The project unfolded in two distinct phases, engaging different stakeholders. The initial proof-of-concept phase saw active participation from experts within the department, external contractors and workshops to assess effective approaches. The decision on which algorithm to implement was guided by a business intelligence analyst and a contracted user interface designer. The second phase involved testing with representatives from eight municipal entities, refining the algorithm’s learning capabilities.