AI-enabled Budget Slack Prediction and Explanation for the Public Sector: How can AI help improve public budgeting?
Start Date
April 1, 2026
End Date
March 31, 2029
Description
Administrative divisions of the Swiss public sector have been found to project significant deficits but have substantial surpluses, undermining trust in the budgeting process and suggesting a lack of fiscal discipline. This budgetary slack can lead taxpayers to overpay for services not rendered or delay necessary investments, distorting public perceptions of government efficiency and eroding trust in the public sector. For the public sector, accurate budgeting is essential for fiscal responsibility and efficient resource allocation. Compared to the private sector, the public sector has received relatively little research on their budgeting practices and potential improvements. This is despite the public sector’s unique complexities of managerial reforms, promising emerging opportunities for processual improvements and large share of spending in most Western countries amounting to 40% to 50% of a country’s gross domestic product.
This collaborative project aims at examining the possibilities in improving budgeting in the public sector by using AI. Together with public organizations, we explore collaborations between humans and artificial intelligence (AI) in budgeting aiming at reducing unfortunate outcomes of heightened budgetary slack. We formulate our guiding research question as follows: How can an AI help reduce public sector budgetary slack? The project is structured in two phases. Phase 1 involves the development of a Budget Slack Prediction and Explanation System for public budgeting in close cooperation with public sector officials. The system has two key components: a Budget Slack Prediction Model and an AI Explainer, that is, a fine-tuned Large Language Model (LLM) explaining the budget prediction and potential budgetary slack (“PublicBudgetingGPT”) that we open source. The AI Explainer component provides information about slack predictions that aim at helping individuals increase understanding of a budgeting context and increase the efficacy of an AI slack prediction. Phase 2 involves experimental validation of the two key components of the system. In Phase 2, we intend to enhance our understanding of how AI can be embedded in budgeting by providing evidence on the role of perceived AI budgeting competence, added efficacy of an AI Explainer, and the role of risk aversion and process accountability when budgets are formed.
The project contributes to the literature in accounting on budgeting and AI. While accounting academics are starting to show a strong interest in the way decision-makers use AI systems in accounting tasks, the accounting literature is in its early stages and has employed hypothetical cases to explore this question. Conceptually, we contribute to the participation literature in accounting and budgeting by examining how AI can be effectively embedded in budgeting processes considering decision-maker’s possible AI aversion/appreciation, risk, and accountability considerations.
The project further contributes to explainable AI (XAI) research by proposing a novel approach enabling laymen to better understand budget predictions, that is, time series predictions for which we face a scarcity of explanation methods. We reach that goal via the development of a Budget Prediction and Explanation System that features a conversational interface with which users can pose their explanation needs in natural language instead of investigating static statistical explanations that reflect the status quo. Creating a novel instruction data set for budget explanation, we fine-tune a pre-trained LLM for explaining budget predictions. Ultimately, our work advances the understanding of human-AI collaboration in public administration, offering theoretically grounded research insights into explainable AI’s role in fostering trust and accountability in governmental decision-making
This collaborative project aims at examining the possibilities in improving budgeting in the public sector by using AI. Together with public organizations, we explore collaborations between humans and artificial intelligence (AI) in budgeting aiming at reducing unfortunate outcomes of heightened budgetary slack. We formulate our guiding research question as follows: How can an AI help reduce public sector budgetary slack? The project is structured in two phases. Phase 1 involves the development of a Budget Slack Prediction and Explanation System for public budgeting in close cooperation with public sector officials. The system has two key components: a Budget Slack Prediction Model and an AI Explainer, that is, a fine-tuned Large Language Model (LLM) explaining the budget prediction and potential budgetary slack (“PublicBudgetingGPT”) that we open source. The AI Explainer component provides information about slack predictions that aim at helping individuals increase understanding of a budgeting context and increase the efficacy of an AI slack prediction. Phase 2 involves experimental validation of the two key components of the system. In Phase 2, we intend to enhance our understanding of how AI can be embedded in budgeting by providing evidence on the role of perceived AI budgeting competence, added efficacy of an AI Explainer, and the role of risk aversion and process accountability when budgets are formed.
The project contributes to the literature in accounting on budgeting and AI. While accounting academics are starting to show a strong interest in the way decision-makers use AI systems in accounting tasks, the accounting literature is in its early stages and has employed hypothetical cases to explore this question. Conceptually, we contribute to the participation literature in accounting and budgeting by examining how AI can be effectively embedded in budgeting processes considering decision-maker’s possible AI aversion/appreciation, risk, and accountability considerations.
The project further contributes to explainable AI (XAI) research by proposing a novel approach enabling laymen to better understand budget predictions, that is, time series predictions for which we face a scarcity of explanation methods. We reach that goal via the development of a Budget Prediction and Explanation System that features a conversational interface with which users can pose their explanation needs in natural language instead of investigating static statistical explanations that reflect the status quo. Creating a novel instruction data set for budget explanation, we fine-tune a pre-trained LLM for explaining budget predictions. Ultimately, our work advances the understanding of human-AI collaboration in public administration, offering theoretically grounded research insights into explainable AI’s role in fostering trust and accountability in governmental decision-making
Leader contributor(s)
Member contributor(s)
Kotzian, Tobias
Funder