Automate, Assist, Avoid: Caseworkers’ Perspectives on Applying Large Language Model-Based Assistance in Public Sector Decision-Making Processes

Paper Title

Automate, Assist, Avoid: Caseworkers’ Perspectives on Applying Large Language Model-Based Assistance in Public Sector Decision-Making Processes

Publication Info

• Topic area: Application of Large Language Models (LLMs) in public sector decision-making processes.
• Keywords: Large Language Models, public sector, decision-making, caseworkers, discretion, human-computer interaction, generative AI, qualitative study, Finnish public institution, AI-assisted tools.

Background and Problem

• Problem / challenge: Limited understanding of how LLMs can effectively support complex public sector decision-making tasks, including addressing legal, cultural, and procedural variables.
• Significance: Public sector decision-making involves high-stakes, sensitive citizen cases that require balancing efficiency with fairness and discretion. Misuse or overreliance on AI can lead to harmful outcomes, as seen in past scandals.
• Motivation and related work: Prior research has explored AI in public decision-making but often focuses on automation or predictive analytics, neglecting nuanced, context-dependent tasks. There is a lack of consensus on how AI tools can meaningfully assist caseworkers without undermining their discretion or professionalism.

Solution

• Proposed approach: Investigating caseworkers’ perspectives on integrating LLMs into their decision-making processes through qualitative methods, including interviews and workshops.
• Novelty:
  1. Conceptualizing decision-making as a multi-step process (case comprehension, information search, discretion use, case solution) rather than a singular task.
  2. Identifying specific sub-tasks where LLMs could assist, such as translation, information retrieval, and administrative-to-spoken language conversion.
  3. Highlighting the importance of maintaining human discretion and accountability in decision-making.
  4. Proposing design pathways for LLM-based tools tailored to public sector needs.
• Procedure and key techniques:
  • Conducted 9 interviews (45–90 minutes each) and 2 workshops (3 hours each) with 10 caseworkers from a Finnish public institution.
  • Used thematic analysis to identify challenges, perspectives, and potential uses of LLMs in decision-making.
  • Iteratively developed findings through participant feedback and qualitative coding.

Results

• Concrete findings:
  • Decision-making involves four steps: case comprehension, information search, discretion use, and case solution.
  • Caseworkers face challenges such as scattered information, language translation, and adapting administrative guidelines to unique cases.
  • LLMs could assist in:
    • Translating between languages and administrative/spoken language.
    • Retrieving and summarizing relevant, up-to-date guidelines.
    • Confirming simple questions or providing prompts for overlooked information.
• Advantage over baselines:
  • LLMs offer potential for nuanced assistance in sub-tasks that require linguistic and contextual flexibility, unlike traditional rule-based or machine learning tools.
  • Caseworkers emphasized the importance of LLMs supporting, not replacing, their discretion and judgment.
• Experiments / evaluation:
  • Data collected from diverse departments (customer service, benefit handling, debt collection).
  • Workshops facilitated idea generation and validation of findings.
  • Caseworkers’ prior experience with generative AI influenced their perceptions of potential tool uses.
• Limitations and future work:
  • Small sample size (10 participants) limits generalizability.
  • Study focused on a single Finnish public institution; results may differ in other cultural or organizational contexts.
  • Future work should explore cross-institutional comparisons, AI literacy impacts, and ethical considerations in LLM deployment.

Summary

This study explores how LLMs can assist public sector caseworkers in their decision-making processes, emphasizing that decision-making is a multi-step process requiring discretion and contextual understanding. Caseworkers identified potential uses for LLMs in tasks like translation, information retrieval, and administrative-to-spoken language conversion, while stressing the importance of retaining human judgment and accountability. The findings provide actionable insights for designing LLM-based tools that align with the nuanced needs of public sector workflows, paving the way for more effective and ethical AI integration.