Using AI to Improve Resource Referrals and Reduce Caseworker Burden: Evidence from a Randomized Experiment

• In Event: AI in Public Sector Service Delivery

Abstract

Introduction/Background: Caseworkers serve as a critical point-of-contact for clients seeking assistance by assessing needs and connecting individuals to resources. These include “internal” resources delivered by caseworkers’ own organization (e.g., Goodwill) and “external” resources provided by community-based organizations (CBOs) and government agencies. Maintaining comprehensive, current knowledge across this complex landscape is challenging, and caseworkers facing pressure may default to heuristic-based decision-making, relying on familiar "go-to" referrals that may inadvertently overlook newer or better-fitting services. Identifying suitable resources also incurs considerable cognitive and administrative burden, exacerbating caseworker fatigue, burnout, and turnover — ultimately harming service quality and client outcomes. To address these challenges, we developed an open-source, generative AI-powered tool combining web search with a curated knowledge base of vetted resources to produce tailored referrals and plain-language action plans matched to each client's needs.

Purpose/Research Question: This study evaluates the impact of our AI-powered referral tool on caseworkers' perceived administrative burden, referral quality and speed, confidence, trust in AI and its use in case management, and burnout — with implications for responsible AI adoption in human services settings.

Data: We collected pre-post surveys from 67 caseworkers at two Goodwill affiliates (Texas and Pennsylvania). Surveys included validated measures of administrative burden (learning, compliance, psychological, and time costs), burnout, and trust in AI. In-depth interviews following the 4-month pilot (November 2025-February 2026) captured caseworkers’ experience, and log data tracked their usage and interactions with the tool.

Study Design and Methods: We conducted a randomized stepped-wedge experiment in which caseworkers received access to the AI tool at staggered intervals. Descriptive analyses of survey and log data have been completed, and inferential analyses using an intent-to-treat approach along with qualitative interview coding are currently underway.

Results/Findings: Descriptive analyses surfaced promising findings, especially for external referrals to CBOs and government agencies. Learning costs declined by 12%, compliance costs by 10%, and psychological costs by 20%, with overall external referral burden decreasing by 13%, all statistically significant. For internal referrals to Goodwill programs, however, burden reductions were less consistent and of smaller magnitude. Surprisingly, learning costs increased in some instances, likely reflecting short-term adjustment as caseworkers integrated a new tool into their workflows. Subgroup analyses revealed that “novice” caseworkers (0-2 years of experience) reported larger reductions in external referral burden, while “seasoned” caseworkers (>2 years) reported small but consistent burden reductions across both internal and external referral types. Despite these mixed patterns, 66% of caseworkers reported increased confidence in making high-quality referrals, and marginal increases in trust in AI were observed.

Conclusion/Implications: These findings have important implications for practitioners, policymakers, and researchers exploring AI adoption in human services. When aligned with existing systems and processes, AI tools can reduce administrative burden, support caseworker decision-making, and enhance service delivery. However, uneven effects across caseworker tenure and referral types highlight the importance of workforce development, change management, and human oversight. This study contributes timely evidence on how AI, deployed as a retrieval and decision-support tool, can augment rather than displace human judgment in complex service environments.

Author

• Yu-Fu Michael Chen, Nava PBC
  Presenting Author