Open-source AI assistant shows promise for California caseworkers’ service delivery

A pilot program to test an open-source AI tool aimed at enhancing the benefit application process for caseworkers and clients is garnering early positive feedback from local social workers in California. 

Developed by the public benefit corporation Nava and tech nonprofit Amplifi, the organization’s form-filling assistant aims to mitigate manual, paper-based processes that often stymie agencies’ work to connect residents with critical assistance programs. 

The tool’s development comes as states race to comply with federal rule changes to Medicaid and the Supplemental Nutrition Assistance Program. For the former program, new work requirements will largely impact how applications are designed and administered, and it will likely create increasing workloads for caseworkers. Benefit agencies are also poised to face burdens under the SNAP changes, as states’ share of program costs will increase based on their payment error rate to beneficiaries. 

Addressing such challenges is the target of the form filling assistant, which leverages generative artificial intelligence capabilities. A prototype of the product launched late last year for usability sessions, after the project received a $1.5 million grant from Google’s Generative AI Accelerator program in June. 

Now in the second phase of the pilot program, the form filling assistant is being leveraged by about a dozen staff members at the Riverside County Children and Families Commission,  according to Jillian Hammer, senior designer and researcher at Nava, who spoke during a webinar hosted by Nava last week. 

The generative AI agent aims to expedite the benefit application process for caseworkers and clients, increasing form accuracy and submissions while enabling staff to prioritize more meaningful tasks when assisting clients, she explained. 

While a caseworker meets with a resident to initiate the benefit application, the tool works by searching an organization’s databases and benefit systems to automatically fill in application forms, while allowing the worker to manually correct generated information as needed and ultimately review and approve the application. 

Social workers’ main priority is “helping families navigate [benefit] systems that can sometimes feel overwhelming for them at home,” Avila said. 

But because social workers interact with people that have varying degrees of familiarity with benefit programs, sometimes client interactions last from minutes to hours, and they can persist over multiple days if staff and applicants have to go back and forth to find the right information and documents. 

The AI tool streamlines the application completion and submission steps, improving workers’ flexibility to juggle dozens of cases a week, she said.

Using the form filling assistant helps reduce “the time spent manually [having] to fill out every single section of every single form, so it allows us to have more one-on-one rapport with the clients and focus on conversation and engagement,” Avila explained. For instance, staff can dedicate more time to educating themselves and their clients on additional assistance programs and resources within the community.  

Caseworkers’ positive experience with the AI agent reflects the nonprofits’ ongoing work to improve the performance and outcomes of the form filling assistant based on user feedback, Hammer said. 

The latest iteration of the tool, for example, does not display all of its processing details while it is running because staff reported that such visuals can distract caseworkers during conversations with a client “who deserves their undivided attention,” she said. The agent’s technical details now fall into an accordion design that users can opt into, if they’d like to view more in-depth data.  

The form filling assistant also generates flags on form fields that were filled by the AI tool and an explanation of why it pulled certain data, which encourages users to double-check applications before approving them for submission, Hammer said.

Another improvement was adjusting the tool to conduct an “upfront gap analysis” of an application form by scanning it and notifying the caseworker of missing information before automatically filling in the form, said Foad Green, a software engineer at Nava. This feature helps the caseworker and client jumpstart the process to retrieve the missing data, such as a medical record, instead of waiting for an interruption to occur. 

Those improvements have helped boost staff’s confidence in leveraging the AI agent, which in turn helps residents feel more comfortable with caseworkers’ use of AI products when dealing with sensitive personal information, Avila said. 

Indeed, Hammer noted that a major takeaway from the staff’s experience is that “an agentic AI-driven tool can improve the delivery of human services, but only when that tool is built upon the human expertise and the human relationships that are at the core of this work.”