Policy & Practice | Fall 2025

The Challenge: Limited Levers in a High-Stakes Environment State agencies have tradition

What is Agentic AI? Agentic AI refers to AI systems that can operate autonomously to achieve specific goals, often by orchestrating a series of tasks or decisions. Unlike tra ditional workflows or pipelines, which follow predefined rules and sequences, Agentic AI can adapt to changing conditions, learn from new data, and make context-aware decisions. Some state agencies already have AI agents today that complete targeted tasks very well, whether they are genera tive AI agents or traditional AI agents. Now, imagine if these agents could work together as a team to accomplish a larger task, guided by Gen AI agents trained in orchestration. This concept of orchestration is what sets Agentic AI apart from the AI agents organiza tions already have. It combines the structured logic of workflows with the flexibility and intelligence of advanced AI models to deliver robust and dynamic solutions. In human services, we automated the straightforward cases more than a decade ago, building in safeguards to pause automation whenever human judgment was needed. But for those complex cases, we still rely on agency staff for answering the question: What should I do now? Agentic AI could be used to manage complex eligibility determinations. Consider a case involving a family with fluctuating income, multiple benefit programs, and recent changes in household com position. Within an Agentic AI system, one agent could gather relevant data from various sources, another agent could interpret policy rules, with addi tional agents to assess eligibility across programs, and others to generate personalized communication for the applicant, all without human interven tion. The agents would document their work along the way, ensuring transpar ency and traceability. This level of automation not only reduces the burden on caseworkers but also improves consistency, reduces errors, and accelerates service delivery. OBBBA imposes a gradu ated penalty on states with error rates higher than 6 percent, requiring them to fund a portion of the outgoing Supplemental Nutrition Assistance Program (SNAP) benefits for their

after initial enhancements have been made. Furthermore, policy changes are typically slow and can be complicated by political dynamics, delaying the implementation of necessary reforms. As a result, the traditional methods are becoming less effective in addressing the growing demands and complexities faced by public services agencies. Technology, by contrast, offers a dynamic and rapidly evolving toolkit. AI, in particular, is reshaping what is possible in human services. In a world that is more online than ever, administering human services benefits continues to be a paper-heavy process. Document processing is one of these AI-powered tools that has quickly pro gressed from simple optical character recognition, to solutions that can com prehend and classify unstructured data sources (even handwriting!). Without manual data entry, these tools take your paper applications into a format that enables the same downstream automation as electronic applications. From automating routine tasks to providing predictive insights, AI is enabling agencies to stretch their limited resources further. But not all AI is created equal, and not all implemen tations yield the desired outcomes. To truly move the needle, agencies must look beyond isolated automation solu tions and embrace a cohesive, strategic approach to technology adoption. The Evolution Toward One-Touch Processing The concept of one-touch eligibility processing—where an application is received, evaluated, and approved or denied with minimal human interven tion—has long been a goal in human services. For years, agencies have been able to automate simple cases, such as those involving straightforward income verification or categorical eligibility. However, life is rarely simple, and many applicants present complex circum stances that defy easy categorization. Each advancement in AI brings us closer to expanding the range of cases that can be handled through automa tion. Natural language processing, machine learning, and predictive analytics have all contributed to this progress, but the latest frontier, Agentic AI, promises to take us even further.

ally relied on three primary levers to improve service delivery: people, process, and policy. While optimizing these areas remains important, many agencies have exhausted the major relief points. Hiring and retaining skilled staff is increasingly difficult due to labor market constraints and budgetary limitations. Students in college today are learning an entirely different way of working. If we thought millennials were turned off by mainframe green screens, Gen Z and Gen Alpha are going to be even less accepting of tedious data entry and manual paperwork. Process improvements, while benefi cial, often yield only marginal gains

Lauren Aaronson is the Health and Human Services Technology and Innovation Lead at Amazon Web Services (AWS).

Eyal Darmon is the Public Service Data, AI, and Gen AI Lead at Accenture.

Heidi Reed is the Integrated Eligibility Lead at Accenture.

Policy & Practice Fall 2025 32