Beyond Chatbots: Where AI Can and Cannot Go in State Benefits Administration

BLUF — Bottom Line Up Front

The Center for Civic Futures (CCF) has awarded $11 million to 11 projects in the second round of its Public Benefit Innovation Fund (PBIF), backed by anchor funder Ballmer Group along with the Gates Foundation, Recoding America and Arnold Ventures. The round deliberately moves money away from applicant-facing chatbots, which made up close to half of first-round applications, toward the administrative work behind eligibility: SNAP recertification, unemployment insurance application completeness, error prediction, legacy system mapping, procurement and computer-readable eligibility rules. For state human services executives, the practical test for any benefits AI proposal is which side of the SNAP certification function it sits on, because federal regulation reserves certification to state merit-system personnel and prior federal guidance treats bots as non-merit staff.

What the First Round Revealed

PBIF launched in 2025 with $8.5 million across seven projects focused on the immediate effects of changing federal reporting requirements. CCF opened the open call for the second round on April 13, 2026, and announced the 11 awards on October 22, 2026, bringing its commitment to nearly $20 million across 18 projects, according to the Center for Civic Futures.

CCF reports early first-round results. A SNAP work requirement verification and change reporting tool that Promise launched in Mississippi helped the state reduce its payment error liability by $40 million. New Jersey's AI tools have, combined, helped deliver an estimated $19 million in public benefits to more than 97,000 residents, against the state's goal of $24 million for more than 203,000 people. These are CCF's own reported figures, not independent evaluations, and they come from first-round projects aimed at federal reporting changes rather than the backend focus of the second round.

The first-round application pool also surfaced a mismatch. "Close to half of the applications that came in were [to develop] chatbots," Cassandra Madison, executive director of the Center for Civic Futures, told StateScoop. "If you are a builder and want to build something for the public good, but you don't really understand what the day-to-day struggles of government are, that's the easiest thing to go after."

A chatbot for benefits applicants is a visible deliverable, and it addresses a real problem: applications are long, instructions are unclear, and eligible residents abandon them. It is not, however, where caseworker capacity goes. Madison framed the second round around that gap, telling StateScoop that CCF "very purposely landed an open call that was much more backend-focused" because of how much time the average caseworker spends on paperwork instead of supporting people one-on-one. Eleanor Davis, director of PBIF, made the complementary point that civic technology has historically concentrated on the applicant's experience while giving less attention to the systems government employees use to process applications.

"There's infinite reasons why that backend work is really hard to do," Davis told StateScoop. "And now is a critical moment to be driving innovation towards where we can leverage AI to make a huge difference in these spaces where we have not had as much opportunity to improve service delivery."

What the $11 Million Is Funding

The funded project list gives a concrete map of where AI interventions are being tested. Grants run from six months to two years; some stay in controlled testing environments while others are designed for pilots with real systems and users, and Davis told StateScoop that everything will be open source so other states can reuse the tools and lessons.

North Carolina will develop and test how AI can assist SNAP recertification, a project CCF titles "Agent-to-Agent Service Application." It targets the steps in recertification that are most burdensome and prone to error, and it uses synthetic data during the pilot to enable rigorous evaluation before any launch. Starting with synthetic data is the right sequencing for work that touches personally identifiable information and program eligibility; Madison noted that Medicaid and SNAP verification relies on names, Social Security numbers, dates of birth, addresses and phone numbers.

Stanford RegLab, in partnership with Maryland and Colorado, will assess whether AI can spot missing or unclear information in unemployment insurance applications at the time of application. The design point is that the tool works on the application, not the decision: catching gaps at intake should reduce the back-and-forth between agency staff and claimants that slows processing.

Other funded projects include predictive modeling by Yale's SCALE Lab and the Better Government Lab to identify which SNAP cases are most likely to contain errors, working with five states; a Recidiviz data bridge connecting state corrections data with Medicaid and SNAP systems; Nava PBC tools intended to cut the time to map legacy systems from months to weeks; and an Open Contracting Partnership effort to help state Medicaid agencies build more flexible technology contracts. Davis described procurement and rules work as "upstream barriers," adding: "You can't test and launch an amazing new AI tool that's going to solve all your problems if you can't even buy it."

The Axiom Foundation project on SNAP eligibility rules is worth noting separately. It will test open encodings of SNAP eligibility rules, written as software that people can read and computers can run, with each provision linked to the law it comes from. SNAP eligibility currently lives in regulatory and administrative language. Converting it into structured, traceable form is a prerequisite for a range of automation and requires close collaboration between agency policy staff and technical teams. Funding that foundation directly is different from funding a tool that assumes the structure already exists.

The Federal Constraint That Procurement Teams Miss

The second round's emphasis lines up with a federal regulatory boundary that shapes every SNAP AI deployment: merit staffing.

Under 7 CFR 272.4(a), state agency personnel used in the SNAP certification process must be employed under a merit system of personnel administration, and volunteers and other non-state-agency employees may not conduct certification interviews or certify applicants. The regulation lists narrow exceptions (SSI households, households in rural Alaska, and disaster programs) and expressly encourages states to use volunteers for outreach, prescreening, helping applicants through the application and certification process, and securing needed verification. That last provision is the regulatory outline of the pre-determination space.

The American Public Human Services Association's November 2024 analysis of AI-powered SNAP modernization summarizes the practical effect: only merit staff can perform eligibility and certification functions, including interviews, eligibility determination, and accepting and processing appeals, complaints or reported changes that could affect eligibility or benefit levels; and in prior federal guidance and interpretation, bots are considered non-merit staff and prohibited from performing those functions. APHSA noted that there was not yet federal guidance specifically detailing what AI can and cannot do in a state SNAP agency, but that the merit-staffing rules "will undoubtedly influence allowability of AI-enabled solutions."

Two further points from that analysis matter for procurement. First, FNS's January 2024 advanced-automation guidance, as APHSA summarized it, required any use of AI to be approved as a waiver or demonstration project, with some advanced automation and robotic process automation handled as a major change. Second, even functions that non-merit staff may perform, such as scanning documents or data matching that requires no household contact, require at least a major-change notification when done with AI. FNS has since published a revised version of its advanced-automation guidance; agencies should scope projects against the current text, not a 2024 summary.

The practical implication is that SNAP AI tools must be designed so that certification is performed by a merit-staff caseworker, not the AI. The AI can prepare, flag, retrieve and summarize; the decision belongs to the employee. That boundary is set in regulation, and changing it would take rulemaking or a waiver or demonstration approval from FNS, not a vendor's product roadmap. States should plan as if it holds.

This boundary is not always visible in vendor demonstrations. A system that "automates eligibility determination" and a system that "automates document collection to support caseworker determination" may look similar in a demo and carry very different compliance profiles. State procurement teams without benefits operations expertise can miss the distinction.

Where AI Has Space to Work

The PBIF projects and the federal rules together trace the shape of the pre-determination territory:

Recertification workflows. Much recertification work involves collecting and organizing documentation that households return. A tool that generates renewal notices, organizes returned documents and flags missing information before a caseworker reviews the case sits upstream of certification. It still needs the FNS notification or approval path that applies to its function.
Application completeness. The Stanford RegLab project targets the gap between application receipt and staff review. An application missing income documentation or with unclear household composition needs follow-up before a decision can proceed; identifying those gaps at intake reduces time per case without touching the determination.
Policy navigation for caseworkers. In May 2026, Code for America announced a partnership with Anthropic whose initial tool, the SNAP Policy Navigator, aims to give caseworkers real-time access to federal, state and county SNAP guidance. This is worker-assist AI: it helps the merit-staff employee find which rules apply to the case in front of them. APHSA's summary of the USDA framework lists AI-enabled search tools that direct caseworkers to the relevant section of an official policy manual among uses not inherently rights- or safety-impacting.
Legacy system mapping. Many state human services systems carry undocumented integrations, data schemas that have drifted from their documentation, and business logic buried in legacy code or vendor-specific configuration. AI-assisted documentation of what a system actually does, the focus of the Nava project, reduces the risk that a migration breaks something nobody knew existed.
Error pattern identification. Predicting which cases are likely to contain errors, as the Yale project does, is a review-prioritization function. It directs merit-staff attention; it does not make or change an eligibility decision. That distinction must survive into the workflow design, so a high-risk score triggers human review rather than an adverse action.

The Agentic AI Question

Madison warned that many states are "not yet ready" for the expected influx of AI agents, including agents that applicants might use to get through the benefits process, as reported by Route Fifty. States have "measured hopefulness around what AI can do, but healthy skepticism of how to do it well," she said, and are starting to experiment with agents in closed environments.

The merit-staffing boundary gives that caution a concrete design test. An agent operating in SNAP administration has to stay in pre-determination space. An agent that initiates a recertification document request, monitors for returned documents and routes completed cases to a caseworker queue has a different risk profile than an agent that changes eligibility status. North Carolina's agent-to-agent recertification project, tested on synthetic data first, is a useful early case to watch. The design requirement does not change with model capability: the certification must be made by a merit-staff employee, and the system must make that unambiguous and auditable.

H.R.1, the 2025 budget reconciliation law, requires states to prepare for more frequent Medicaid eligibility checks and new work requirements while shifting more responsibility for SNAP administrative costs and payment errors to states, as StateScoop summarized. That pressure raises the urgency to reduce caseworker workload. It does not move the boundary, and pressure to move quickly is the condition under which compliance gaps are most likely to appear.

The First Question Before Any Benefits AI Proposal

For state human services executives evaluating an AI proposal for SNAP or Medicaid administration, the first question is not which model, which vendor or what accuracy rate. It is which side of the certification function the system sits on. A short decision sequence follows from that:

Locate the function. Map each output of the proposed system to a step in the case lifecycle. If any output performs or substantively conducts an interview, determination, or the processing of a reported change, the analysis starts at 7 CFR 272.4 and the waiver and demonstration paths, not at the vendor's demo.
Identify the FNS path. For worker-assist functions, determine whether the use requires a major-change notification, explicit FNS approval, or a waiver, using FNS's current advanced-automation guidance.
Define the failure mode. Specify what happens when the tool is wrong: who sees the error, how a caseworker overrides it, and how the household is protected from an adverse action driven by a bad flag.
Test before live data. Require evaluation on synthetic or controlled data first, then validation on the population the tool will actually encounter, with results broken out by the case types and document types that matter for the program.
Decide what evidence would change the plan. If pilot results show caseworkers spend more time checking AI output than they save, or that error flags are concentrated in ways the agency cannot explain, the credible alternative is to fund the underlying work first: clean data, encoded rules and better contracts, the PBIF "upstream barriers," before adding another tool.

The PBIF second round, with its synthetic-data testing, error prediction framed as review prioritization, and investment in rules encoding and procurement, is a reasonable model for that sequence. Its open-source evaluations over the next six months to two years will show which of these bets reduce caseworker burden in practice. The merit-staffing constraint shaping the project designs is not a flaw in the approach; it is the boundary that makes benefits AI deployable at all.

Spartan X provides AI strategy, architecture and implementation advisory to state leadership, including Colorado's CDLE, OIT and the Governor's office. In benefits administration, that capability applies directly to the design questions above: where caseworker-assist AI can remove administrative burden without crossing into determination, and how to build oversight that stays valid as the systems evolve.