How should AI agents help people manage benefits? (Part 2)

This article explores what responsible use of AI agents looks like in the benefits ecosystem, evaluating past applications, challenges agents can genuinely help with, limitations for implementation, and how to ensure users maintain control of their cases. This is part 2 of a series about AI agents in the safety net. You can find part 1 here.

More than a decade ago, in 2015, Intuit's TurboTax started submitting SNAP applications on behalf of its users. The company had built a feature called “ Benefit Assist” to help low-income filers screen for food stamp eligibility while doing their taxes (a great idea!). Without public sector coordination, Intuit chose to list itself as the "Authorized Representative" on the applications and submit them via fax. States received applications filled with nothing but special characters and symbols. The Food and Nutrition Service (FNS) had to issue emergency guidance on how to handle the mess. Intuit shut the system down by January 30th, leaving countless families in the dark.

Though Intuit dedicated a creative and talented technology team to help consumers smoothly access SNAP by leveraging their own tax data, the interface was murky, and lacked coordination across fragmented state programs. The attempt resulted in confused families, overburdened agencies, and a missed opportunity to leverage automation for good. We’re now in a similar moment: AI agents – the software that navigates websites, fills out forms, and completes multi-step workflows autonomously – will begin to rapidly roll out to tens of millions of Americans. With this comes an opportunity to transform the safety net, improve customer experience and reduce administrative burden for consumers and state workers.

In Part 1 of this series, we made the case for why agents could help build a stronger social safety net, and tested what they can do today:

Americans spend millions of hours each month filling out paperwork for SNAP and Medicaid alone. These hours are often confusing, stressful, and high stakes; a single mistake can derail getting food on the table or critical medical care. During the post-pandemic Medicaid "unwinding", 17 million people lost Medicaid coverage for such “procedural” reasons alone. AI agents could ease much of this undue burden and stress, while giving states streamlined applications, better inputs, and better outcomes.

We then tested if the leading AI agent products could meaningfully help people apply for SNAP. Here's what we found:

[The agents] found the right websites, the right program, and successfully started the SNAP application on our behalf. At their best, they successfully navigated complex websites, uploaded documents, and proactively identified key discrepancies in our application. But they were also presumptuous, entered incorrect information, and agreed to legal terms on behalf of the user.

The agents were able to successfully navigate through the process by finding the right websites, programs, and applications. The agents also entered applicant information wrong and agreed to legal terms on behalf of the user. Neither of which help solve the problem of high stakes, high burden paperwork. They were impressive but not yet fully up to the task. Helping people navigate safety net services demands a much higher bar of accuracy and reliability than the products delivered. In this post, we're going to explore how they should work in the future so we can deliver on the opportunity and avoid the pitfalls of the past.

The Federal Government supports use of AI but states lack operational guidance#the-federal-government-supports-use-of-ai-but-states-lack-operational-guidance

America's safety net is composed of large, federally funded and governed, state-administered programs like SNAP, TANF, and Medicaid. States look to the federal government for guidance on how to administer them. Currently, the federal government is aggressively promoting and supporting AI adoption throughout the government. The March 2026 White House National Policy Framework for AI directs states to "...not unduly burden Americans' use of AI for activity that would be lawful if performed without AI." They then offered a more specific positive vision a few months later in Winning the Race: America's AI Action Plan: "With AI tools in use, the Federal government can serve the public with far greater efficiency and effectiveness…Taken together, transformative use of AI can help deliver the highly responsive government the American people expect and deserve."

The policy landscape is putting further pressure on states to adopt new technology. Under H.R.1, states must now cover 75 percent of SNAP's administrative costs, up from 50 percent, and almost half of states may face $100 million+ in benefit cost-sharing. Human services agencies are simultaneously facing staffing shortages across programs, such as the nearly 400 open positions at Pennsylvania DHS alone. Arizona's SNAP caseload dropped by nearly half since H.R.1, marking a true operational crisis. Agentic AI powering both caseworkers and beneficiaries provides a compelling operational strategy to help states combat these new burdens.

FNA is further encouraging states to adopt advanced automation and AI to address these challenges:

"These advanced automation technologies could be useful tools in assisting State agencies with workload and staffing challenges and improving customer service. FNS encourages and supports State agencies’ use of advanced automation technologies to enhance the administration of SNAP and foster public trust, both in SNAP and in the State agencies’ systems."

Despite enthusiasm from the top and immense pressure on states to do more with less, states lack specific guidance about exactly where AI agents fit in.

What role should agents play? #what-role-should-agents-play

How should an AI agent help someone manage their benefits? What exactly should they do, or more importantly, not do? For example, should AI agents fill out and submit SNAP applications? Should voice agents conduct eligibility interviews? Should they monitor payroll data and report changes to the state? Much of this is technically within reach today, but we lack a principled framework to answer such questions.

Fortunately, people have been helping other people navigate complex government services for decades, and there are many legal and operational precedents that we can look to. Surveying many government programs, there are three distinct categories of third-party helpers:

1. Assisters who explain programs, help gather documents, and prepare applications, with no legal liability of their own. Two common examples are SNAP outreach workers or VITA tax volunteers.
2. Preparers who take on accountability and legal liability for what they submit. These include tax return preparers or ACA navigators and brokers.
3. Representatives who act fully in place of the applicant and absorb the legal risk that comes with it. These include SNAP Authorized Representatives or SSA Appointed Representatives.

Each role reflects a different point on the spectrum of accountability and legal liability. Under current U.S. law, AI agents can't incur liability, which leaves us a clear starting point: AI Agents as Assisters. The Assister model has worked for decades in a variety of government programs, which makes it a promising framework to extend into this new domain.

Building a standard for benefits navigation agents#building-a-standard-for-benefits-navigation-agents

If agents are to become the next generation of safety net Assisters, we'll need a common understanding of what they should do and how they should do it. Toward this end, we've started work on a public "Benefits Navigation Agent Standard," a framework for beneficiary consent, safe interfaces within eligibility systems, and auditable interactions. We believe the right standards can make agent-assisted cases the best cases for both beneficiaries and states. Inspired by other voluntary public-private industry standards such as the General Transit Feed Specification and Financial Data Exchange, we will join a cross-sector coalition of states, navigators, and technologists to design standards for AI agents who interface with the safety net. To start, we've crafted a set of minimally viable design principles for industry and governments engaging with agents in the safety net:

1. Identify and track AI-assisted cases. Find a way to identify the application as AI-assisted. Some systems like BenefitsCal support this natively, while others will require creative workarounds, like uploading explanatory cover pages. Agents should be transparent and should not impersonate the user. Governments should track performance and outcomes of these cases for any variations in administrative denials or emerging issues.
2. Assist, don't sign. Users must have the opportunity to understand the legal commitments and risks they are accepting. Agents should never sign or check legal acknowledgments unless the user also sees the complete text.
3. Enable users to maintain control of their accounts. Even if agents help create or rely on user accounts, the end user should always be able to access their accounts directly. Users may revoke agent access at any time, and this should not obstruct their ability to manage their case.
4. Ensure users review information that impacts their eligibility. Modern SNAP applications and periodic reports ask many ancillary questions that do not impact eligibility, which users may reasonably want to skip. But if any piece of information may impact eligibility or benefit amount, users should have the chance to review it. The appropriate role for agents is to generate meticulous pre-populated forms for users to review and verify (as is already common with Medicaid and SNAP renewals).

Where we go next#where-we-go-next

With these early principles, we have an even longer set of open questions without clear answers yet. Questions such as: Should AI assisted cases be treated identically to human-assisted cases? What's the right technical approach to manage access and consent?

AI agents will spread through the safety net with or without common standards. Without them, we risk repeating the Benefit Assist debacle of the last decade; a well-intended but uncoordinated effort stressed states and confused consumers. With them, the same technology promises more efficient program operations, easier access for families, and a stronger safety net. That's why we're starting working toward an open standard. If you're building agents for the safety net, shaping a standard, running a company that helps navigate benefits, or working on policy, we'd love to hear from you.

Want to help design a Benefits Navigation Agent Standard? Have ideas or questions? We're part of a growing coalition of states, navigators, and technologists. Reach out to alicia.rouault@joinpropel.com.