CDT maps where AI can help and harm people navigating public benefits

Why benefits go unused

CDT opens by describing public benefits as some of the most effective tools the United States has for reducing poverty. It cites Social Security, which served 28.7 million people in 2024, the Earned Income Tax Credit (EITC), which served 4.4 million, and the Supplemental Nutrition Assistance Program (SNAP), which served 3.6 million.

Yet the programs are not always used by those who need them. According to the latest available SNAP data cited by CDT, only 88% of eligible individuals received SNAP benefits. For Temporary Assistance for Needy Families (TANF), only around 22 in 100 eligible families in poverty receive support.

CDT says researchers have already mapped how governments and vendors build technology into benefits systems, but that fewer resources exist from an applicant's perspective. The new piece is meant to fill that gap. It is organized around five phases: awareness, eligibility, enrollment, use and maintenance. It also covers risks that span all phases.

Awareness and eligibility

In the awareness phase, CDT points to siloed programs spread across dozens of agencies with no single front door. Awareness gaps are often worst among immigrants, the elderly and people going through a major life change, who may face language or literacy barriers.

On the AI side, CDT lists tools that flag people as likely eligible for programs they have not applied to, based on data from programs they are already in. It cites a 2024 SNAP technology grant from the Department of Agriculture to Dallas County, Texas, for software to identify potentially eligible people. Multilingual chatbots and SMS tools are described as available around the clock. Many states moving to unified Integrated Eligibility and Enrollment (IEE) systems are exploring AI to suggest programs and translate information.

For eligibility, CDT notes that rules are complex, differ across programs and change yearly. It says the 2025 federal expansion of SNAP work requirements caused a tremendous amount of confusion as states implemented changes with limited guidance. Calculators such as New York's Access NYC and Colorado's MyFriendBen can flag eligibility across dozens of programs before someone applies. CDT says many rely on AI-powered infrastructure, for example the Axiom Foundation, which uses AI to process and encode regulations as code. Tennessee's Department of Human Services uses a virtual assistant called Ava to answer common questions about SNAP, TANF and other programs before redirecting people to a caseworker if needed.

Enrollment and use

Enrollment is where application friction and missed interviews cause trouble. CDT says SNAP applications and recertifications require an interview within 30 days of submission, and over 40% of Missouri's rejected SNAP applications in 2024 were due to failure to complete it.

Proposed AI uses include flagging missing information before submission, carrying data across programs, and extracting names, dates and addresses from uploaded documents to pre-populate forms. CDT also describes predicting who is at risk of missing an appointment and targeting outreach by text, email or mail. It says this is not widely used in public benefits yet, but is well established in adjacent settings such as predicting outpatient no-shows.

In the use phase, CDT highlights that recipients often cannot easily see their balance and can fall victim to card skimming and fraud, often discovering it only after benefits are gone. Opportunities include AI voice agents that answer Electric Benefits Transfer (EBT) balance questions, and a meal-planning assistant that Propel built on top of LLMs to help SNAP recipients make affordable meals from what they have. Some apps, including Propel, already offer suspicious transaction monitoring that uses AI to flag higher-risk transactions. CDT notes Mastercard has recommended AI to learn a cardholder's normal purchasing patterns and flag anomalies.

What it means for applicants and agencies

The text provided for this report ends as it begins the fifth phase, maintenance, which covers ongoing renewal and reporting, so CDT's findings on that phase and on cross-cutting risks are not covered here.

What the resource does establish is a framing: AI tools are being tested at nearly every step between a person's first awareness of a benefit and their day-to-day use of it. For applicants, CDT's examples suggest the potential gains are in earlier discovery, clearer eligibility answers, fewer missed deadlines and better protection against fraud. For agencies and vendors, the applicant-side view offers a way to judge whether a tool addresses a real barrier, such as language gaps or missed interviews, rather than only improving internal efficiency.