Propel in-app SNAP benefits assistant

Propel Inc., a private Brooklyn company, runs a free benefits app used by people on SNAP, the US food-benefit program. Its missed-deposit alert flags a benefit deposit that did not arrive, and its AI help tells the recipient why and what to do next. The tools read the state's deposit history but never write to it.

The app and the programs

SNAP stands for the Supplemental Nutrition Assistance Program. Benefits arrive as deposits into an Electronic Benefit Transfer (EBT) account.

More than five million people use Propel's app to check their balances for SNAP and two other public benefit programs, WIC and TANF. The app connects to state benefit portals with each user's permission.

The problem the tools target

Propel says about 200,000 of its monthly users have a benefit deposit interrupted each month. Most interruptions follow the periodic reports due about six months after applying, or the yearly recertification.

An interruption can end with benefits stopping and the person reapplying. This is program churn, and it is the failure Propel's AI tools target.

How the tools work

A detection algorithm reads the recipient's deposit history from the state benefit portal. It flags an expected deposit that did not arrive. The state verifies that history, so the signal is almost certainly right.

Two tools then help the recipient find out why, and what to do next. One is a multi-step self-diagnosis flow, whose code was written with AI and then edited by hand before launch. The other is an AI chat assistant. It answers from a curated state-by-state SNAP knowledge base, and it runs on another company's commercial customer-support platform.

When the chat detects that it cannot help, it hands the case to a named Propel staff member. That person contacts the user directly.

The 2025 pilots

In spring 2025, Propel ran two deliberately small pilots, each over one deposit cycle. The self-diagnosis flow served about 1,300 CalFresh recipients. CalFresh is California's SNAP. The chat helped about 1,000 SNAP recipients nationwide with missed deposits. Propel held the pilots to about that size so its team could monitor results and handle every escalation by hand.

Propel reports that 53 percent of eligible users offered the tools used them. It reports that randomized testing against a control group showed "modest but meaningful" improvements. They were in the days until the next deposit, and in the share of users restored in the same month. Propel did not publish the size of either effect.

The chat gave state-specific answers and recognized state-specific shorthand. It switched to Haitian Creole when a user began replying in it, which the team had not designed in advance. Users could rate conversations. Most did not, and among those who did, very few ratings were negative.

How far the tools have gone

As of mid-2026, no source documents a full rollout of the AI chat to the app's millions of users. The deployments the sources describe are the capped pilots and one pipeline built for the November 2025 federal shutdown.

That shutdown delayed SNAP payments to about 42 million recipients. Propel checked all fifty state SNAP agency websites every hour. A language model sorted real updates from formatting changes before people reviewed them.

Where the facts come from

Every outcome figure here comes from Propel itself. The case file found no independent evaluation, replication, or outside coverage of the pilots.

NPR independently reported the app's scale of about five million users. Propel says more than 25 percent of EBT cardholders nationwide use the app. As of July 2026, Apple's App Store listed about 574,000 ratings for it, averaging 4.9 out of 5.

What this case asks

Most cases in the Lab follow a mistake from an AI into a record, and back out into someone's decision. Here that loop is cut at the source, because the help tools never write to the deposit history. AI-written text is never saved into it and read back later as fact.

The risk moved somewhere quieter. It moved onto the staff who catch what the AI misses, a channel sized for a pilot. It also moved onto a gap: no one outside Propel checks any of it. No state agency, regulator, or independent auditor governs the tools.

What this network is drawn from

This network follows the pattern the Propel case file describes. It is not a reconstruction of the actual app. The people using the app are outside the network, which draws Propel's own support and escalation staff. It computes no restored benefits and no difference in who the tools help.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Closing a pathway means mistakes stop passing along it. The work along it goes on.

This case has a budget of 8 units. Each tool costs the same at every target level.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets are met before any tool is used or pressure added. Any single tool on offer also keeps them met.

Under Service and Safety Targets and All Governance Targets, the targets can be met within the budget. Both levels ask you to close every failure pathway, among other targets. Before any tool is used, three failure pathways are open: Results and hand-offs to staff, Alert reads deposit history, and Staff read deposit history.

Escalate checks closes the first. Mark AI-written records closes the other two. Every way of meeting these targets uses both tools, which cost 4 units together at their standard settings. Within the budget, 14 different sets of tools meet the targets, or 35 counting the stronger settings.

More is not better here. Using every tool on offer, each at its strongest setting, costs 31 units, nearly four times the budget. It meets the targets at none of the three levels that set them, because the help tools then add too little to the work.