AI-Powered Benefits Navigation System for SNAP Recipients

propel · Government · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Public benefits navigation

Propel wanted to help SNAP recipients understand and resolve missing benefit deposits quickly, without accidentally worsening delays or leaving users unsupported in high-stakes benefit interruptions.

Constraints:

• millions of monthly users
• benefit interruptions create food insecurity
• state-specific eligibility and process differences
• wrong advice can extend a user's wait for benefits
• small pilots required close monitoring and escalation

Architecture

Propel combines missing-deposit detection with tightly monitored help flows: affected users are notified, routed into either a state-specific triage form or knowledge-base chat tool, and escalated to humans when the AI cannot help safely.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  missing-deposit-detector["Missing-deposit detector"]
  mobile-notification["Mobile notification"]
  california-triage-flow["California triage flow"]
  decagon-chat-tool["Decagon chat tool"]
  knowledge-base["Knowledge base"]
  monitoring-dashboard["Monitoring dashboard"]
  human-escalation-team["Human escalation team"]
  missing-deposit-detector --> mobile-notification
  mobile-notification --> california-triage-flow
  mobile-notification --> decagon-chat-tool
  decagon-chat-tool --> knowledge-base
  decagon-chat-tool --> human-escalation-team
  decagon-chat-tool --> monitoring-dashboard
  monitoring-dashboard --> human-escalation-team
  california-triage-flow --> human-escalation-team
  human-escalation-team --> mobile-notification

Components

Data flow

The AI help layer sits on a knowledge base with real-time monitoring, escalation, and performance tracking, while small pilots keep affected users close to human support.

Why this architecture

The architecture is conservative because wrong advice can restart benefit timelines and make a missing-deposit problem worse.

Models used: generative AI platform · AI code generation

Scale

• The pilots were run at about 1,000 users each, with one CalFresh flow assisting about 1,300 recipients in a deposit cycle.
  “We ran these tests at a relatively small scale (on the order of ~1,000 users each) so that we could carefully monitor results and manually handle any escalations where necessary.”article, stated

  “We assisted about 1,300 CalFresh (California SNAP) recipients with this flow during one month’s deposit cycle.”article, stated
• Propel ran each missing-benefit pilot at roughly 1,000 users so results and escalations could be monitored closely.
  “We ran these tests at a relatively small scale (on the order of ~1,000 users each) so that we could carefully monitor results and manually handle any escalations where necessary.”article, stated
• The California flow assisted about 1,300 CalFresh recipients during one deposit cycle.
  “We assisted about 1,300 CalFresh (California SNAP) recipients with this flow during one month’s deposit cycle.”article, implied
• More than half of eligible users, 53%, used the help tools when offered.
  “more than half of eligible users (53%) used the help tools when offered”article, stated

Alternatives considered (roads not taken)

Broad release of benefit-navigation AI before monitoring

Rejected because: Propel recommends small pilots with close monitoring, iteration, and human escalation support.

User reapplication as quick self-service advice

Rejected because: The source warns reapplying can restart a 30-day timeline and extend benefit delays.

AI chat without escalation to a person

Rejected because: The tool escalates to Propel staff when it cannot safely resolve the user issue.

Patterns

rag-retrieval (retrieval)

The chat assistant layers generative help over a benefits knowledge base so users can receive state-aware guidance.

human-in-the-loop-gate (safety)

The pilot escalates unresolved or unsafe cases to Propel staff rather than forcing the AI to answer every user.

canary-rollout (operations)

Propel tested the AI assistance at small scale first so the team could monitor outcomes and handle escalations manually.

Tradeoffs

How to build a state-specific triage experience quickly.

Chose AI-generated code from a written logic tree with manual edits before launch over fully manual implementation of every branch.

Rationale: Propel found code generation useful for rapidly creating a complex multi-page form flow, while still manually reviewing the generated implementation.

How broadly to release benefit-navigation AI.

Chose small randomized pilots over immediate broad rollout to the full user base.

Rationale: The pilots let the team measure impact while keeping manual escalation feasible.

Gotchas (what bit them)

Fast but incorrect guidance can make a missing-benefit problem worse.

Root cause: Reapplying too early can restart a processing timeline even when the user is still eligible.

Resolution: Detect expected missing deposits and guide users through the right state-specific next step instead of a generic reapply instruction.

Public-benefit AI support needs live monitoring and a human backstop.

Root cause: The system operates in a high-stakes domain where the AI may not be able to resolve every issue safely.

Resolution: Start with monitored pilots and route hard cases to staff for direct outreach.

Evaluation

Propel used small randomized tests with control groups, monitoring uptake and benefit-restoration outcomes for around 1,000 eligible users in each pilot.

Metrics: over 5 million monthly users · about 200,000 users with monthly deposit interruptions · about 1,300 CalFresh recipients assisted with the California flow · 53% of eligible users used the offered help tools · modest but meaningful improvements in two key outcomes

Outcomes

• More than half of eligible users used the AI-assisted help when offered.
• Both pilots showed modest but meaningful improvements for benefit restoration.
• Propel identified multilingual personalized support as a promising way to scale benefits navigation.

The work is intentionally described as small-scale pilots rather than a full deployment across Propel's five million monthly users.