Customer support · Production
Propel prototypes AI tool to help SNAP recipients understand and act on notices
The problem
SNAP notices are frequently confusing to recipients due to legal language, cognitive overload from mandatory disclosures, and unclear urgency levels — causing eligible people to miss required actions and driving high unnecessary call volume to state agencies.
Workflow diagram · grounded in source
1
User submits SNAP notice
trigger
“we can enable people to easily take pictures of notices they receive on their phone (or screenshots from their portal) and share those”
2
LLM processes notice
ai_action
“You are a legal aid attorney specializing in SNAP benefits. You are given a SNAP notice and an optional question about it.”
3
Importance and action output
output
“Importance: HIGH importance
This notice is telling you that you missed an interview that's required to keep or start getting SNAP/food assistance benefits.”
4
User asks specific question
trigger
“we also think it's important for people to be able to ask specific questions about a notice”
5
LLM answers user question
ai_action
“Here's another example of a notice running through the tool, but this time with a real question from a user who posted this on social media”
Reported outcome
Early prototype tests with real SNAP notices show the AI tool provides more helpful output than reading the notice alone; the tool is not yet deployed to users.
Reported metrics
SNAP recipients calling agency when they have a question50%
Reported stack
Claude 3.5 SonnetStreamlitChatGPTGeminiMicrosoft Presidio
Frequently asked questions
What did this team achieve with this AI workflow?
Early prototype tests with real SNAP notices show the AI tool provides more helpful output than reading the notice alone; the tool is not yet deployed to users.
What tools did this team use?
Claude 3.5 Sonnet, Streamlit, ChatGPT, Gemini, Microsoft Presidio.
What results were reported?
SNAP recipients calling agency when they have a question: 50% (source-reported, not independently verified).
How is this customer support AI workflow structured?
User submits SNAP notice → LLM processes notice → Importance and action output → User asks specific question → LLM answers user question.