–
In East Boston, Link Health works with hundreds of patients, helping them to enroll in public benefits programs, like SNAP. Most of them are Spanish-speakers. Arianna Doss, a Patient Navigator and student at MIT, has been working with Link Health for about one year. For her, one of the biggest barriers to enrollment is the language barrier.
From a distance, this looks like an ideal use case for artificial intelligence. AI can be available 24/7. It can process rules quickly, translate languages, and scale far beyond the capacity of any nonprofit call center or volunteer program.
Public benefits systems are notoriously complex. Eligibility rules vary by program, by state, by household composition, by residency requirements, by income fluctuations that don’t neatly align with monthly reporting periods. Applications are long, documentation-heavy, and written in language that assumes a level of time, stability, and bureaucratic fluency many people simply don’t have.
In theory, as eligibility rules shift, especially around work requirements and documentation requirements, AI can help surface up-to-date information quickly, reducing confusion for both volunteers and families navigating an already unstable system.
However, for Doss and other Patient Navigators, AI tools often fall short of their expectations. The chatbots were “only helpful when doing something very simple,” where a quick Google search or consultation with a supervisor would have sufficed.
How chatbots are supposed to help, and where they fall short
At Link Health, student volunteers regularly help people access benefits like SNAP, WIC, Medicare Savings Programs, and Lifeline. These are programs that already exist yet go unused by millions of eligible people. These are students who care deeply about equity and access. They’re trained, motivated, and supported by a nonprofit that has invested in tools meant to make their work easier, including AI-powered chatbots designed to answer eligibility questions and guide enrollment.
For student volunteers, chatbots are meant to serve as instant policy references: a way to quickly answer questions like “Does a part-time worker qualify for SNAP in Massachusetts?” or “Can a senior receive both MSP and SNAP?” The idea is consistency, accuracy, and reduced cognitive load.
But in practice, many volunteers default to Google or official agency websites. Chatbots, by contrast, can feel opaque. “Immigration policy is always changing,” Doss shared as an example. “I don’t think AI is ever up-to-date enough to be able to answer those kinds of questions.”
For patients attempting to enroll in these programs, chatbots are designed to lower barriers. They offer privacy and convenience. But many patients disengage early, or are distrustful of AI altogether. Enrollment is a deeply private matter, and it can be emotionally and cognitively taxing.
Designing for reality
Despite these challenges, there are moments when technology can genuinely help.
Chatbots can perform well for narrow, well-defined tasks: quick eligibility pre-screens, plain-language explanations, and translation. Most importantly, AI works best when it is paired with humans. When a chatbot helps someone get 60% of the way there – and a person helps them cross the finish line – follow-through can improve.
If nonprofits want AI tools that people actually use, design has to center on real behavior. For volunteers, this means integrating chatbots into existing workflows rather than treating them as separate platforms. For patients, it means shorter interactions, focusing on one benefit at a time rather than overwhelming users with comprehensive but exhausting pathways. It means offering human help early and often, not as a last resort.
The goal of AI in public benefits should not be to replace people. AI can handle repetition, complexity, and scale. Humans build trust, navigate nuance, and support people through moments of vulnerability. AI is powerful, but we must design systems that acknowledge how people actually live, decide, and engage.
Author: Karie Shen
