Domain Atlas / Benefits navigation & public-facing chat
GetCalFresh: the nonprofit front door that carried most of California's online SNAP intake
Work with this case in the PAN Lab ↗
Between 2019 and 2025 more than 70% (about 73% per its ten-year retrospective) of California's online SNAP applications were submitted through GetCalFresh, a deterministic, structured-workflow application assister built and operated by the nonprofit Code for America, which reports helping 6.2 million people obtain more than $12.8 billion in food benefits from 2017 to 2025 (organization-published figures that are not independently audited); the node made no eligibility determinations, and in 2024 and 2025 the California Department of Social Services coordinated a dated, phased transfer of its functions into the state-owned BenefitsCal portal.[3]
What happened
GetCalFresh was a digital application assister for CalFresh, California's SNAP food-benefits program, built and operated by Code for America, a nonprofit, in partnership with the California Department of Social Services (CDSS) and county welfare departments. It is important to be precise about what it was and was not: it is not a machine-learning, generative, or scoring system, and it makes no eligibility decisions. It is a deterministic, structured-workflow tool - a mobile-first guided intake form (one question per page), a phone-camera document uploader, an SMS and email reminder scheduler keyed to application and renewal milestones, multilingual live chat, and a portal for community-based-organization (CBO) assisters. Completed applications are transmitted into county and state eligibility systems, where human caseworkers conduct the intake interview and make every determination; an applicant could bypass the node entirely by applying at a county office. Its influence on outcomes runs through completion, timing, and churn on the applicant side, never through a decision.
It launched in 2014 as a Code for America pilot with the San Francisco County Human Services Agency, initially in five counties, replacing a legacy online application that had spanned more than one hundred screens across three different county systems; it reached more than half of California's counties by 2018. On May 31, 2019, CDSS adopted GetCalFresh as the statewide application assister across all 58 counties, timed with a June 1, 2019 expansion of CalFresh eligibility to SSI recipients that made roughly 500,000 older adults and people with disabilities newly eligible. By Code for America's own account (an organization-published figure that predates the statewide era), application completion time had fallen from about 45 minutes to under 10.
The scale figures are Code for America's own, organization-published and not independently audited, and two different cumulative counts coexist that should not be conflated: the program page states GetCalFresh helped 6.2 million people obtain more than $12.8 billion in food benefits from 2017 to 2025, while the August 27, 2024 ten-year retrospective states that more than seven million Californians have used the service (users of the site, not people receiving benefits). Between 2019 and 2025, the organization reports, more than 70% (about 73% per the retrospective) of California's online SNAP applications were submitted through it. California's SNAP participation rate rose from 66% of eligible people in 2014 (fourth-lowest in the country) to 81% by 2022 - a trend Code for America cites alongside GetCalFresh's growth and presents as context, explicitly not as a proven causal effect.
Two measured completion levers are unusually well-documented for this domain, and they matter because they show the node's influence runs entirely through the applicant's path. First, an in-house experiment on renewal retention: clients received up to three reminders to submit the SAR 7 semi-annual renewal form, with about 60% of invited clients submitting, and a tested last-minute deadline reminder raised submissions among prior non-responders from about 1.5% to roughly 12% (a nearly eightfold increase), which once deployed organization-wide Code for America estimated would yield about 315 people renewing benefits each month; the organization reports these figures without sample sizes or confidence intervals. Second, a flexible-interviews pilot with Los Angeles County (launched October 2018, full access May 2021) was evaluated by a randomized controlled trial designed with Tatiana Homonoff (NYU) and Jason Somerville (New York Fed). Two effect framings coexist and should be cited with their definitions rather than averaged: Code for America's operational write-up reports 61% versus 58% approval (a 3-percentage-point intent-to-treat contrast) and 15 versus 19 days to determination, while the peer-reviewed version (Giannella, Homonoff, Rino, and Somerville, American Economic Journal: Economic Policy 16(4), 2024, DOI 10.1257/pol.20220701; NBER Working Paper 31239, 2023), using roughly 65,000 Los Angeles applicants, found that access to applicant-initiated flexible interviews increased approvals by 6 percentage points, doubled early approvals, and increased long-term participation by over 2 percentage points, identifying the intake interview as a key procedural-denial barrier. Code for America estimated the LA change would mean about 10,000 additional households a year, with gains concentrated among applicants from county offices with the lowest baseline approval rates.
The wind-down was, by design, the opposite of a collapse. CDSS documented a dated, phased sunset of GetCalFresh's functions: the SSA joint-application tool on September 30, 2024; the SAR 7 flow (including SAR 7 LaterDocs) on December 2, 2024; the application assister, CBO portal, and in-application document uploader on June 30, 2025; the standalone LaterDocs feature on September 3, 2025; and the data dashboards and CBO tools on September 30, 2025 - with digital advertising, informational landing pages, and multilingual live chat (answering thousands of texts and emails per month statewide) persisting, and applicants redirected to BenefitsCal.com. GetCalFresh's design principles - plain language, mobile-first design, streamlined document uploads - were embedded into BenefitsCal, California's permanent state-owned multi-benefit portal (rolled out 2021-2023), making the wind-down an intentional transfer of the navigation function into government-owned infrastructure rather than a discontinuation. The live counterfactual for what an ungoverned removal of such a node looks like is Benefits Data Trust, a Philadelphia benefits-navigation nonprofit that had delivered more than $10 billion in benefits since 2005 and served more than 120,000 people in 2023, and that abruptly announced its closure in June 2024; its SNAP partnerships with seven states ended and the fate of its screening tools, chatbots, and call center was left uncertain - an analogical contrast about a different organization with a different funding structure, not evidence about GetCalFresh itself.
The sociotechnical reading
The Atlas is full of benefits systems whose danger is a decision: a fraud score, an eligibility cutoff, a risk tier, a generated answer someone might act on. GetCalFresh is the instructive inversion, and the lesson is about criticality without authority. Here was a node that came to carry more than 70% of a state program's online intake - by any measure critical public infrastructure - and it held no authority over any outcome whatsoever. It scored nothing, decided nothing, predicted nothing. Every determination stayed with a county caseworker who conducted an independent interview and verified the application against administrative records; an applicant could route around the node entirely. On the system map that is a very specific shape: the load-bearing check is not a habit inside the user (the US copilots) or a supervisor inside the organization (Caddy) but a statutory determination floor drawn strong and placed entirely outside the node. That externality is the whole safety case. It is precisely because the node could not decide anything that it was safe to let most of a state's intake flow through one nonprofit web form - and it is why the measured levers here (a reminder that lifts lapsed renewals nearly eightfold, an interview-flexibility change that raises approvals by several points) act on completion friction, never on any judgment. Influence through the applicant's path, authority held elsewhere.
But criticality is criticality even when authority is absent, and that is the second half of the lesson. Concentrating a navigation function into a single voluntary node - sustained by philanthropy and a state partnership rather than statute - is a dependency risk that has nothing to do with the tool's accuracy, because there is no accuracy to audit. For six years there was no redundant intake path and no statutory backstop: if the node vanished, over 70% of online intake had nowhere to go. The Atlas carries the live counterfactual in the same domain - an intermediary nonprofit that abruptly announced its own closure, stranding partners with no successor. GetCalFresh is the governed answer to exactly that failure mode, and the contrast is the point. Its dependency was unwound not by a board vote in weeks but on a dated, five-phase schedule that transferred each function into a state-owned successor and embedded the node's design principles into permanent government infrastructure before the lights went off. So the governance instruments that matter for a node like this are not the ones the rest of the domain reaches for - no verifier, no risk tiering, no better model, because none of those touch a system that makes no decision. They are the succession instruments: a built backstop authorized deliberately and on the record, a dated hand-off schedule someone can be held to, and the protection of the statutory floor that made the concentration safe in the first place. The distinct lesson the Atlas draws here: the safest place to concentrate the throughput of a public program is a node that holds no authority over its outcomes - and the only governance that then matters is whether the concentrated dependency is unwound on a schedule with a built successor, or dropped as a cliff. A critical node you can safely run is one you must still plan to safely remove. The honest boundary throughout: served people, and the benefits they do or do not ultimately receive, are not modeled in the paired Lab; the completion and approval effects are served-outcome evidence measured outside any system diagram, the scale figures are organization-published and not independently audited, and this was never an AI system - any reading that implies algorithmic decision-making misreads it.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.