$140 billion in federal benefits goes unclaimed every year, and the effort of finding out what you qualify for is a key reason. GAO
BennyFit screens a household against 21 federal and state benefit programs from a few plain-language sentences, under each state’s own SNAP, Medicaid and tax-credit rules. Benny, our agent, asks a follow-up only when the answer would change what the household gets.
No sign-up. Try it with a made-up household: BennyFit keeps nothing you type, but it is sent to our AI provider to be read.
The dollar figure is computed from the program rules applied to our 150 validation households, not from a model. Those households are synthetic, so it describes the validation set rather than the population. How this is measured.
Not because they are ineligible, and not because the money runs out. Because nobody ever told them, and because finding out is a day’s work.
in Earned Income Tax Credit goes unclaimed each year, by ~5 million eligible people. Tax Policy Center, citing TIGTA
of eligible non-participants are unaware they qualify. Finkelstein & Notowidigdo, NBER WP-18-03
In a randomized trial of about 30,000 likely-eligible households, pairing outreach with help completing the application moved enrollment from 11% to 18%. Finkelstein & Notowidigdo, NBER WP-18-03
That trial measured people helping people. It is the reasoning BennyFit is built on, not a result BennyFit has produced.
Not a chatbot walking someone through a form. Benny reads the whole situation, works out which rules are already settled by it, and finds the single unanswered fact that would move the outcome most, then says what it is worth.
For this household, thirty-one of the thirty-two rules were already decided by the paragraph. One was not.
In their own words, the way they would tell a person. No forms, no sequence of screens, no vocabulary to learn.
Every eligibility rule across all twenty-one programs is checked against that one description together, not program by program. Income, deductions and thresholds are calculated from the official tables, never estimated.
If something they did not mention would change the result, Benny asks about that one thing and explains why. If nothing would, it asks nothing and gives the answer.
Screening tools are easy to build and hard to trust. Everything here is designed so that a wrong answer would be visible.
Every dollar figure records the official page it came from and the date it was read. Nothing is transcribed from a summary or a third-party site.
Each result comes with the deduction-by-deduction arithmetic behind it and the specific rule that decided it, with the regulation cited.
Accuracy is reported against a validation set whose correct answers are worked out from the rules independently, and the method is published in full.
No database, no session, no retention of what a household types, and no analytics on the text of it.
Every claim below is something each product publishes about itself, linked to the page it came from. Two of them screen for more programs than we do, and the table says so.
| Single StopBenefits screener sold to colleges, credit unions and utilities | findhelpSocial care network with a benefits eligibility check | mReliefNonprofit SNAP screener, web and SMS | BennyFitThree programs, in full | |
|---|---|---|---|---|
| Programs screenedMore than any of them. The difference is what a program being "covered" means: these twenty-one are implemented against the regulations, with the deductions, the thresholds and the arithmetic behind every figure, and each applies the rules of the state the household actually lives in. | 20+ | 10 core federal and state programs | SNAP only | 21, each against its full rule set |
| Estimated dollar value for this householdmRelief is explicit that it does not estimate amounts "due to the complexity it would add to the screening questions in order to have an accurate estimate". That is the trade BennyFit is built to avoid: the arithmetic happens in code from the official tables, so an amount costs no extra questions. | Aggregate only: "$20K average annual support found" | Population average: "$2,800 annually" | None. States that amounts would add too many screening questions | Per household, per program, with the arithmetic shown |
| Questions asked | A questionnaire, about 15 minutes | "A few direct questions", 60–90 seconds | A short fixed set of yes/no and multiple choice | None, unless one would change the answer. Never more than three |
| Accuracy method publishedmRelief says it "indexes towards being more inclusive on who may be eligible, so as to generate as few false negatives as possible", a deliberate, reasonable choice that accepts false positives in exchange. BennyFit reports both rates separately so you can see which way it errs rather than having to assume. | 96% claimed; no method or validation set published | No accuracy figure published | No figure; states it deliberately skews toward over-inclusion | Both error directions reported, method and validation set published |
| Shows why it reached the answer | Referrals and an eligibility estimate | Programs you may qualify for | Likely qualify, and where to apply | The deciding rule, the deductions, and the regulation cited |
| Programs that qualify each other | Not described | Not described | Single program | Followed automatically: SNAP enrollment qualifies Lifeline |
| CoverageCovering every state and applying the same rules in each are different things. 44 of the 51 jurisdictions set a SNAP income limit above the federal 130%, and most impose no asset test at all; 8 have not expanded Medicaid; 33 run their own earned income credit. Those differences are read from USDA’s own chart, from KFF and from ITEP, so the same household gets a different and correct answer in Michigan than in Ohio. | Not stated on the screener page | All 50 states and DC | All 53 states and territories | All 50 states and DC, each with its own SNAP limits, Medicaid status and state credit |
Read from each product’s own pages on 21 September 2026. Where a figure is a marketing claim rather than a published measurement, the cell says so. Our own speed and cost per screening are still in validation and are not claimed here.
Benefits Data Trust shut down in August 2024 after twenty years, laying off all 273 employees. It had helped hundreds of thousands of people across seven states under contracts with state agencies and health insurers, and was one of the largest drivers of SNAP applications in Pennsylvania. The Philadelphia Inquirer
Screening people one at a time works. It has never scaled, and the funding for it is not reliable.
A calculator asks a fixed list of questions and stops. BennyFit reads the household description once, works out what it already knows, and then asks only about the facts that would actually change the result. Many households are asked nothing at all, and none is asked more than three questions.
Directly from the federal sources: the USDA cost-of-living memos and the SNAP regulations, the IRS revenue procedures for the Earned Income Tax Credit, the HHS poverty guidelines, and the USAC Lifeline tables. Every dollar figure records the page it was read from and the date it was read, and those documents are kept alongside the rules so any number can be traced back.
Every figure is stored against the period it applies to, so the system picks the table in force on the date of the screening. The SNAP fiscal year turns over on 1 October and the change takes effect on its own. When a new table has not been published yet, the screening says so on screen rather than quietly using a stale number.
No, and it never presents itself as one. Every result is labelled as a screening estimate and links to the program’s own application. SNAP, Medicaid and state tax credits are screened under the household’s own state rules. The other programs use federal limits, which some states and local agencies raise, so a household told it is unlikely to qualify for one of those may still qualify where it lives, and the result says so.
Nothing. There is no database, the screening keeps no session, and household descriptions are not retained or used for analytics. The validation set the accuracy figures are measured on is entirely synthetic.
As three separate things, because they fail in different ways. Whether the rules match the federal sources, checked against figures worked out by hand from USDA, IRS, Medicare and VA publications. Whether the system reads facts correctly from descriptions written the way people actually write, on a set held back from development. And whether it reaches the right answer overall, which needs the live service and is reported as pending until it has run. The full method is published, including where the system is weakest.
SNAP, the Earned Income Tax Credit and Lifeline, each against its full federal rule set rather than a simplified version. These three were chosen because they interact: qualifying for one can qualify a household for another, and BennyFit follows that automatically.
It works as a screening step in front of whatever intake you have. A household describes their situation once, and your team receives the programs they are likely to qualify for, the estimated annual value of each, and the specific reason behind each result.
Type a situation in your own words and watch every rule across twenty-one programs get checked at once, under your state’s own limits. It takes about thirty seconds.
Try the demo