InnovateUS · AI and Public Health · Benefits access and delivery

What does AI tell residents about benefits?
Ariel Kennan · President, the Axiom Foundation
October 22, 2026 · Online
1

AI as a new front door to benefits

People and agentsask what the law saysAI modelTodayA guessfrom training dataand web searchWith AxiomEncoded lawopen, cited, executableThe rulewith its citation
If AI is going to do public good, it has to get the rules right.
2

How accurately AI models calculate household taxes and benefits

PolicyBench 50 models, 100 households, 18 outputs, snapshot October 14, 2026.
Model rankings by exact match: GPT-6 Sol 95.2%, Claude Opus 5.5 94.1%, GPT-5.6 Sol 93.1%, Claude Sonnet 5.5 91.7%, GPT-6 Astra 91.7%.
policybench.org, October 14, 2026 · no tools · PolicyEngine, which our CEO Max Ghenis also leads, builds PolicyBench and its reference answers3

AI answers for households that qualify for SNAP

42%

of AI answers told a household that qualifies for SNAP it would get $0.

268 of 643 answers · 50 AI models · 13 households that qualify

policybench.org, October 14, 2026 · 50 models, no tools · a household qualifies when PolicyEngine’s reference SNAP is above $04

Where models still break

Share of AI answers that got the amount right, by program

Right means within 10% of PolicyEngine’s reference amount,
or within $1 when the household is owed $0

Households owed moneyHouseholds owed $0
0%25%50%75%100%State refundable credits18 households owed moneyState refundable credits: 17.6% of answers within 10% for households owed money (18 households, 50 models)18%State refundable credits: 98.6% of answers right for households owed $0SNAP13 households owed moneySNAP: 36.7% of answers within 10% for households owed money (13 households, 50 models)37%SNAP: 96.9% of answers right for households owed $0Federal income tax36 households owed moneyFederal income tax: 45.6% of answers within 10% for households owed money (36 households, 50 models)46%Federal income tax: 93.8% of answers right for households owed $0State income tax36 households owed moneyState income tax: 47.8% of answers within 10% for households owed money (36 households, 50 models)48%State income tax: 91.1% of answers right for households owed $0Federal refundable credits13 households owed moneyFederal refundable credits: 52.4% of answers within 10% for households owed money (13 households, 50 models)52%Federal refundable credits: 96.1% of answers right for households owed $0Payroll tax60 households owed moneyPayroll tax: 88.5% of answers within 10% for households owed money (60 households, 50 models)88%Payroll tax: 98.6% of answers right for households owed $0Self-employment tax8 households owed moneySelf-employment tax: 94.4% of answers within 10% for households owed money (8 households, 50 models)94%Self-employment tax: 99.3% of answers right for households owed $0
policybench.org, October 14, 2026 · 50 models, no tools · income taxes before refundable credits · TANF (1 household) and SSI (2) omitted5

One household, as every model sees it

The facts
  • Arizona, 2026
  • Lives alone
  • Age 80, disabled
  • Social Security $23,736 a year
  • Private pension $6,720 a year
  • $58,700 in the bank
  • Plus health premiums, a mortgage balance and small investment income
SNAP for 2026
$240
PolicyEngine’s reference
$0
from all 50 AI models
$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0$0
Scenario explorer, policybench.org, October 14, 2026 snapshot · household 13 · $240 is SNAP’s $24 monthly minimum, March to December6

What the models miss

How the top two models explain $0
GPT-6 Sol · rank 1

“Monthly Social Security and pension income exceeds the amount that would produce a SNAP allotment for one person after the listed deductions.”

Claude Opus 5.5 · rank 2

“Countable income is about $2,545 per month. After the standard and small medical deductions, net income is about $2,330, which exceeds 100% of the poverty line (about $1,304).”

The rule they skip

Since March 2026, Arizona has extended SNAP’s categorical eligibility to every household with gross income under 200% of the poverty line. A categorically eligible household skips SNAP’s income and asset tests, and a household of one or two gets SNAP’s minimum benefit.

$240
PolicyEngine’s reference SNAP for this household in 2026: $24 a month, March to December
7 CFR 273.2(j)(2) · USDA FNS broad-based categorical eligibility chart, June 29, 2026 · PolicyBench, October 14, 20267

The same AI model, on its own and with the Axiom rules engine

The question

“I'm a single mom in Fresno, California with two kids (8 and 5). I make $3,400 a month before taxes, pay $1,500 rent, and pay about $120 a month for utilities, including heat. Can we get CalFresh, and how much would we get each month?”

On its own

The model answers from its training data and web search.

With the Axiom rules engine

The model calls the encoded rules and returns the answer with the provisions behind it.

Open the live demo axiom.org/gallery/chatbot
A demo · answers come from OpenAI’s GPT model, unedited · OpenAI does not sponsor or endorse this page8

The status quo in government

Government publishes the law as archaic websites and scanned documents, not code that any system can run.

Pennsylvania SNAP policy manual on a frames-era websiteLive today
A state SNAP policy manual
Redacted scan of a Medi-Cal eligibility procedures formLive today
Procedures manual — a scan of a scan
9

Rules encoding happens out of sight with limited documentation

Today’s rules engines hide their reading of the law
The law
Statutes, regulations and manuals
→
Someone's interpretation
Authored into the engine, out of sight
→
The rules engine
Executes that reading
→
The answer
No way to check it against the law
The current range of rules engines
A spreadsheet

Formulas with no provenance: one person's interpretation.

No provenance
A decision standard

A notation for expressing a decision, silent on where the rule came from.

Silent on the source
A business-rules engine

A generic engine that executes whatever rules someone authors into it.

Runs whatever is authored
An eligibility platform

Logic buried in a vendor system: the closed re-implementation problem.

Closed inside a vendor
A new rules language

Faster execution of the same unanchored rules. Speed doesn't fix provenance.

Faster, still unanchored
10

What if we had a shared foundation to power all digital tools?

One open, validated set of rules
  • Every value cites the provision it comes from
  • Every change carries its effective date
  • Every release is versioned, so you can see what changed
  • AI tools
  • Community screeners
  • Government systems
  • Staff training
  • Policy changes
  • Research

Every system that helps people discover and enroll in economic security programs runs on the same rules,
reducing cost, duplication and error at once.

11

The encoding pipeline

We take the law from its official sources, encode it rule by rule, and apply each rule only after its required checks pass. An AI judge flags issues for review. Every value keeps the citation it came from. One artifact then serves every surface, from a chatbot to a government rules engine.
Intake
Hundreds of official sources
Scraping pipelines
eCFRUSLMstate codes ×47policy manuals
↓
Database
Goal: keep every version
→
Corpus
The whole law, word for word
us:statutes/7/2017/a
§ 2017 · Value of allotment
“…30 per centum…”
→
Encoding loop
Drafted, checked, redrafted
Encoders
“30 per centum”
→ 0.30 rev 43
↓
Validators
✓ deterministic✓ comparison models, where availableAI judge, advisory
↺ Applied only when required checks pass
→
RuleSpec
One rule, fully cited
¶ rulespec-us
snap_allotment
imports:
us:statutes/7/2017/a
entity: Household
formula:
max(0, tfp − 0.30 × net)
0.30 ← “30 per centum”
→
Graph
The spec becomes a node
Typed · cited · executable
→
Compile and serve
One artifact, every surface
Program artifact
SNAP · US · 2026
✓ compiled
↓
ChatbotAI labs
Rules enginegovernment
API/MCPbuilders
12

What does an open, validated set of rules enable?

For governments
  • Determinations that cite the law they apply, so caseworkers and residents can see the reasons
  • Caseworker guidance and training built on the same rules
  • Payment accuracy checks before benefits go out
  • A policy change tested before it takes effect
  • Screeners and AI tools that give your agency’s answer
HOW THE RULES CROSS-REFERENCEeligibilityutilitieseligibilitycertifieseligibilityeligibilityTANFLIHEAPSSISNAPMedicaidSchool mealsWIC
For individuals

Accurate information to plan for their futures and receive assistance on any channel they use.

For AI systems

One source of record, so models stop estimating benefits amounts and start citing them.

For builders

Tools that guide people through benefits and taxes, built on rules they don’t have to re-encode.

For researchers

A computational model of the rules, for estimating what a policy does and reproducing the analysis.

13

Thank you

Your partnership is the path to solving this nationally.

Accurate AI answers

AI tools that give residents and staff the right answer for your agency.

Government systems

Your eligibility systems, screeners and caseworker tools on open, validated rules.

Ariel Kennan · President, the Axiom Foundation

Get in touch at hello@axiom.org

14
← → navigate · H home · C contents · P print to PDF