Yes, within limits. An AI caseworker policy assistant can help mitigate errors that start with a misread or missed rule, provided it answers from the agency's current policy with the passage cited and the caseworker still makes the call. It cannot fix a wrong pay stub or a job the household never reported. A general-purpose chatbot with no access to the agency's manual will add errors of its own.
The question got expensive this year. USDA's fiscal year 2025 payment error rates, published in June, put the national SNAP rate at 10.62%. That figure counts overpayments (9.28 points) and underpayments (1.33 points) above a $57 tolerance, and only ten of the 53 state agencies came in under 6%.
The 6% line now carries a price. Under H.R. 1, the budget law passed in July 2025, states at or above 6% will pay between 5% and 15% of SNAP benefit costs from fiscal year 2028, a bill the federal government has always covered in full. State agencies are looking hard for ways to cut errors, and AI tools sit near the top of the list.
Which errors an AI agent assist tool can catch
AI agent assist is software that works beside a person handling a case and surfaces what they need in the moment. The term comes from call centres. A caseworker policy assistant applies the same pattern to eligibility work: the caseworker asks a policy question and gets back the relevant rule.
The questions are rarely simple. One test question from Nava's work with Benefits Data Trust asked whether an applicant's 20-year-old child, a college student who lives at home only in summer, belongs on the SNAP application. Answering it means reading household composition rules alongside student rules, then checking for a state twist.
In most offices that job falls to the colleague two desks over who has worked there since 2009. They know everything, and they are also the person every other caseworker is already asking.
SNAP quality control sorts errors by cause, and a policy assistant reaches only some of them:
| Error source | Example | Can a policy assistant help? |
|---|---|---|
| Policy misapplied by the worker | A household member counted who should be excluded | Yes, this is the core use case |
| Rules explained unclearly | A household never told it must report a pay rise | Partly, through plain-language explanations |
| Data keyed in wrongly | Monthly income typed as weekly | No |
| Household information wrong or missing | An unreported second job | No, that needs verification data |
| Intentional fraud | Income deliberately concealed | No |
So the realistic ceiling is the share of your errors that begin with a rule. Quality control reviewers already record what went wrong in each sampled case, so pull that number before you talk to a vendor. Policy lookup is a bounded and checkable job, the kind our guide to which tasks suit generative AI recommends starting with.
What the AI expansion in government in 2026 looks like for benefits
States are past asking whether to try AI in benefits work. Whether the pilots reduce errors is the open question.
Code for America's second annual Government AI Landscape Assessment, released on May 1, found more states moving from piloting into experimentation, while putting AI into daily operations and measuring its impact remain in the early stages. Pennsylvania is scaling a tool that scans application documents for legibility, and Maryland partnered with Anthropic on an AI agent that helps residents apply.
A week later, Code for America and Anthropic announced the SNAP Policy Navigator, a Claude-powered tool that gives caseworkers answers to case-specific policy questions. It is built on the Model Context Protocol, an open standard for connecting AI applications to trusted data sources, so answers draw on verified policy from the federal level down to counties. Nava's Benefit Navigator chatbot takes a similar line, showing its sources on every answer.
None of the announcements above includes a before and after error rate, so for now the evidence comes from benchmarks and early pilots.
How accurate is an AI policy assistant on SNAP rules?
Out of the box, general-purpose models fall short. With the agency's own policy behind them, the evidence is better, though thin.
The clearest public measurement is Vals AI's Public Benefits Bench, funded through the Center for Civic Futures. It poses 459 SNAP scenarios, each graded against a rubric that SNAP policy experts validated. As of October 7, 2026, the top model scored 76.93% and the weakest 41.41%. The authors' verdict is blunt: no general-purpose model performs well enough to be trusted with SNAP guidance.
A 77% score earns a respectable grade in most classrooms. Applied to benefit amounts, it earns a corrective action plan.
Where the models failed matters more. They struggled most with answers that depend on local office procedures and with recent H.R. 1 changes such as immigrant eligibility. Accuracy also fell when a question combined rules, such as household composition and income deductions. Those are the questions that send a caseworker to the manual.
Two findings point at the fix. Letting a model ask follow-up questions raised scores by about 24.5 points on average, while web search added about 4.5. The authors conclude that the knowledge models lack sits inside state agencies, in internal manuals and local procedures, and that reliable answers need better grounding in those documents. Grounding means retrieving the agency's own text and answering from it. Public web search cannot reach internal manuals, and our note on why LLM web search answers can be inaccurate covers the other ways it slips.
Nava's experience lines up. In 2024, its first chatbot generated answers from Benefits Data Trust's policy database, and its accuracy was highly variable. The team switched to having the model summarise the retrieved policy, which let navigators check each answer against its source. Later, navigators told Nava they feared outdated rules and wanted a "last updated" date on every response.
AI governance failures that turn one error into thousands
A caseworker who misreads a rule makes one mistake. A system that misreads it makes the same mistake on every case it touches, at machine speed.
Michigan learned this with unemployment insurance. Between October 2013 and August 2015, an automated claims system flagged more than 50,000 potential fraud cases, and an initial state review found a 93% error rate. The 2017 settlement of a federal lawsuit required that claimants be notified of fraud accusations and given time to respond. That system predates generative AI, and the lesson transfers intact: automated decisions that no person reviews scale their mistakes.
A policy assistant fails more quietly. The first failure is automation bias, the habit of accepting a machine's answer because it arrives fast and sounds sure. A caseworker with a full queue will often skip the citation. The second is a stale index. When a policy memo supersedes a manual section and the assistant keeps retrieving the old text, every office using it applies the old rule until someone notices.
Both are governance problems, fixed by decisions about who checks what, made before launch. Agentic AI in Business covers how to set those autonomy boundaries for AI systems inside an organisation.
AI governance and compliance for a caseworker policy assistant
The controls that make an assistant safe also make it useful. None of them needs a research breakthrough, which is good news for anyone whose procurement cycle is measured in fiscal years.
- Answer only from the agency's own policy corpus, from federal regulations down to county memos, and cite the section and effective date on every answer.
- Show when each source was last updated, and re-index the corpus the day a memo is issued.
- Tell the caseworker when the retrieved policy does not cover the question, and route it to a policy specialist.
- Ask about the household before answering. Follow-up questions were the biggest single gain in the Vals results.
- Keep the determination with the caseworker. The assistant explains the rule, and the person applies it to the case.
- Log every question with its answer and cited sources, so quality control can trace an error back to the answer behind it.
- Build a test set from last year's quality control findings and rerun it whenever policy changes.
Most of these controls are design decisions. Citations and the refusal path live in how the assistant's tools and system prompt are written, the ground Agentic AI Engineering covers for teams building their own. If you are still choosing a use case, a problem-first approach to building AI applications starts from the error data you already have.
A plan for next quarter: count last year's quality control findings that trace back to a misapplied rule, and turn twenty of them into test questions. Run any candidate assistant against them before a caseworker sees it. If it answers with the right citations, pilot it in one office and compare error findings with a similar office.
If it passes, the colleague two desks over might finally finish a coffee while it is still warm.