Before You Deploy AI for Benefits Fraud Detection, Solve These Three Problems

The Financial Pressure Is Structural

Section 10105 of the One Big Beautiful Bill Act amends the Food and Nutrition Act of 2008 to establish tiered SNAP benefit cost-sharing obligations that scale with payment error rates: 0 percent below 6 percent, 5 percent for the 6–8 percent range, 10 percent for the 8–10 percent range, and 15 percent at or above 10 percent. The provision takes effect starting in federal fiscal year 2028, using either FY2025 or FY2026 error rates to determine initial liability, depending on which the state elects. States with rates at or above 13.34 percent receive a delayed phase-in to FY2029 or FY2030 under a separate high-error provision.

This creates a narrow window and a clear incentive. FY2025 payment accuracy data show that the vast majority of states exceeded the 6 percent threshold where cost-sharing begins, with a national average payment error rate of 10.62 percent. A state with a 9 percent error rate faces a 10 percent cost-sharing obligation, recalculated annually from its lagged error rate — a recurring exposure that grows with caseload size.

An important distinction before turning to what states are deploying: SNAP payment error rates measure the gap between what the program paid and what it should have paid. Those gaps are often administrative — income changes not timely reported, verification errors, eligibility miscalculations — rather than intentional fraud. The AI fraud detection discussion most often centers on Medicaid provider fraud: upcoding, phantom billing, and other intentional billing manipulation. These are different problem profiles, with different data signatures and different detection approaches. Procurement decisions shaped by one problem type tend to perform poorly on the other.

The Indiana CMS/Oracle pilot — a 90-day partnership announced in June 2026 — gave Indiana free access to Oracle AI software to flag upcoding and phantom billing, structured to also assess whether Oracle's models could be shared with other states. CMS has said it will publish the Indiana pilot's results; it is also working with five other states on Medicaid fraud detection pilots. Nextgov/FCW, October 2026

What States Are Deploying

The activity across states in 2026 reflects genuine urgency but uneven implementation depth.

Indiana piloted AI specifically for Medicaid claims review: pattern detection for upcoding and ghost billing in provider claims. Per WRTV reporting, FSSA Deputy Chief of Staff Marcus Barlow said the tool would not be able to initiate regulatory action, even while flagging potential abuse. That design constraint matters and is discussed below.

Arizona announced a Medicaid fraud detection tool in May 2026 for a July launch, using the Alivia 360 platform for AI-enabled prepayment review and risk ranking of provider claims. Performance data are not yet public.

Virginia issued Executive Directive 13 in August 2025, requiring local social services benefits staff to verify applicant expenses and household composition rather than accept self-attestation — an administrative step aimed at reducing the state's 11.5 percent SNAP error rate at the time of issuance before the FY2028 cost-sharing obligations take effect.

Maryland announced a partnership with Anthropic to deploy a Claude-powered benefits assistant to help residents apply for SNAP, Medicaid, temporary cash assistance, and WIC, and to help caseworkers verify documents and validate eligibility. Separately, Code for America and Anthropic launched a pilot of Code for America's SNAP Policy Navigator — a caseworker tool (not specific to Maryland) aimed at giving benefits workers real-time access to SNAP policy guidance and reducing administrative burden.

Minnesota passed a roughly $90 million package to modernize the software counties use to administer both SNAP and Medicaid benefits, with both programs funded simultaneously. The Minnesota approach reflects what several other states have not yet acknowledged: fraud detection AI running on fragmented legacy systems will not perform reliably, regardless of model quality.

Three Prerequisites That Determine Whether This Works

1. Data quality before model quality

Benefits fraud detection models reason over application records, income verification data, claims histories, and cross-agency identifiers. In most states, these data sources live in different systems, were collected under different schemas, and have no unified person-level identifier that persists across programs. A SNAP eligibility record may carry a Social Security Number, a Medicaid record may carry a state-issued beneficiary ID, and the two may reference the same person through a fuzzy match that resolves differently depending on which system runs the query first.

An AI model running over this environment does not detect fraud — it detects anomalies in data that are themselves partly artifacts of system fragmentation. The state's fraud investigation unit then receives leads that mix genuine fraud signals with data quality errors. Investigators who cannot distinguish the two lose confidence in the tool's output over time.

Minnesota's investment in modernizing intake and back-end integration is the kind of groundwork that fraud detection depends on. Data integration work — unified person IDs, normalized income verification fields, consistent case status codes — is unglamorous, slow, and not visible in a press release. It is also the work that determines whether the fraud detection investment pays off.

Before deploying a fraud detection model, a state should audit: what percentage of its SNAP or Medicaid enrollment records can be matched with high confidence to its identity proofing records? What percentage of income information is self-attested with no verification flag? How many cross-agency data matches are run through automated fuzzy matching versus a confirmed unique key? The answers to those questions predict AI reliability better than any model specification.

2. Explainability and caseworker workflow integration

Indiana's design decision — AI flags, humans must review before any action — is correct. It is also the minimum. A flag that tells an investigator "this claim scored 0.87 on our fraud model" is not actionable. Consider, as a hypothetical, a flag that reads: "this provider billed for 48 hours of services in a single day across 12 patients in three counties; similar billing patterns have preceded license revocations in other states." That gives an investigator a starting point.

The difference between those two outputs is not model accuracy; it is model design. Fraud detection in benefits programs requires that the system surface the specific anomaly and the evidence underlying it, not just a score. When flags are opaque, investigators either over-rely on the score (acting on insufficient evidence) or discount it (ignoring flags they cannot evaluate). Both failure modes produce waste and exposure.

Caseworker integration also means designing the appeals workflow into the tool from the start. When a flag leads to a proposed adverse action — reduced benefits, termination, referral — the recipient's notice must include the basis for the action in enough detail to contest it. That requirement comes from federal due process standards and from decades of administrative law, not from AI-specific legislation. It can be overlooked in procurement specifications that focus on what the AI detects rather than what happens after it detects something.

3. Due process architecture before deployment

Michigan's MiDAS unemployment fraud system, launched in October 2013, is the reference case for what happens when automated flagging and adverse action are coupled without adequate human review or due process. A review of its October 2013–September 2015 determinations found that 40,195 fraud findings had been robo-adjudicated — issued by the system without human involvement — and that the error rate on those findings was approximately 85 percent. The Unemployment Insurance Agency acknowledged that while MiDAS operated without human review, from October 1, 2013 to August 7, 2015, it falsely accused about 20,000 people of fraud — a 93 percent error rate on those robo-adjudicated determinations. Recipients received penalties with inadequate notice and an appeals process that was difficult to exercise in the timeframe allowed. Following a 2017 agency review of more than 62,000 cases, the state reversed about 70 percent of fraud findings and refunded approximately $20.8 million to affected claimants; a separate class action, Bauserman v. Unemployment Insurance Agency, reached a $20 million settlement in October 2022, with final court approval in January 2024. Daily Press, March 2026

MiDAS failed on data quality (the system conflated minor discrepancies with fraudulent misrepresentation), on explainability (recipients received fraud accusations with no explanation of what triggered them), and on due process architecture (penalties issued before recipients had meaningful opportunity to respond). These are systems design failures — not AI failures in the narrow technical sense — that any automated adjudication system would have replicated.

The same design failures appear in SNAP and Medicaid contexts when AI is deployed as a fraud enforcement tool rather than a fraud investigation tool. The distinction matters: investigation flags generate leads for human review; enforcement flags generate adverse actions. Deploying AI for enforcement without solving data quality and explainability first does not reduce fraud — it displaces fraud losses with legal costs.

What to Decide Before the Next Procurement

Before issuing an RFP for AI fraud detection in SNAP or Medicaid, agency executives and their IT leads should be able to answer four questions:

1. Can the AI's recommendations be explained to an investigator in operational terms? Not "the model scored this 0.9" — what specifically did it detect, and what does the investigation need to verify?
2. What is the data quality baseline for the records the AI will reason over? If a significant share of enrollment records has unverified income data or unresolved cross-agency identity conflicts, the AI's fraud signals will be indistinguishable from data quality noise. The specific threshold depends on program and system context; the point is that an audit is necessary before deployment, not after.
3. What is the adverse action workflow? Who reviews a flag before action is taken, what does the notice to the recipient say, and how are appeals integrated into the case management system?
4. What does the false positive rate need to be for this to be cost-effective? A fraud detection system with a high false positive rate imposes investigator review burden and appeals costs that may exceed the improper payments it prevents. Defining that threshold before procurement — not after — is what the requirements document is for.

These questions belong in the requirements document, not the vendor assessment stage. A vendor can describe a model's precision and recall on historical data; they cannot tell you whether your data is clean enough for those numbers to hold, or whether your agency's adverse action process complies with constitutional minimum requirements.

Sources and further reading

• SNAP cost-sharing provision (Section 10105), CALT Iowa State
• SNAP payment error rates and cost-sharing analysis, AEI/COSM
• Indiana FSSA to use AI to detect Medicaid fraud, Indiana Capital Chronicle, June 24, 2026
• Indiana looks to AI to root out Medicaid fraud, WRTV, 2026
• CMS hopes to quadruple provider suspensions over next six months, Nextgov, October 2026
• Governor Hobbs underscores Arizona's leadership in Medicaid integrity, azgovernor.gov, May 2026
• Arizona AHCCCS Alivia 360 prepayment review platform, AHCCCS Program Integrity
• Virginia Executive Directive 13, Digital Government Hub, August 2025
• Virginia ED 13 and 11.5% SNAP error rate, WAVY, August 2025
• Minnesota invests $90M in human services IT modernization, StateScoop
• The State of Maryland partners with Anthropic to better serve residents, Anthropic
• Code for America and Anthropic build AI tools to help SNAP caseworkers, StateScoop, May 2026
• Michigan's MiDAS Unemployment System, IEEE Spectrum
• Michigan MiDAS 93 percent error rate, GovTech/Detroit Free Press
• Michigan reverses 70% of unemployment fraud cases, Michigan Public Radio, August 2017
• Class action settlement approved, Michigan AG, January 2024
• Michigan's use of AI to process SNAP applications draws concerns, Daily Press, March 2026

Spartan X works with state agencies on AI program architecture — specifically the institutional layer beneath the technology: data environments, operational workflows, and governance structures that determine whether AI tools produce reliable outcomes in production. Benefits fraud detection is a domain where deployment speed and program accountability are genuinely in tension, and resolving that tension at the design stage is the difference between a program that reduces improper payments and one that creates a new category of legal and operational risk.