Loading...
Loading...
No state in the country has a more cautionary AI story than Michigan's, and every serious conversation about government AI here starts with MiDAS. The Michigan Integrated Data Automated System — deployed by the Unemployment Insurance Agency between 2013 and 2016 — used an automated fraud-detection algorithm that generated over 40,000 false fraud accusations against unemployed workers, incorrectly flagging legitimate claimants, triggering penalties, and in some cases pushing families into financial crisis. A 2016 state audit confirmed the system operated for months with a known false-positive rate above 90%. The legal settlement cost Michigan taxpayers more than $20 million, and the political fallout shaped how DTMB (Department of Technology, Management and Budget) and legislative oversight committees approach every subsequent AI procurement. Governor Whitmer's 2024 Executive Directive on Responsible AI Use — issued after a statewide AI in Government Task Force convened in 2023 — establishes a framework that requires algorithmic impact assessments, public comment periods for high-risk AI deployments, and annual third-party audits of automated decision systems that affect benefits or civil rights. The directive is one of the more operationally specific state AI governance documents in the country, reflecting hard lessons from MiDAS that most states haven't had to learn firsthand. AI vendors entering Michigan government must understand both the technical standards DTMB enforces and the political sensitivity that now surrounds any AI system touching benefits, licensing, or enforcement.
The MiDAS settlement did more than generate a liability — it established an evidentiary record that Michigan's Legislature, civil-rights attorneys, and GAO auditors cite when evaluating new AI proposals. Any AI vendor pitching Michigan state agencies on fraud detection, eligibility determination, or benefits processing will encounter questions that reference MiDAS explicitly: What is the false-positive rate at the 99th confidence interval? How does the system handle appeals? What human review step exists before a determination is communicated to a claimant? DTMB's post-MiDAS procurement standards require that any AI system in a high-risk category — defined as systems that can restrict, deny, or delay a citizen's access to government services — must pass a Human-in-the-Loop (HITL) sufficiency review before go-live approval. The University of Michigan's School of Information, working with the Ford School of Public Policy, has published a widely-cited algorithmic accountability framework that DTMB's IT governance team uses as a reference document alongside the NIST AI RMF. In practice, this means Michigan agencies spend more time on pre-deployment testing than most comparable states — 4-6 months for medium-risk systems — but go-live incidents are rare, and the post-MiDAS liability exposure calculus strongly incentivizes that investment.
Below the state level, Michigan's AI landscape is dominated by the scale and fiscal complexity of Detroit and its inner-ring suburbs. The City of Detroit's Department of Innovation and Technology — relaunched and expanded under Mayor Mike Duggan and continued under Mayor Sheffield — has deployed AI in 311-call routing, blight prediction modeling, and tax-foreclosure risk assessment. The blight-prediction model, built in partnership with Data Driven Detroit and with University of Michigan researchers, has been cited in academic literature as a rare example of a municipal AI model that was tested for racial disparate impact before deployment and adjusted accordingly. Grand Rapids has piloted an AI-assisted permit-review system through its Planning Department, targeting the commercial building-permit backlog that had reached 90+ days in 2023. DTMB's Shared Technology Services platform, which hosts applications for roughly 70 state agencies on a common Azure Government tenant, is the infrastructure backbone — AI projects that integrate with shared DTMB services move faster than agency-standalone builds because authentication, logging, and data-classification layers already exist. Operators report that the DTMB platform onboarding process, while bureaucratic, eliminates a class of security and compliance work that can consume 30-40% of a typical state AI project budget in states without centralized shared infrastructure.
The Unemployment Insurance Agency has been rebuilding its fraud-prevention capability under legislative scrutiny since 2016, and the COVID-19 pandemic — which drove Michigan UI claims from roughly 5,000 per week to 1.5 million in a two-month span in spring 2020 — stress-tested every improvement made since MiDAS. Michigan lost an estimated $8.5 billion to fraudulent UI claims in 2020-2021, according to a 2022 OIG estimate, despite the post-MiDAS reforms. The current fraud-detection system, rebuilt with federal CARES Act administrative funding, uses a tiered approach: velocity-pattern detection flags suspicious claim clusters; identity-verification checks (Knowledge-Based Authentication, ID.me integration) create a second gate; and human adjudicators review flagged claims before any fraud determination is issued. The system is explicitly designed around the MiDAS lesson — no automated determination without a documented human-review step. On the citizen-services side, the Michigan Department of Health and Human Services (MDHHS) is the state's largest service-delivery agency, managing Medicaid, SNAP, cash assistance, and child welfare programs for more than 3 million Michiganders. MDHHS has deployed NLP-assisted application processing on its MI Bridges portal, reducing average document-review time on SNAP applications from 11 days to 6 days. The NLP model handles benefits-verification document extraction; eligibility determination remains with human caseworkers — a design choice explicitly informed by MiDAS.
Strategic planning for AI adoption, readiness assessment, and roadmap development
Workflow automation using AI, including Make.com-style automation and RPA
Predictive models, data analysis, and ML pipeline development
Text analysis, document automation, sentiment analysis, and language processing
Every Michigan agency procurement team with institutional memory — and DTMB's IT governance staff specifically — will ask directly about false-positive rates, appeals processes, and human-in-the-loop design for any AI system touching benefits, licensing, or enforcement. The 2024 Executive Directive codified this into formal requirements: algorithmic impact assessments, public comment for high-risk systems, and annual third-party audits. Vendors who treat this as compliance theater get screened out early; vendors who can document their testing methodology against MiDAS-type failure modes — particularly on demographically correlated false-positive rates — move to shortlist.
The directive requires agencies to classify AI systems by risk tier, conduct algorithmic impact assessments for medium- and high-risk systems, post public notices before deploying high-risk AI in citizen-facing contexts, and conduct annual third-party audits of live systems. It also requires agencies to report AI system performance data to DTMB on a semi-annual basis. The risk-tier classification mirrors NIST AI RMF categories, with the Michigan-specific addition that any system capable of affecting a benefit determination or civil-rights-related decision is automatically classified as high-risk, regardless of the vendor's own risk self-assessment.
Agencies building on DTMB's Azure Government shared tenant — which covers authentication, data classification, logging, and network security — typically save 3-4 months versus agency-standalone builds. The tradeoff is DTMB's Architecture Review Board process, which adds 6-10 weeks for new AI system approval. Net-net, DTMB-native projects tend to reach go-live 1-2 months faster than standalone builds once the architecture review is completed, and they carry significantly lower long-term maintenance burden. Municipal agencies in Detroit and Grand Rapids that connect to DTMB shared services through Michigan's Local Government Technology framework get similar infrastructure benefits.
Detroit's blight-prediction model — built with Data Driven Detroit and University of Michigan researchers — remains active and has been updated with 2022 and 2024 parcel data. The key lesson other Michigan municipalities cite is the pre-deployment disparate-impact testing: the original model showed statistically significant racial correlation in predictions that tracked historical redlining geography. The team adjusted feature weighting and added a human-review step for parcels in historically redlined areas before deploying. Grand Rapids, Flint, and Lansing have each cited the Detroit model as a design reference for their own blight and infrastructure-maintenance pilots.
311 and constituent-contact routing via NLP, permit-review automation (particularly building and zoning), tax-assessment anomaly detection, and predictive road-maintenance scheduling are the four categories with the most active procurement activity. Detroit's DIT, Grand Rapids Planning, and the City of Ann Arbor IT division have all issued RFIs or RFPs in at least one of these categories since 2024. Vendors with Michigan references — particularly post-MiDAS-era deployments with documented human-review architectures — are at a significant advantage over vendors with only private-sector AI experience.