7 State and Local Government AI Pilots in 2026: What's Working, What's Failing

If you only follow the federal AI conversation — the executive orders, the NIST documents, the DoD budget lines — you are missing where the actual experiments are happening. State and local governments are running real AI in production right now, with real budgets and real constituents on the other end. Some of those pilots are working. Several are quietly failing. A few are useful warnings for everyone else.

I have been tracking these pilots through the National Association of State Chief Information Officers (NASCIO), the Government AI Coalition, and the public dashboards a handful of states have published. Here is the honest tour as of April 2026.

Why state AI pilots matter more than federal ones right now

The federal AI conversation in 2026 is dominated by big-dollar programs at DoD and the intelligence community — Anduril's $20 billion Lattice contract, the DIA Task Force Sabre work, classified efforts I cannot describe. That is significant, but it is not where most of the country lives. Most Americans interact with state and local government — the DMV, the unemployment office, child welfare, public schools, city permits — far more than they interact with any federal agency.

That means the AI tools that will shape ordinary Americans' first impression of "government AI" are state and local pilots. Whether those pilots are competent and trustworthy will set the political ceiling for everything that follows.

California: the most ambitious program in the country

California issued Executive Order N-12-23 in September 2023, then expanded its GenAI guidance in 2024 and again in 2025. By spring 2026, the state has multiple GenAI pilots running concurrently across several departments.

What is live: a Caltrans pilot using GenAI to summarize and triage traffic-incident reports. A DMV pilot drafting plain-language explanations of vehicle-registration rules for the call center. A Health and Human Services Agency pilot to translate constituent communications into a dozen languages, with human review.

What is working: the translation pilot. California is an unusually multilingual state, and the cost of professional human translation has been the binding constraint on equitable access to services. Adding GenAI translation with a bilingual reviewer in the loop has measurably reduced the average response time for non-English-speaking constituents.

What is rough: the Caltrans pilot has had inconsistent accuracy on incident severity classification. The state has not shut it down — they have iterated — but the public dashboard shows the metric still wobbling.

Texas: enterprise-wide TexasGPT

Texas took a different approach. Rather than department-by-department pilots, the Texas Department of Information Resources (DIR) built a state-wide GenAI offering — informally called TexasGPT — that any agency can subscribe to. The platform is a Microsoft Azure OpenAI deployment with state-specific fine-tuning, hosted in a sovereign Azure environment.

What is working: rapid deployment. Once the central platform was live, twenty-plus agencies stood up internal use cases in weeks rather than months. The Comptroller's office uses it for policy-question lookup. The Department of Public Safety uses it for incident-report drafting (with human review). The Health and Human Services Commission uses it for SNAP eligibility-letter drafting.

What is rough: governance. With twenty-plus agencies running their own use cases on the same platform, the state has had to retrofit governance after the fact — clarifying what data can flow through, who reviews outputs, what counts as a public record. The state is now publishing a uniform AI policy, but the gap between platform launch and policy backfill was real.

New York: cautious, audited, and constituent-first

New York has taken the most conservative posture of any large state. The state's Office of Information Technology Services issued a generative AI directive in 2024 that requires every pilot to (a) have a clear use-case statement, (b) document a measurable benefit, (c) include a human-in-the-loop unless explicitly exempted, and (d) be reviewed by the state CIO's office before launch.

What is live: a fairly small set of pilots — a Department of Labor chatbot for unemployment-insurance FAQ, a Department of State licensing-renewal explainer, a Department of Education policy-search tool for school administrators.

What is working: trust. New York has not had a public AI failure that made the front page. That is partly luck, but mostly procedure. The pilots are small, well-scoped, and reviewed before they touch a constituent.

What is rough: speed. The same procedures that prevent failures also slow innovation. New York is launching pilots that other states launched twelve months earlier. Whether that is a virtue or a vice depends on your priors.

Florida: focused on permitting and licensing

Florida's pilots are tightly scoped on a single problem: the permitting and licensing backlog that affects construction, professional services, and small business formation. The Department of Business and Professional Regulation (DBPR) and the Department of Environmental Protection (DEP) both run AI pilots that automate the routine portions of license review and permit checking.

What is working: throughput. Routine license renewals that used to sit in a queue for two to three weeks are now turning around in two to three days when no human review is needed. Renewals that need review are flagged earlier and routed to the right examiner.

What is rough: appeals. When the system flags an application as incomplete, the resulting appeals are slower than the underlying processing improvement. Florida is now building an "explain why this was flagged" feature into the pilot — exactly the kind of feature most states forgot to scope on day one.

Illinois: child welfare and the lesson of overreach

Illinois is the cautionary tale. The state's Department of Children and Family Services piloted a predictive-risk model for child-welfare cases — not generative AI, but classical machine learning — in the late 2010s, ran into accuracy and bias problems, and rolled it back. The state's GenAI pilots in 2025 and 2026 have stayed deliberately far from any consequential decision-making about people.

What is live: an internal staff knowledge tool, an interpreter-assist tool for caseworker visits, and a draft-letter tool for routine correspondence. None of them make a decision about a child or a family.

What is working: the discipline. The Illinois experience taught the state that AI in child welfare is not a place to move fast. The current generation of pilots is small, internal-facing, and explicitly excluded from case-decision authority.

The lesson: the right place for AI in child welfare is helping a human worker do their job better, not replacing the worker's judgment. Every state agency should re-read the Illinois case before deploying AI anywhere near a consequential decision about a person.

Virginia: the executive-order playbook

Virginia issued Executive Order 30 (2024) creating an AI task force and a state-level AI policy. The order is short, well-structured, and has been quietly copied (with variations) by several other states. Virginia's pilots include AI-assisted procurement document drafting, a Department of Motor Vehicles plain-language explainer, and a higher-education advising pilot.

What is working: the procurement pilot. Drafting RFPs and contract documents is repetitive work. Adding an AI drafting layer with attorney review in the loop has cut the time to first draft significantly without breaking the legal review process.

What is rough: the higher-education advising pilot has been criticized by some faculty as a partial replacement for human academic advisors. The state has clarified the pilot is supplemental, not a replacement, but the perception fight took six months to settle.

Washington: bottom-up agency experimentation

Washington state has not centralized AI procurement the way Texas did. Instead, the state CIO published guidelines and let individual agencies experiment within them. The result is a portfolio of small pilots — Department of Licensing, Department of Health, Department of Revenue, the courts — each picking a tool that fits its workflow.

What is working: agency ownership. Because each agency selected its own tool, agencies feel responsible for outcomes. There is no central scapegoat when a pilot underperforms.

What is rough: cost. Without a central procurement, the state is paying for redundant capability across agencies. A consolidation effort is rumored for late 2026.

The patterns: what works, what fails

After reading every state pilot announcement and every public outcome I could find, the patterns are clear.

What works. Pilots that pick a small, specific problem with a measurable metric. Pilots that keep a human in the loop on every consequential output. Pilots that publish their results — including the failures — so the next agency can learn from them. Pilots that start internal-facing (helping employees) before going constituent-facing (interacting with the public).

What fails. Pilots that announce a launch without saying what success looks like. Pilots that lock in a single vendor without a clean exit path. Pilots that touch a high-stakes decision (custody, eligibility, benefits) without a human reviewer. Pilots that run six months without publishing a single metric.

For small AI vendors and for federal contractors looking to expand into the state market, the practical takeaway is this: state work is real, and the budgets are growing, but the procurement path is slow. The fastest entry points are NASPO ValuePoint cooperative agreements, state-level master AI service contracts, and partnerships with primes who already hold a state contract.

Common questions

Are these pilots transparent enough? Mostly no. Only a handful of states publish pilot dashboards. The rest disclose pilots in press releases without follow-up data. Public-records requests can fill the gap, but the friction is real. The states with the best public dashboards (California's GenAI inventory, New York's directive log) deserve to be copied.

What about cities and counties, not just states? The action there is meaningful but harder to track. New York City, Los Angeles, San Francisco, Seattle, Austin, and Boston all have AI initiatives. The most concrete city-level success in 2026 has been multilingual constituent-services chatbots — a clean win when paired with a human review queue.

How is federal funding affecting state pilots? A growing share of state pilots are seeded with federal grant dollars (HHS, Department of Transportation, Department of Education). That brings the federal AI policy framework — NIST AI RMF, the Office of Management and Budget's M-24-10 memorandum — into state operations whether the state wants it or not.

Where can I read the underlying documents? California's GenAI Executive Order N-12-23, Virginia's EO 30 (2024), the Texas DIR AI guidance, and NASCIO's annual State CIO Top Ten Priorities are good starting points. NIST AI RMF 1.0 is the technical baseline most states cite.