AI Case Studies / Case study
How State and Local Governments Run on AI
State agencies, cities, and counties are already putting AI to work on benefits casework, 311 requests, permits, and translation. Here is what real deployments look like and what they actually deliver.
State and local governments use AI mainly to speed up back office work and widen public access to services. California negotiated a 50 percent discount on Anthropic's Claude for its agencies, Pennsylvania rolled ChatGPT Enterprise out to 3,000 employees and reported about eight hours saved per person each week, Code for America and Anthropic built a SNAP Policy Navigator for benefits caseworkers, and cities like Denver and San Jose run resident chatbots and staff built AI tools. The common thread is human review kept in the loop, not full automation.
Key Takeaways
- The real buyer is not one office. State CIOs, county sheriffs, city IT teams, and benefits agencies each adopt AI for their own workflow, so a consultant sells to many doors, not one.
- The measurable wins are internal. Pennsylvania reported roughly eight hours saved per employee per week from ChatGPT Enterprise across 35 agencies before it expanded access to 3,000 workers.
- Translation is the fastest public facing win. Minnesota's Enterprise Translations Office and Santa Clara County's Sheriff both adopted AI translation to reach residents who do not speak English.
- Caseworker copilots are the frontier. The Code for America and Anthropic SNAP Policy Navigator grounds Claude in verified federal, state, and county policy so a worker gets a real answer, not a guess.
- Permitting is where AI meets money and speed. California made the Archistar plan review tool free for Los Angeles fire rebuilds to turn a multi week check into hours.
- Failures are public and instructive. New York City kept its MyCity business chatbot live after it gave wrong legal advice, a reminder that governance and disclaimers are part of every real deployment.
- Procurement discounts are opening the door. Anthropic's 50 percent California deal and its $15 million cyber defense credits for state, local, tribal, and territorial governments lower the cost of the first pilot.
Government is buying AI the way businesses did two years ago
Public agencies moved slower than private firms, and for good reason. A wrong answer from a benefits system or a permit tool has legal weight. But the caution is over. By late 2024, employees in 82 percent of state CIO organizations were using generative AI tools daily, up from 53 percent a year earlier. What changed is not the hype. It is that a handful of real deployments now have names, dates, and numbers attached, which means the work has moved from pilots to procurement.
That matters for anyone advising the public sector. The buyer is fragmented. A state IT office, a county sheriff, a city permit counter, and a benefits agency all adopt AI on separate budgets and separate timelines, and each one needs help scoping the work, checking the vendor, and writing the guardrails. That is consulting work, and it is the reason we built [the Leveraged Consultant course](/leveraged-consultant) around exactly this kind of engagement. Below are twelve real deployments, grouped by the job they do.
Statewide access deals: California and the Anthropic model
The clearest signal came on June 29, 2026, when Governor Gavin Newsom and Anthropic announced a deal giving California state agencies and local governments access to Claude at half price, plus training and technical support (TechCrunch, June 2026). Newsom framed it plainly: AI should help government workers move faster, not replace them.
Anthropic paired that with a separate program announced June 12, 2026, committing up to $15 million in Claude credits for state, local, tribal, and territorial governments, with as much as $100,000 per smaller entity for cyber defense work, runbooks, and cohort training. California and Texas joined the first cohort (StateScoop, June 2026). For a consultant, these two moves lower the single biggest barrier to a first project, which is the cost of the pilot.
Workforce rollouts: Pennsylvania, Colorado, and San Jose
Pennsylvania ran the most cited experiment. A year long pilot with ChatGPT Enterprise, starting at 175 employees across 14 agencies, reported that participants saved an average of about eight hours per week. On April 15, 2026, the Shapiro administration expanded access to 3,000 employees across 35 agencies, with another 6,500 enrolled in training. Real use cases included drafting communications, summarizing long documents, and processing records with poor image quality (StateScoop, April 2026).
Colorado's Office of Information Technology took a smaller, measured path, running a structured 90 day Google Gemini pilot in summer 2024 with roughly 150 participants drawn from multiple agencies, then folding the lessons into a statewide AI framework (Colorado OIT). San Jose went a different direction and taught staff to build their own tools. Working with San Jose State University, the city trained more than 1,000 employees, about 15 percent of its workforce, through a voluntary upskilling program. Staff shipped a tool that verifies emergency vehicles are equipped before deployment, one that reviews contractor submissions for missing information, and an assistant that scores projects against the city's carbon neutrality goal (StateScoop, 2026).
Benefits casework: the Code for America and Anthropic SNAP tool
The most consequential deployment is the least flashy. On May 8, 2026, Code for America and Anthropic announced the SNAP Policy Navigator, a Claude powered tool that gives food assistance caseworkers real time access to federal, state, and county policy. It is built on the Model Context Protocol so answers stay grounded in verified guidance rather than general model output, and the roadmap covers reviewing eligibility documents and drafting plain language notices for recipients (Code for America). This is the pattern to watch. A benefits worker does not need a chatbot that sounds confident. They need one that cites the right rule, and that is a design and oversight problem a consultant can own.
Translation and 311: Minnesota, Santa Clara County, and Denver
Translation is where AI reaches the public most directly. Minnesota's Enterprise Translations Office adopted ChatGPT to widen access to state services for communities that do not speak English, working translation into an office whose whole job is precise, timely language support (OpenAI). In the Bay Area, the Santa Clara County Sheriff's Office launched a one year pilot of the Pocketalk translation device so deputies could communicate in the field, timed ahead of major events including the FIFA World Cup (Santa Clara County Sheriff).
On the 311 side, Denver's chatbot Sunny, built by Citibot, handles resident questions and service requests in more than 45 languages across web, SMS, and WhatsApp. The vendor reports Sunny has engaged more than 95,000 residents, answered over 100,000 questions, and filed more than 5,000 service requests at a fraction of the cost of a staffed call (Citibot). New Jersey took the internal route with the NJ AI Assistant, a secure generative AI platform for state employees run through the New Jersey Innovation Authority (NJ Innovation Authority).
Permitting: California's Archistar rollout
Permitting is slow, expensive, and universally hated, which makes it a prime target. On April 30, 2025, California launched the Archistar plan review tool to speed rebuilding after the Los Angeles fires. It uses computer vision and automated rulesets to check designs against local zoning and building codes, letting owners pre check plans before they file. The state offered it free for the Eaton and Palisades recovery and made it available statewide by contract (Governor of California). Similar tools are live in more than 25 cities across North America and Australia. The honest caveat: the launch promised turning weeks into hours, but published performance numbers are still thin, so a consultant should treat the speed claim as a hypothesis to measure, not a fact to repeat.
The cautionary case: New York City's MyCity chatbot
Every real portfolio needs the deployment that went wrong. New York City launched the MyCity business chatbot in October 2023 to answer questions from more than 2,000 city business pages. In early 2024, testing found it giving answers that, if followed, would have owners break the law, including telling employers they could take workers' tips. The city did not pull it. It added disclaimers, kept iterating, and a later Comptroller audit picked apart the development process (OECD AI Incidents). The lesson is not that government AI fails. It is that governance, testing, and clear disclaimers are line items in every serious project, which is a point we make in detail in [how federal agencies run on AI](/how-federal-agencies-run-on-ai) and across the rest of our [AI case studies](/ai-case-studies/).
What this means for the work
Read these twelve together and a consulting map appears. The entry points are cheap now because of vendor discounts. The reliable wins are internal time savings and translation. The high value, high risk work is benefits casework and permitting, where accuracy and legal exposure demand real oversight design. And the failures are public, which means the market rewards advisers who can build the guardrails as confidently as they build the tool.
| Organization | What they use AI for | Reported result or status | Source |
|---|---|---|---|
| California (statewide) | Claude access for agencies and localities | 50 percent discount plus training, June 2026 | TechCrunch 2026 |
| Anthropic SLTT program | Cyber defense credits for governments | $15M in credits, CA and TX joined, June 2026 | StateScoop 2026 |
| Pennsylvania | ChatGPT Enterprise for staff work | About 8 hours saved per week, expanded to 3,000 staff | StateScoop 2026 |
| Colorado OIT | Google Gemini productivity pilot | 90 day pilot, about 150 participants, summer 2024 | Colorado OIT |
| San Jose | Staff built AI tools | 1,000 plus employees trained, multiple live tools | StateScoop 2026 |
| Code for America and Anthropic | SNAP caseworker policy navigator | Claude grounded in verified policy, May 2026 | Code for America 2026 |
| Minnesota ETO | Translation of state services | Wider access for non English speakers | OpenAI |
| Santa Clara County Sheriff | Field translation for deputies | One year Pocketalk pilot | Santa Clara County Sheriff |
| Denver | Sunny 311 resident chatbot | 95,000 plus residents, 45 plus languages | Citibot |
| New Jersey | Secure staff AI assistant | Statewide platform for employees | NJ Innovation Authority |
| California and Los Angeles | Archistar building permit review | Free for fire rebuild, statewide by contract, 2025 | Governor of California |
| New York City | MyCity business chatbot | Gave wrong advice, kept live with disclaimers | OECD AI Incidents |
Frequently Asked Questions
Is any of this producing real savings, or is it all press releases?
Pennsylvania is the strongest evidence, reporting about eight hours saved per employee per week before expanding ChatGPT Enterprise to 3,000 workers across 35 agencies. Denver's vendor reports tens of thousands of resident interactions at a fraction of call center cost. Permitting speed claims are still mostly unmeasured, so treat those as promising rather than proven.
Who actually buys AI inside a government?
There is no single buyer. State CIO offices run workforce pilots, benefits agencies build caseworker tools, sheriffs and cities buy translation and 311 bots, and permit departments buy plan review software. Each has its own budget and approval path, which is why public sector AI work spreads across many small engagements.
Are governments building this themselves or hiring outside help?
Both. San Jose trained its own staff to build tools, while most agencies partner with vendors like Anthropic, OpenAI, Citibot, or Archistar and lean on outside advisers to scope pilots, vet vendors, and write governance. The gap between a signed contract and a safe, working system is where consultants earn their fee.
What is the biggest risk in a government AI project?
Wrong answers with legal consequences. New York City's MyCity chatbot told business owners things that would have broken the law. The fix was not abandoning the tool but adding disclaimers, testing, and human review. Any serious proposal has to budget for oversight from the start.
How are agencies affording this?
Vendor discounts are doing heavy lifting. California secured Claude at half price, and Anthropic put up $15 million in credits for state, local, tribal, and territorial governments. Those deals make the first pilot cheap, which is often the hardest budget line to get approved.
Is state and local different from federal AI adoption?
Yes. Federal agencies operate at national scale with different procurement and security rules. State, county, and city buyers are closer to residents, run smaller budgets, and adopt tools faster and more independently, which changes how you scope and price the work.
Sponsored Training
Informational analysis for working professionals, not professional advice.