Most conversations about AI in government stay abstract. The agencies below have moved past pilots: they have tools in production, measured results, and lessons other jurisdictions can borrow.
The nine use cases come from cities, counties, a regional water district, and several state agencies whose work sits close to local government. Each one names the problem, what the agency deployed, and what changed. Results are as reported in the technology vendors' published customer case studies, with the vendor named in each section.
We found these stories the same way we find most of what goes into Clearbox Source: by following the evidence of what governments actually buy and deploy. Every agency and product below links to its page on Marketplace.city, so you can see the agency's broader technology footprint and where else the same tools are in use.
Harris County, Texas saved roughly $2.5 million by using AI to read its own purchasing records. The county's Universal Services team built PRISM (Purchasing Review and Intelligence for Software Management) on Microsoft Foundry in three weeks.
Procurement systems are full of free-text line items that resist standard reporting. PRISM analyzed more than one million of those entries to find overlapping software licenses and consolidation opportunities. According to Microsoft's case study, the savings came from license optimization alone.
What makes the approach repeatable is the governance around it. PRISM went through the county's responsible AI, investment, cybersecurity, and architecture committees. Oversight is proportional: high-risk uses need strong evidence and a narrow scope, while low-risk tools such as public chatbots move faster.
The county also sorts AI into three "swim lanes": personal productivity (Microsoft 365 Copilot), production chatbots (Copilot Studio), and custom enterprise agents (Foundry). Each lane is measured differently, with qualitative surveys for productivity tools and hard numbers for enterprise agents.
The City of Raleigh, North Carolina cut IT service desk costs by 66% with AI agents, according to ServiceNow. Four help desk staff support 4,400 employees and more than 80,000 devices, while the city adds about 75 residents a day.
The city's agent, Ral-E, answers employee questions and routes requests across IT, HR, and facilities using the city's existing knowledge base. ServiceNow reports it routes 98% of tickets correctly on first contact and returns more than 1,300 staff hours a year. AI-generated incident summaries saved 137 hours of writing and reading time in the first three months, and up to 98% are accepted without edits.
Raleigh did not start with AI. It spent roughly a decade standardizing change management, ITSM, HR, and facilities workflows on one platform, which gave the agents reliable knowledge to draw on. A second agent, Alli, now completes more than 20 workflows end to end, and the city plans to bring both agents to its resident-facing Ask Raleigh portal.
One State of Maryland agency had staff build an app with Google Gemini in five weeks at no added cost, against a vendor proposal of $400,000 over eight months. Maryland's Senior AI/ML Program Manager, Lauren Maffeo, described the result in a Google customer video.
The smaller wins add up too. A summer intern used Gemini to build a no-code chatbot that manages the public response-to-comments process, which had been fully manual. Maffeo's team transcribes user interviews with Gemini and analyzes them in NotebookLM, saving hours of manual transcription.
Adoption followed the state's first AI roadmap, published in January 2025, and a vetting process that confirmed data stays in the state's environment. Within six months, one in three eligible state employees was using Gemini regularly. A voluntary monthly AI community of practice now draws about 150 people.
For procurement teams, the takeaway is not that agencies will stop buying software. It is that internal build capacity is becoming a real benchmark against which vendor proposals get judged.
Covered California, the state's health insurance marketplace, reports that Glean users save four to six hours per week each and more than 100 hours collectively per RFP review. The agency handles large volumes of personal data, so its first requirement was that any AI tool mirror its existing role-based security model rather than duplicate it.
Staff now build their own agents for specific jobs:
The rollout started as a 50-person pilot with policy and documentation teams. Users reported time savings by week four, and the agency expanded to 1,500 employees across 14 divisions, gating each wave on readiness and change-management review.
Attackers now use generative AI to write convincing phishing and business email compromise messages, and agencies are answering with AI-based filtering.
Sutter County, California has had zero account compromises since deploying Abnormal AI, according to the vendor. A 16-person IT team supports about 1,100 users, and an account takeover was what pushed the county to act. Abnormal reports the system stops more than 200 advanced attacks per month and saves 56 staff hours a month by automatically triaging employee-reported phishing, a job one person used to do by hand. General Services Director Ken Sra summed up the reasoning: the county has to "fight AI attacks with AI."
The California DMV, with more than 10,000 employees, blocked 1.8 million unwanted and malicious messages in 30 days using Proofpoint, including more than 380,000 advanced threats. Its security team also uses data loss prevention to watch for a newer risk: employees uploading sensitive documents to consumer AI tools outside policy.
North Dakota Information Technology (NDIT) cut open security alerts from 16,000 to about 50, a 99.6% drop, according to Palo Alto Networks. NDIT secures more than 250,000 endpoints across 600+ entities, including cities, counties, K-12, and higher education, with a fixed number of staff.
Automation now resolves 60% of incidents without an analyst, which Palo Alto Networks equates to the work of 8 to 10 SOC analysts. Managed threat hunting shortened the time to find a true positive from weeks to minutes. The state also reports security staff tenure has doubled to about three years, which it credits to less manual alert work.
The model extends past state lines. NDIT helped stand up a Joint Cybersecurity Operations Command Center where nearly 20% of U.S. states share threat data and compare responses.
Dallas County Health and Human Services eliminated 2,000 hours of manual record review a year by automating how it matches and merges public health data, according to Informatica. At the start of COVID-19, lab results still arrived by fax, and staff spent up to 15 hours a day removing duplicates and verifying addresses.
The department built a central disease surveillance system on Informatica Master Data Management, connected to Salesforce, the national disease surveillance system, and the Texas immunization registry. It processes more than 4.5 million data points and creates a single trusted record per person. Informatica reports 20% of contact-tracing data was deduplicated and the reliability of critical data elements improved by 50%.
This is the least flashy use case on the list and arguably the most important. The department is now exploring AI features such as natural language queries, so epidemiologists can analyze data without writing SQL. Those features depend on the clean data it built first.
The South Florida Water Management District (SFWMD) cut wetland analysis times from hours to under a minute, according to Google Public Sector. The regional agency manages water resources for more than nine million residents across 16 counties.
Its scientists track hydroperiods, the timing, duration, and depth of flooding, across 40 square miles of Kissimmee River floodplain to guide Everglades restoration. The new Hydroperiod Calculator combines Google Earth Engine with Google Cloud AI and data services to process overnight water depth data and run comparisons interactively.
The district went from concept to full deployment in 10 months. Google engineers also used Gemini to help district scientists refine large volumes of existing code. SFWMD plans to extend the tool to other South Florida water resources.
The City of Parkland, Florida reduced network-related trouble tickets by 95% after replacing its network with an AI-managed one, according to HPE. Parkland has about 38,000 residents and expects roughly 15,000 more, and its aging network went down in thunderstorms and drew near-daily complaints.
Working with partner MORSECOM, the city installed a wired, wireless, and SD-WAN network run by HPE Mist AI, cutting over in a single weekend. The system monitors itself and recommends fixes. Its AI assistant, for example, flagged bad cabling during a security camera installation so the team could fix it immediately.
For a small IT team, the benefit is fewer problems to chase. Residents also gained free, reliable Wi-Fi at city parks and recreation facilities.
Several agencies are not yet using AI at scale but are clearing the way for it by retiring fragmented legacy systems.
The City of Akron, Ohio replaced decades-old HR, finance, and payroll systems with Workday, going live in 45 weeks and avoiding more than $1 million in recurring HR upgrade costs, according to Workday. Six hundred employees who had never used a PC for work gained self-service and mobile access. The city is now exploring Workday's AI agents, and its CTO, Darren Rosenek, describes the training model as simply typing a question.
The State of Idaho consolidated more than 75 legacy applications into a single Infor ERP on AWS GovCloud, integrating data from 86 agencies for more than 20,000 employees, according to AWS. The state says it can now answer legislators' questions about contracts and inter-agency costs from one source.
Neither case is a finished AI story. Both show the step that Raleigh and Dallas County point to as well: AI performs only as well as the data and workflows underneath it.
Across these agencies, the results came from narrow, well-defined problems rather than broad AI mandates. Four patterns repeat:
For vendors selling to government, the bar is rising. Agencies now compare proposals against what their own staff can build, and they expect governance, security, and measurable outcomes from the start.
Every implementation on this list surfaced in the research behind Clearbox Source. Our team tracks how governments actually buy and use technology, from purchasing records and contract activity to the case studies vendors publish about their public-sector customers, and ties each one to a specific agency and a specific product. Clearbox Source now connects more than 100,000 real-world implementations this way.
Vendor case studies are a useful starting point, but on their own they are marketing. What makes them useful to a buyer is the context around them: how big the agency is, what it already runs, what else it has purchased, and how other agencies have bought the same product. When we find a deployment worth learning from, we connect it to that context and put it in front of the agencies using Source, on the agency and product pages where they already do their research.
We share it because government learns best from government. In our 2026 customer survey, 92% of agencies said they look to peers for ideas, but only 60% had actually collaborated directly. Making real implementation evidence easy to find is how we help close that gap, and it is the fastest way to spread AI that actually works. That is the idea behind a Government-Only Vendor Intelligence Network (GOVIN): real implementation data, shared among agencies, instead of scattered across vendor blogs and the open internet.
If your agency is weighing an AI project, start with who has already done it. Look up any agency or product in Clearbox Source to see where it is deployed, what peers bought, and how to buy it. Any state or local government employee can get started with a free Clearbox Starter account. And if your agency has an AI implementation with results worth sharing, tell us. We would like to feature it.