The City Hall AI Nobody Approved

The city of Hopkinsville, Kentucky, had a peculiar problem: its employees were already experimenting with artificial intelligence, but nobody could say exactly where. Some used familiar chatbots. Some encountered AI inside software the city had already bought. The municipality had tax information, health details and other resident data to protect. A ban might sound tidy on paper. It would also leave the actual use, already under way, out of sight.

Hopkinsville did something less theatrical than banning AI. It formed an advisory board of department leaders, asked employees how they were actually using the technology, and wrote a deliberately lean policy: keep private information out of public tools, use sanctioned systems, and have a person review AI-assisted work before it goes out. Darwin AI supplied the visibility and controls to make those instructions more than an annual training slide.

Then came a far less abstract test. The fire department had to place qualified people on the right trucks and at the right stations while accounting for leave, sick calls and minimum staffing. According to the city’s account, battalion chiefs had spent roughly three to four hours on the task in a day. The first scheduling workflow built with Darwin LaunchPad brought that to about two minutes after staff supplied availability. A human can still inspect and override the result. The number is striking; the sequence matters more. First the city learned what its people were doing. Then it set rules. Only then did it automate a job worth automating.

The machine was already in the building

Darwin was founded in 2024 by Noam Maital and Asaf Fadida. Maital is chief executive; Fadida is chief technology officer. Their first market was not a blank slate waiting for an AI product. It was state and local government, where software procurement is deliberate but staff can open a chatbot in seconds. At the company’s March 2025 emergence from stealth, it said its own data indicated more than 40% of the public-sector workforce was already using AI tools, often through personal accounts or unapproved platforms. That is the kind of figure a security officer reads twice.

The initial product, Darwin Govern, tries to turn that hidden activity into a living inventory. It detects browser tools, native applications and AI features embedded in other products. An agency can classify what it finds, approve tools, attach rules to users or departments, and keep records of what happens. Controls can warn, block, redirect or redact. A sensitive file upload is no longer merely a clause in a policy; it becomes an event the software can catch.

That order is Darwin’s particular pitch. General enterprise security products can flag risky traffic, and an agency can write a policy without buying Darwin. But public organizations also contend with records retention, citizen data, state rules, elected oversight and purchasing processes. Darwin says it was designed for those demands. It sells a connected pair of products: Govern for visibility and enforcement, LaunchPad for specific tasks such as scheduling, HR, procurement or permitting. The company also offers a free AI Policy Wizard to help public bodies draft or benchmark their starting rules.

“We didn’t want to take a reactive approach to AI.”Amanda Brunt, technology and public information director, City of Hopkinsville

The first failure was a missing map

In Hopkinsville, the advisory board’s first discovery was that it did not know the extent of AI use across city departments. That ignorance shaped the response. The city surveyed staff and kept its rules short enough to be remembered. City Administrator Troy Body had returned from an AI conference at Harvard convinced the city needed a deliberate strategy. The city’s account describes a shift from anxiety about shadow use to a practical question: which real municipal problems could be solved safely?

The fire schedule was a good answer because it already had constraints people understood. A paramedic must be on every fire truck; specific skills may be needed ahead of storms; holidays and absences do not excuse a staffing gap. Software can calculate a proposal, but the duty officer still owns the decision. This is the part of the story another agency can copy without buying the same platform: form a cross-department group, measure existing use, write a few clear rules, choose one repetitive process with checkable inputs, and keep a named human responsible for its output.

The result should not be mistaken for a universal ROI calculation. Darwin does not publish a general price list for its core platform, and the Hopkinsville account does not disclose its contract cost. A scheduling model also depends on clean availability data, agreed staffing rules and a team willing to test its suggestions. If those ingredients are absent, the headline number will not travel. The city’s two-minute claim belongs to that city’s first phase, not to every fire department.

A company learning government’s buying language

The market response has been visible in funding and in contracts. Darwin announced a $5 million seed round in March 2025, co-led by UpWest and Resolute Ventures. In October it announced a $15 million Series A led by Insight Partners, with the earlier backers participating. The company said the money would support product development, deployments, policy work and training. It has not disclosed revenue or a valuation.

Customers and partners suggest why that implementation work matters. CapMetro in Austin used Darwin to understand AI activity across departments. Georgia’s Technology Authority announced a June 2026 partnership to build statewide governance groundwork. Darwin named government deployments that include Hopkinsville, the City of Fishers, and the Central Ohio Transit Authority. The users are IT and security teams, policy leaders and operational staff, each interested in a different part of the same question: can this tool improve work without losing control of data or decisions?

Buying the answer is its own government problem. In June 2026, Darwin announced a TXShare cooperative purchasing route through Civic Marketplace. In August it announced a Texas Department of Information Resources bulk purchase agreement through Carahsoft, offering eligible public organizations state-negotiated pricing. The price itself was not published. These agreements matter because a public agency can spend months establishing a legal route to buy software. A product that shortens that route can get tested while the policy problem is still fresh.

Darwin’s website now describes Govern and LaunchPad as two layers of one platform. It claims more than 1,000 AI tools covered, over 5,000 policy violations prevented and more than a million personally identifiable information redactions. Those are company-wide figures, useful for scale but without the agency-by-agency detail needed to judge each outcome. Recognition in the 2026 GovTech 100 and a TX-RAMP Level 2 certification add signals that the company has been building for this market’s standards and scrutiny.

Permission with an audit trail

The most revealing thing about Darwin may be that its sales story starts with a concession: staff did not wait for headquarters to finish an AI strategy. In some offices, that feels like a failure of discipline. In reality, it is how new tools often enter work. Darwin’s bet is that agencies will pay for a map of that reality, rules that operate in real time, and a way to use the technology openly for jobs that can be measured.

Hopkinsville’s fire schedule is a neat illustration because it is so unromantic. Nobody asked a chatbot to reinvent the city. The city asked for a roster that meets its own requirements, arrives faster and leaves a person in charge. In public service, that may be the more interesting AI trick: a tool that can show its work well enough for somebody else to answer for it.