Artificial intelligence may be one of the most important productivity-enhancing technologies of the coming decades. That promise extends to city hall. AI could give municipal governments more administrative capacity without requiring commensurate increases in spending or head count.
Whether it does, however, will depend not only on what the technology can do but also on the rules cities write around its use. Consider New York City. Its official guidance specifies that employees should access AI systems only after receiving approval from “responsible agency personnel.” It also instructs employees to consider whether the benefits of any given use outweigh the energy and water consumed in producing it.
Compare that approach with San Francisco’s. The city has given roughly 30,000 employees access to an approved tool and holds each accountable for what it helps produce. An employee reading San Francisco’s guidance would reasonably conclude that the city wants its workforce to use the technology. An employee reading New York’s would likely conclude the opposite.
That contrast captures an emerging divide among America’s cities. A growing number of major American cities have issued AI policies, but their approaches vary widely. Some frameworks make responsible use easier. Others make it harder.
What divides the wheat from the chaff is a distinction between a presumption of use and a presumption of control. Some frameworks begin by asking how employees can deploy AI safely. Others begin by asking how officials can prevent misuse. Bureaucracies respond more reliably to defaults than to mission statements, so those starting points can produce very different results.
Municipal government is especially well suited to AI-assisted productivity gains. Personnel costs dominate most city budgets, and much of the work is precisely the work current AI systems do well: processing permits, drafting reports, answering 311 calls, translating documents, and reconciling records.
Even modest productivity gains in these fields could yield big benefits. San Francisco’s six-month AI pilot involving more than 2,000 employees reportedly found savings of up to five hours per worker each week. That’s consistent with a preregistered public-sector experiment that found that employees completed document-heavy tasks roughly one-third faster and at higher quality using AI (though it found no gains for data analysis). That distinction matters because municipal administration is overwhelmingly document-heavy.
Cities are in fact unusually revealing laboratories for studying how a new technology like AI spreads. Unlike private firms, municipal governments can’t relocate or shed many unproductive responsibilities. Their gains will depend heavily on institutional choices rather than market pressure.
Those institutional choices can have profound consequences. The political scientists Pamela Herd and Donald Moynihan have shown that administrative burdens can suppress adoption of government programs. The same logic applies inside government. An employee’s willingness to adopt a sanctioned AI tool depends in part on what each use costs in terms of time, approvals, and personal exposure to the technology itself.
Five questions help reveal those costs:
Is employee use presumed permitted with guardrails, or prohibited absent approval?
How many approvals stand between an employee and sanctioned use?
Does the city provision an approved tool at all?
Who bears accountability for outputs: the user or a gatekeeping office?
Does the framework distinguish between low- and high-risk uses, or apply similar scrutiny to both?
The answers carry real weight. A city that regulates AI without giving its workforce any sanctioned option has created a de facto prohibition. A framework that applies the same scrutiny to basic uses as it does to a system that helps determine eligibility for public benefits has not governed risk so much as rationed permission.
I reviewed publicly available policies, governance frameworks, procurement guidance, and employee-use directives in 12 major jurisdictions. I compared jurisdictions across five dimensions: default permission, approval requirements, provisioning of approved tools, accountability mechanisms, and risk-tiering. The resulting spectrum is qualitative rather than numerical and reflects written rules rather than the full reality of agency-level practice.
The results reveal a spectrum of municipal approaches, from cities that make approved use routine to those that burden it with excessive procedural hurdles.
At the permissive end sit San José, Los Angeles, San Francisco, and Miami-Dade. Los Angeles has deployed Google Workspace with Gemini across its entire 27,500-person workforce in 45 departments. San Francisco provided Copilot Chat to nearly 30,000 employees after its pilot and updated its guidelines to enable use while holding staff responsible for outputs. San José built a citywide platform for its 7,000 employees and founded a coalition to share governance templates with other governments. Miami-Dade has adopted countywide AI guidance, workforce training, and a deployment strategy.
The permissive cities also had inexpensive options. San Francisco’s Copilot access came bundled with its existing Microsoft license at no new cost. Los Angeles was a longtime Google Workspace customer. This accidental acquisition may be a benefit. Research on how cities actually buy AI finds that tools that are acquired through cooperative “piggyback” contracts or under routine cost thresholds can bypass much of the normal procurement gauntlet.
A middle tier—Boston, Austin, Seattle, Phoenix, and Washington—permits adoption while attaching reporting and oversight requirements of varying weight. These cities are the swing group, and the direction of their revisions over the coming two years will indicate which presumption ultimately wins out. Boston is closest to the permissive group: it has switched on Google’s Gemini across its Workspace environment and holds employees accountable for their output, but it still gates access behind mandatory training.
At the control end sit New York, Philadelphia, and Chicago. Philadelphia and Chicago are there mainly because they are still building advisory structures ahead of broader deployment.
New York is different. Its framework is fully developed, sophisticated, and unusually cumbersome. Employees need official clearance to use AI. The city has largely declined to provide a default tool. Responsibility rests heavily with gatekeeping offices rather than the employee who owns the work. And the framework does little to distinguish low-stakes internal tasks from systems affecting benefits, rights, or public services.
Its first guiding question subjects the technology to an unusually demanding necessity test: “Is GenAI the only feasible way to achieve the stated purpose?” No government applies a necessity test to spreadsheets, word processors, or search engines.
Employees are also directed to consider whether AI use justifies its energy and water consumption. Systems that affect worker productivity may require consultation and pilot testing. The guidance even warns that the burden of verifying AI outputs may erase anticipated productivity gains.
Municipal administration is full of document-heavy, process-bound work that current AI systems can handle well. A city that makes responsible use the path of least resistance could recover meaningful capacity across permitting, casework, correspondence, and other municipal functions.
Embracing AI requires cities to do more than buy software. It requires them to adopt a posture of permission by default for low-stakes work, clear accountability for employees, and real scrutiny reserved for the systems that affect benefits, rights, and public decisions. The divide that will matter, in the end, is between governments that have made adoption cheap and those that have made it expensive.
> A city that regulates AI without giving its workforce any sanctioned option has created a de facto prohibition.
What will *de facto* happen is that employees will end up using unauthorized AI and then lying about it.
The cities may learn, as many private companies are learning , that the cost outweighs the benefits