91,729 Governments. One AI Playbook Won't Work.

Anthropic is having a good autumn in the public sector, and it deserves credit for it.

In June, California announced an agreement making Claude available to state agencies at a 50 percent discount, with free workforce training and technical assistance. Cities and counties could access the same discounted offer. On October 27, Anthropic expanded its agreement with the General Services Administration, offering 40 percent off Claude Enterprise, Claude for Government, and the Claude API to federal, state, local, tribal, and territorial governments, with no minimum spend and additional implementation support planned.

Some of that support is available today, and some is still coming. But making powerful AI more affordable, accessible, and easier to implement in government is exactly the kind of progress we should celebrate.

Having spent much of my career in senior leadership roles with the City of San Diego, and now working in enterprise technology and AI, I genuinely want this to succeed.

But I also have a friendly challenge for the technology industry.

We have a habit of building products for private enterprise, adapting them for the federal government, and then assuming the same approach can scale across state and local government.

I'm not suggesting Anthropic is making that mistake. I'm suggesting that, as AI adoption accelerates, the industry needs to be careful not to repeat it.

Because the United States has 91,729 local governments, according to the Census Bureau's 2026 government organization counts.

That's not 91,729 interchangeable customers. It's tens of thousands of institutions with different responsibilities, budgets, laws, labor agreements, political environments, organizational cultures, and community expectations.

The challenge isn't simply implementing AI technology. It's integrating AI into the institution of government.

The Operating System of Everyday Life

Most people don't spend much time thinking about local government until something stops working.

The water runs. Someone answers 911. Roads get maintained. Parks open. Garbage gets collected. Building permits get issued. Public health inspectors help keep our communities safe.

Local government is the operating system of everyday life. Much of what makes our communities function depends on public employees doing work most residents will never see.

And those governments are extraordinarily diverse.

Of the tens of thousands of local governments in America, many aren't cities or counties at all. They're school districts, water districts, fire districts, and other special-purpose governments responsible for essential public services.

Even neighboring cities can operate in fundamentally different ways.

One might run its own police, fire, and utility departments. Another might contract for those services. One might have numerous bargaining units, a large technology department, and a strong mayor. Another might have a small administrative staff, a council-manager government, and limited internal technical expertise.

They can share a border while having remarkably little in common operationally.

Local government is not a smaller version of federal government. And it certainly isn't just a slower version of private enterprise.

It's also a significant part of the American workforce. Approximately 14.7 million people worked for local governments in March 2025, and Bureau of Labor Statistics figures put local government at nearly one in ten nonfarm payroll jobs.

In many communities, government is a major employer, providing stable careers, benefits, economic opportunity, and institutional knowledge that is difficult to replace.

None of this means local government is beyond criticism. Budgets are strained, infrastructure is aging, permitting can take too long, and residents understandably expect better service.

Those are compelling reasons to embrace innovation.

But they're also reasons to be thoughtful about how we introduce it.

Labor Isn't the Obstacle. The Pitch Is.

In the private sector, AI is frequently marketed through the language of productivity, operating margins, cost reductions, and fewer employees.

Bring that same pitch to city hall, and you may quickly find yourself at odds with the workforce, labor organizations, elected officials, and the community.

Government has a legitimate obligation to control costs and steward taxpayer resources. But government isn't a profit-maximizing enterprise.

Government is fundamentally a service organization. Its success should be measured by how effectively it serves the public.

During my years in city leadership, I rarely encountered employees who opposed technology simply because it was new.

In fact, many were eager for tools that could take administrative work off their desks.

Anyone who has worked inside a large government organization knows how much time can be consumed by reporting, documentation, compliance, coordination, and the countless unexpected demands that arise when running complex public institutions.

That work often pulls skilled employees away from the very responsibilities that motivated them to enter public service.

Where I observed more resistance was when technology moved beyond supporting employees and began potentially replacing the professional judgment, expertise, and human relationships at the center of their work.

There is a real difference between automating the paperwork around public service and automating public service itself.

That distinction helps explain why the narrative that public-sector unions are inherently opposed to AI is too simplistic.

Consider San Diego County's 2025 to 2028 agreement with SEIU Local 221.

The agreement expressly recognizes the importance of AI advancement while establishing principles around safety, equity, transparency, data protection, employee consultation, and job preservation.

It identifies enhancing employee efficiency, rather than replacing bargaining-unit jobs, as a goal. It also requires that AI-derived information used for employee discipline be corroborated by evidence that was not collected through AI.

Or look at the San Francisco Unified School District's agreement with SEIU Local 1021, which includes protections against AI replacing bargaining-unit work without union agreement, provides for paid employee training, and restricts AI-based employee monitoring and evaluation.

These agreements don't mean every union supports every AI deployment. Concerns about displacement, surveillance, contracting out, deskilling, and the erosion of professional judgment are legitimate.

But they demonstrate that labor organizations can help establish the conditions for responsible AI adoption rather than simply opposing it.

And I think there's an even more important opportunity.

Public employees aren't just stakeholders whose permission we need. They're often the people best positioned to identify where AI could actually improve government.

A planner understands which applications repeatedly arrive incomplete. A social worker knows which administrative requirements consume time better spent helping families. A public works engineer understands the exceptions and operational realities that rarely appear in a formal process diagram.

These employees understand what works, what doesn't, and what happens when something goes wrong.

Their knowledge is invaluable when designing effective AI applications.

Instead of starting with the question of how many government workers AI could replace, what if we started by asking how much more those workers could accomplish for their communities?

AI might help agencies handle growing workloads, reduce overtime, avoid unnecessary consulting expenses, or provide better services without constantly expanding staffing.

Those are legitimate efficiency gains.

But there is a meaningful difference between using AI to expand the capacity of the workforce and making workforce reduction the primary objective.

Just Because AI Can Make a Decision Doesn't Mean Residents Want It To

The workforce is only one part of the equation.

The people government serves have their own expectations about where technology belongs.

The 2026 Carnegie California AI Survey found that only 7 percent of respondents unconditionally supported state and local governments using AI to make decisions affecting them. Another 36 percent supported it if it improved decision quality, while 45 percent opposed it.

That doesn't mean residents reject every government application of AI. It does mean unconditional public acceptance should never be assumed.

And the stakes are different depending on how the technology is used.

Consider drinking water.

AI could help utilities analyze sensor data, identify unusual conditions, and flag potential problems faster than traditional monitoring approaches.

That seems promising.

But how comfortable would residents be with AI determining whether their drinking water is safe? What evidence would they want? Who would be accountable if the determination was wrong?

Now consider police reports.

AI-generated police narratives are already a reality. Tools such as Axon's Draft One can generate draft reports using body camera audio.

This could reduce administrative work and potentially allow officers to spend more time serving their communities.

But those reports can influence criminal investigations and the legal rights of individuals. Questions about accuracy, transparency, and accountability are not minor implementation details.

California has already enacted legislation requiring disclosures and records retention related to AI-generated police reports, illustrating how the technology's use creates new legal and public accountability obligations.

Or take something seemingly less controversial: deciding which roads should be resurfaced.

AI could analyze pavement conditions, traffic volumes, maintenance histories, and projected costs to help public works departments make better recommendations.

But deciding which neighborhood receives investment first is not purely a technical exercise. It also involves community priorities, fairness, available resources, and elected leadership.

In each example, the technology could provide significant value.

But there's an important difference between AI informing a decision, recommending a course of action, creating an official record, and making the final determination.

Those roles carry different risks, responsibilities, and expectations.

My own experience working with residents and businesses has reinforced another point that I think the technology industry often overlooks.

Most people aren't asking for an AI government. They're asking for a government that works.

They want someone to answer the phone. Someone to explain a complicated process. Someone who can help resolve a problem, exercise judgment, and take responsibility when something goes wrong.

They want someone they can appeal a decision to.

They want public employees who know their community and engage with them as partners.

That doesn't mean everyone wants to wait on hold or visit city hall for a routine transaction. AI-powered tools could make government more accessible, particularly for people facing language barriers, transportation challenges, or demanding work schedules.

If someone wants to renew a permit at 10 p.m. without speaking to anyone, that's a good service outcome.

But convenience and human accountability aren't mutually exclusive.

The opportunity is to use AI to make routine interactions easier while preserving meaningful access to people when judgment, discretion, empathy, and accountability matter.

And simply keeping a human in the loop doesn't automatically solve that problem.

A reviewer who lacks the time, information, or authority to challenge an AI-generated recommendation provides little meaningful oversight.

Nor should communities be expected to accept every AI application merely because it meets a technical performance standard.

Sometimes residents will want a public discussion, additional safeguards, or limits on how technology influences decisions.

That isn't necessarily resistance to innovation.

It's citizenship.

Scale the Technology. Adapt the Institution.

All of this brings us back to Anthropic's announcement and the larger challenge facing the technology industry.

Technology companies need scale.

They need products that can be sold repeatedly, consistent support models, standardized technical implementations, and business models that don't require rebuilding everything for every customer.

Those are entirely reasonable requirements.

But local governments need technology that works within their particular legal obligations, labor agreements, governance structures, organizational capabilities, and community expectations.

Those requirements are reasonable too.

The challenge is figuring out what should be standardized and what must remain local.

We don't need 91,729 different AI products.

The underlying technology can and should be scalable.

Foundation models, security controls, identity management, APIs, testing frameworks, and common technical capabilities can serve thousands of institutions.

Many workflows can also be adapted across jurisdictions. Permitting, code research, service requests, public works analysis, and administrative operations often share common underlying patterns.

But the institutional environment in which those tools operate is not interchangeable.

Consider two neighboring cities purchasing the same AI platform.

One operates its own permitting department with represented employees, established labor agreements, and local policies requiring public disclosure of automated systems.

The other contracts out much of its permitting work, has limited internal technical expertise, and operates under different administrative and political expectations.

Both might benefit from exactly the same underlying AI capabilities.

But the process of introducing those capabilities into daily operations will be substantially different.

This is where I think the distinction between technical implementation and institutional integration becomes critically important.

Technical implementation involves configuring the software, connecting data, establishing security controls, integrating systems, and training users.

Institutional integration involves understanding how that technology changes service delivery, employee responsibilities, legal compliance, organizational accountability, and the relationship between government and the communities it serves.

Anthropic's commitment to training and implementation assistance is an important step. But the hardest work may not be getting Claude into government employees' hands.

It may be figuring out how those tools should be used within thousands of different public institutions.

That work requires people who understand both government operations and modern AI technology.

It requires technical expertise, certainly. But it also requires familiarity with public administration, labor relationships, organizational change, policy, and community expectations.

The smallest jurisdictions may need that combination of expertise most, because they often lack the staff and resources to manage complex technology programs internally.

The solution isn't necessarily to build bespoke applications for every government. It's to combine scalable products with adaptable workflows, knowledgeable implementation partners, and locally informed leadership.

There's a broader lesson here for enterprise software.

As AI models become more capable, affordable, and accessible, competitive differentiation increasingly depends on the quality of the data, the workflows surrounding the models, and the ability to reliably solve real problems.

Government is no different.

Except that when government technology fails, the consequences can involve essential services, legal rights, public resources, and community trust.

Start With the Problem, Not the Technology

I would offer two challenges as we enter this next phase of government AI adoption.

To technology companies: recognize that local governments are not interchangeable customers. Build products that scale, but design implementation approaches that accommodate institutional differences. Work with people who understand the complexities of government, not just the capabilities of technology.

And to government leaders: resist the temptation to start with a technology solution.

Start with the problems your employees and residents need solved.

Where are residents waiting too long for answers? Which administrative burdens keep skilled employees from delivering services? Where is information difficult to access? Which workflows are frustrating residents and businesses? What problems have persisted because the organization lacks capacity, coordination, or effective tools?

Bring employees into those conversations. Understand the needs of the community. Then determine whether AI is the appropriate solution.

Sometimes it will be.

Sometimes the answer will be conventional software, a redesigned process, better coordination, additional staffing, or simply fixing a policy that no longer makes sense.

AI should not be the predetermined answer.

Anthropic deserves credit for helping open the door to broader AI adoption across government. Lower costs, better access, training, and implementation support are meaningful contributions.

But getting through that door is only the beginning.

Scaling AI across America's 91,729 local governments doesn't mean doing the same thing 91,729 times. It means creating technology that can scale while respecting the institutions, employees, and communities in which it operates.

The goal shouldn't be to automate government for the sake of automation, or to measure success by how many licenses are deployed or positions eliminated.

It should be to help public employees do more of the work that matters, make government more responsive to the people it serves, and solve problems that have challenged communities for years.

The most successful government AI deployments may ultimately be the ones that make government feel more human, not less.