The single industry that's invisible to current AI tooling.

The single industry that is most invisible to current AI tooling is municipal government. Not federal, not state, not even county in the urban states. Municipal. The two thousand or so American cities and the roughly nineteen thousand smaller towns and villages that sit underneath them. The places where a building permit takes seven weeks because the zoning desk is one person and that person is on bereavement leave. The places where a water main break is logged in a spreadsheet that lives on a single workstation in a basement. The places where the agenda for tonight’s council meeting is a PDF scanned from a typewritten template that has been in use since 1983. The places where the AI conversation in early 2024 has not arrived and is not arriving and will not arrive on the timeline anyone in San Francisco or Boston or Seattle expects.

I want to be precise about what I mean by invisible. I do not mean that nobody is selling AI to cities. There are, by my count, somewhere between forty and sixty startups in early 2024 with a pitch deck that includes the words “AI for local government” somewhere on slide three. I mean that the actual operating reality of how a city works, and the actual operating reality of how the AI tooling shipping in 2023 and 2024 is shaped, do not intersect. The tooling is shaped for individual knowledge workers at companies that have already been through several waves of digital transformation. The municipal operating reality is shaped by procurement rules written before the internet, vendor lock-in to software systems that were specified in the 1990s, elected-official turnover on two-and-four-year cycles, union contracts that govern who can touch which workflow, public-records-act exposure on every keystroke, and a workforce whose median tenure is closer to twenty years than to two. None of the AI products being shipped in 2024 are designed for that environment. The pull on municipal government from current AI tooling is, as a first approximation, zero.

This matters more than the conversation in early 2024 registers, because municipal government is where most Americans actually touch the state. The state, for most people, is the building department, the school district, the parks-and-recreation desk, the property-tax office, the water utility, the police non-emergency line, the public library, the snow-plow dispatch, the trash-pickup schedule, the parking-ticket appeal, the marriage-license clerk. The state is not the IRS. The state is not the Department of Defense. The state is the thirty-five thousand to seventy thousand municipal employees who operate, in aggregate, the largest single category of American public-facing labor. They are the surface area on which the average citizen experiences whether government works. And in 2024 it does not work particularly well, and the AI revolution is not, as currently configured, on a path to fix that.

The reason this is invisible to the AI conversation is not that the AI builders are stupid. The reason is that the AI conversation is happening inside a particular social and economic geography (the coastal-tech-hub circuit) and the municipal operating reality is a different geography that the coastal-tech-hub circuit rarely touches. The founder shipping the document-summary copilot in San Francisco is not, generally, sitting in a Tuesday-night planning-board hearing in a town of twelve thousand in Pennsylvania. The investor underwriting the Series B is not, generally, the person who has spent six years watching a city of forty thousand try and fail to modernize its permit-tracking software. The training corpus that the foundation models are built on contains very little of the actual textual artifact of municipal government, because that artifact lives in scanned PDFs and in proprietary software systems and in physical-paper-only records that no web crawler has ever touched. The model does not know what a CAFR is. The model does not know how a Robert’s-Rules motion to table differs from a motion to postpone indefinitely. The model has been told, in its system prompt, that it is a helpful assistant; it has not been told that the underlying epistemic-and-procedural reality of municipal government is alien terrain.

Consider the specific shape of the work. A mid-sized American city, call it sixty thousand residents, runs on roughly four hundred discrete recurring workflows. Issue a building permit. Process a payroll for a police department. Reconcile a water-utility billing run. Approve a special-events permit. Calendar a planning-board agenda. Mail out a property-tax bill. Each of those workflows touches between three and seven legacy software systems, two to four union jurisdictions, one to three elected boards, a state-mandated reporting requirement, and a public-records-retention rule. The work is not knowledge work in the way that the AI tooling is shaped for. It is procedure work. It is the application of codified rules to specific cases under hard constraints from above and below. It is what a bureaucracy actually is, in the technical sense of the word, and the AI tooling shipping in 2024 is being built for the not-bureaucratic case.

The closest analog in the private sector is mid-market insurance back-office work, which is itself a category the AI conversation has barely touched. Both run on codified procedure, both have decades of legacy software, both have workforces with deep institutional memory and short technological exposure, both are regulated in ways that constrain what can be automated and what cannot. The difference is that the insurance back office has shareholders pushing on cost; the municipal back office has voters who are largely unaware of the cost and elected officials who have no career upside in surfacing it. The pressure differential matters because the AI builders, when they go looking for a market, follow the pressure. The pressure in municipal government is the wrong shape: it is episodic (a budget crisis every six to eight years), it is local (no national procurement, no national standards), and it is political (the people with the budget authority are not the people doing the work). The combination is repellent to the venture-funded software model. So the venture money does not arrive, and the tooling does not get built, and the gap between what AI could do for a city and what AI is currently doing for a city stays wide.

The obvious objection is that this gap will close on its own as the foundation models get better and the municipal-government category becomes more obviously valuable. I think that is wrong, or at least substantially incomplete, and the reason is structural. The foundation models getting better does not solve the procurement problem. It does not solve the union-jurisdiction problem. It does not solve the public-records-act problem. It does not solve the legacy-vendor-lock-in problem. The model getting better from GPT-4 to GPT-5 to GPT-6 is mostly orthogonal to the question of whether a city of sixty thousand can actually deploy a useful AI tool against its building-permit workflow. The bottleneck is not model capability. The bottleneck is the wrapper of institutional, contractual, and regulatory constraint that the model has to be deployed inside, and that wrapper is not getting better on the same curve as the model.

What would actually change the situation is a different kind of company than the venture-funded AI startup. It looks more like a long-cycle, capital-patient, operationally-deep public-benefit corporation that commits to the municipal market for a decade and accepts gross margins in the thirty-to-fifty-percent range rather than the ninety-percent range that SaaS investors underwrite. It looks more like the firms that built out municipal-utility billing software in the 1990s, with a modern AI-native architecture grafted on, and a posture toward the customer that is genuinely operational rather than primarily transactional. It looks like a firm whose founders have actually held municipal jobs, or consulted into municipal jobs for a decade, before starting the company. It looks unlike most of what is being funded in the AI boom of 2024.

I think this is one of the larger civilizational questions of the current moment, and I think it is underappreciated by a factor of about ten in the AI conversation. The reason is the Kardashev-scale framing. If the goal of the AI revolution is to push humanity up the curve of what a civilization can collectively do, the rate-limiting step is not the frontier capability of the models. The rate-limiting step is the diffusion of the capability into the institutions that mediate the average citizen’s actual interaction with the collective. A civilization in which the foundation models can pass the bar exam and write production code and diagnose rare diseases, but in which the average citizen still spends seven weeks waiting for a building permit because the zoning desk has not been touched by any of it, is a civilization that has built an extraordinary engine and bolted it to a horse-cart. The horse-cart is not going to keep up. And the citizen, walking into the building department, is not going to feel the civilizational lift; they are going to feel the same friction they felt in 1994. The AI revolution that does not reach the building department is, from the citizen-experience perspective, not the AI revolution. It is something else, narrower and less consequential, which we are mistaking for the real thing.

I will name three concrete forecasts on a five-year horizon. First, the first municipal-government AI company that crosses fifty million in revenue will be founded by someone who has worked in or alongside municipal government for at least a decade, will not look like a typical AI startup, and will be on a capital structure that the current venture model is unlikely to underwrite. The capital will come from a different pool: public-benefit-oriented family offices, civic-tech philanthropy at scale, possibly a sovereign or state-pension allocation that recognizes the infrastructure-grade nature of what is being built. Second, the dominant interface for municipal AI tooling will not be a chat window or a copilot panel; it will be an integration into the existing government-software stack (Tyler Technologies, Granicus, OpenGov, the category leaders of the legacy era) with the AI doing its work as a layer underneath the interface the clerk already uses. The startups attempting to displace the legacy stack from the front will inevitably lose. The startups attempting to ride the legacy stack from underneath will win. Third, the distribution of the gain will be uneven in a way that the AI conversation has not yet priced in. The cities that adopt early will pull years ahead of the cities that do not, and the gap will be visible in permit-issuance times, in tax-collection efficiency, in public-records responsiveness, in 311-call resolution, and ultimately in the lived experience of being a resident. Some cities will become noticeably easier places to live. Other cities, fifty miles away, will not. The political consequences of that divergence will be larger than the AI conversation currently expects.

The longer-run shape is the one I find most worth sitting with. Municipal government is, in a sense, the last large category of American work that has been unimproved by the past three decades of software. The consumer internet improved consumer experience. Enterprise software improved enterprise productivity. Cloud infrastructure improved the underlying engineering substrate. Municipal government got, in net, almost none of it. The AI revolution is the next chance, and possibly the last chance for a generation, to bring the same kind of compounding improvement to the surface area on which most Americans actually meet the state. If the AI revolution misses this, it will have missed the single highest-leverage civilizational intervention available to it. If it hits this, the building department of 2034 will not look like the building department of 2024, and the citizen walking in to file a permit will, possibly for the first time in their adult life, feel the state working at the speed of the rest of their world. That is a small example with a large meaning. It is what the diffusion of the AI revolution into the institutions of everyday life actually looks like, on the ground, in the place the citizen actually lives.

I do not know who will build the firm that does this. I know it is not, in early 2024, any of the firms that the AI conversation is currently watching. I will be watching for it, and I will be writing about it when it appears, because the company that gets this right is, on a serious civilizational-leverage assessment, one of the more important companies of the next decade. The invisibility of the category to the current AI tooling is the opportunity. Someone will see it. The shape of what they build will tell us a great deal about whether the AI revolution is the thing the optimists hope it is, or whether it is something narrower and less consequential, which the citizen at the building-permit counter would experience as a continuation of the same.

—TJ