City Hall Has a Productivity Problem. AI Is the Most Plausible Fix It Has Seen in Decades.

Sixty-seven percent of OECD countries were already using AI to improve public service design and delivery as of 2024, according to the OECD's "Governing with Artificial Intelligence" report. That figure matters less as a headline than as a baseline: AI has moved from conference-room aspiration into actual government operations across most of the developed world. The more interesting question is why local governments, the layer of government that most residents interact with most often, are still largely watching from the sideline.

The answer starts with a structural problem that predates AI by several decades. Most municipalities are sitting on enormous volumes of data covering traffic patterns, building permits and maintenance records, but that data is scattered across departments that were never designed to communicate with each other. Staff spend hours pulling information manually from systems that cannot connect, then synthesizing it by hand. The result is that decisions get driven more by institutional habit than by evidence. A World Bank Digital Progress and Trends Report published in 2025 noted that public-sector digital transformation has consistently underdelivered relative to expectations, which is a polite way of saying that governments have accumulated technology without building the underlying capacity to use it well. AI does not automatically fix that, but it is the first tool in a long time that directly addresses where the bottleneck actually lives.

The permit counter is the experience most small business owners know by heart. You arrive prepared, documents in hand, morning cleared. What you have not accounted for is the discovery, delivered with bureaucratic serenity, that the form you need is only available from a different office, which operates on a schedule apparently designed to conflict with every other office in the building. Three hours later, you have filled out four forms requesting the same information in slightly different orders, and the next review slot is six weeks out because the scheduling system predates the smartphone. This is not a caricature. It is a description of process friction that accumulates invisibly inside institutions built for a different era and never fully redesigned. The staff involved are usually doing their best inside systems that were not intended for speed or coherence. The problem is structural, which is exactly why individual effort alone cannot fix it.

"Most municipalities are sitting on enormous volumes of data, but that data is scattered across departments that were never designed to communicate with each other. Staff synthesize it by hand. Decisions get driven more by institutional habit than by evidence."

Residents have noticed the gap between what digital services can feel like and what city hall actually delivers. When a banking app resolves a dispute in under two minutes and a retailer can anticipate a reorder before you think to place it, "please call back during business hours" stops reading as a policy and starts reading as a statement about priorities. For a small business owner trying to open a location or pull a permit on a deadline, that gap has a direct dollar cost. The question AI is beginning to answer is whether the gap is actually inevitable, or just familiar.

What makes this moment different from previous rounds of government technology investment is the combination of two shifts happening simultaneously. First, the OECD's analysis finds that 57% of documented government AI use cases focus on automating or tailoring services, with a further 45% involving decision support and forecasting. These are not experimental applications. They target the specific operational failures that have frustrated residents for years: fragmented information, slow processes, and decisions that arrive without explanation. Second, generative AI lowered the technical barrier to entry enough that tools requiring custom development a few years ago can now be accessed through standard platforms that a non-technical procurement team can actually evaluate. A mid-sized city planning department running on a 2004 database is still not in an easy position, but it is in a better position than it was. The vendor ecosystem has transformed, and that changes what is realistically on the table for cities that have historically been priced out of serious technology investment.

The tension worth naming at the outset is this: the national policy environment is pushing hard toward AI adoption, while most local governments are barely organized to respond. A 2024 survey by the International City/County Management Association found that 48% of local government respondents consider AI a low priority, and fewer than 6% call it a high priority. Only 9% have an organization-wide policy governing AI use. Mayors who are still treating AI as a future-agenda item are making a choice, even if it does not feel like one, and the cities that spent the last two years running careful pilots are now in a position to scale while others are starting from scratch under pressure.

What Local Governments Are Actually Deploying Right Now

Start with what is not happening: local governments are not deploying sentient robot administrators. What they are deploying is considerably more useful, which is targeted tools that automate specific, high-volume, low-judgment tasks that have historically consumed enormous amounts of staff time. The gap between AI in the headlines and AI in actual municipal use is wide, and the cities seeing real results are the ones that ignored the headlines and focused on the mundane.

Resident-Facing Chatbots and Permit Assistance

The most common entry point is a conversational AI tool that handles the questions city staff answer fifty times a day: office hours, business license renewals, trash pickup schedules, permit requirements. These are not intellectually demanding questions, but they consume real staff hours and, when answered inconsistently or slowly, erode resident confidence in ways that accumulate quietly. The ICMA's 2024 survey of local governments found that 55% of respondents identified resident engagement as the top area where AI has near-term potential, which tracks with where actual deployments are concentrating. Atlanta is the clearest current example: GovTech's 2024 Digital Cities Survey recognized the city for implementing AI-powered chatbots specifically to reduce wait times for residents interacting with city government, as part of a broader push on digital experience. That is not a pilot. It is a production deployment in a major American city.

Permitting is where the chatbot use case gets particularly valuable for small business owners. Applicants get accurate, consistent guidance at any hour, the process feels less opaque, and staff stop fielding the same clarifying questions on repeat. Multilingual support compounds the benefit further, expanding access for residents who would otherwise have to navigate complex regulatory language in a second language or wait for a bilingual staff member to become available. A tool that can walk a prospective business owner through zoning requirements at 9 PM on a Thursday is not a luxury feature. For someone trying to open a location on a deadline, it is the difference between making progress and losing a week.

Internal Productivity: The Win Nobody Writes Press Releases About

Less visible but meaningfully impactful is what AI is doing inside city hall, away from public view. Meeting minute summaries, process documentation, routine memo drafting: these are tasks that follow predictable patterns and consume hours of skilled staff time every week, producing outputs that are structurally identical from one cycle to the next. A department head who spends three hours every week formatting meeting notes is a department head with three fewer hours for actual department management. Multiply that pattern across finance, public works and licensing, and you have a city where a significant share of its skilled workforce is doing work that does not require skill.

The efficiency gain here is not glamorous enough for a press release, which is probably why it gets less attention than chatbots and predictive infrastructure. But recovered administrative time, consistently redirected toward higher-judgment work, is how organizations actually get more capable. McKinsey's analysis of generative AI in government specifically identifies content summarization and administrative task automation as the areas where productivity gains are most immediate, because the inputs and outputs are measurable in ways that complex policy work is not. The unglamorous wins are still wins.

Planning and Making Sense of Community Input

AI is also beginning to change how cities approach planning decisions and the processing of public feedback. Modeling the potential impacts of a zoning change before implementation, running scenarios against traffic or infrastructure data, identifying patterns in resident complaints across neighborhoods: these are tasks that previously required either expensive consultants or staff time that most planning departments simply did not have. McKinsey's framework for generative AI in government includes drafting initial urban planning layouts and designing optimal public transport routes as concrete near-term applications. Los Angeles has built this into its formal AI Roadmap, with workforce training as a central component rather than an afterthought, as GovTech's 2024 Digital Cities Survey documents. San José, sitting at the center of Silicon Valley, is recognized in the same survey for its leadership in AI in government more broadly.

The community input angle deserves particular attention. The traditional public engagement model, a noticed meeting, a comment period, a vote, was designed around the constraints of pre-digital communication. It systematically underrepresents residents who cannot attend evening meetings, who do not speak English as a first language, or who simply do not trust that showing up will change anything. AI tools that can process written submissions across multiple channels and surface recurring concerns from large volumes of input do not replace deliberative democracy. They give it better raw material. The ICMA survey found that 38% of local government respondents already see significant potential for AI in policy applications like budget modeling and community analysis, which suggests the appetite is there even where the deployments have not yet followed. For a mayor who genuinely wants to know what constituents think, rather than what the twelve people who always show up to council meetings think, that is worth taking seriously.

"The traditional public engagement model systematically underrepresents residents who cannot attend evening meetings, who do not speak English as a first language, or who simply do not trust that showing up will change anything. AI tools that can process written submissions and surface recurring concerns do not replace deliberative democracy. They give it better raw material."

The Numbers: What the Evidence Really Shows

The most-cited figure in the government AI space is a 2023 McKinsey estimate that generative AI could generate up to roughly $480 billion in annual productivity impact across the public sector, primarily through administrative summarization and citizen engagement tasks. That number has traveled far and wide through conference decks and vendor proposals. What it actually is: a consultancy projection of potential, not a measurement of realized gains, published three years ago when generative AI was still in its breakout year. Worth knowing the number exists. Worth being equally clear about what it is not.

The honest picture is that rigorous, independent data on AI outcomes in local government is still thin. Most figures circulating in this space originate with vendors or advocacy organizations that have a stake in the answer. That does not make every claim false, but it does mean the numbers should be read as directional rather than definitive. A city that claims dramatic cost reductions after deploying a chatbot may be measuring call deflection volume against a baseline that was never formally established. The math is not necessarily wrong; it is just not the kind of math that survives a skeptical audit. Any mayor who has sat through a vendor presentation knows the feeling of watching a slide full of percentages and wondering quietly where they came from.

"A city that claims dramatic cost reductions after deploying a chatbot may be measuring call deflection volume against a baseline that was never formally established. The math is not necessarily wrong. It is just not the kind of math that survives a skeptical audit."

What the Evidence Does Support

Strip away the aspirational projections and what remains is narrower but still meaningful. The OECD's 2025 cross-country review of public-sector AI adoption found that 57% of documented government AI use cases focus on automating or streamlining services, with a further 45% involving decision support and sense-making. Those are not projections. They are counts of what governments have actually built and documented. The pattern that emerges is consistent: AI reduces the manual hours consumed by repetitive work, and that reduction compounds when applied across multiple departments simultaneously. Whether it translates into headcount reduction or faster service delivery depends entirely on what the organization does with the recovered capacity, which is a management question, not a technology question.

The trust dimension tends to get buried under the productivity numbers, and it should not be. A 2024 Route Fifty survey of public-sector technology leaders found that trust and security ranked as the top concerns around generative AI adoption, above cost and above implementation complexity. That finding matters for how cities should interpret vendor efficiency claims. An AI tool that processes resident data in ways that create security exposure, or that produces outputs staff do not trust enough to act on, does not deliver its projected savings regardless of what the sales deck promised. The productivity story is real, but it is conditional on a level of organizational trust in the tool that most cities have not yet established.

The Governance Gap in the Numbers

Perhaps the most revealing data point in the current landscape has nothing to do with productivity. The ICMA's 2024 survey of local governments found that 77% of respondents cited lack of AI awareness and understanding as the most significant barrier to adoption, and 70% identified AI-generated disinformation as their top concern. Only 9% of local governments have an organization-wide policy governing AI use. Read those figures together and you get a picture of an institutional environment where most of the people responsible for deploying AI do not yet feel equipped to evaluate it, and most of the organizations considering it have not written down what they are and are not willing to do with it. The productivity potential is real. The governance infrastructure to capture it safely, in most municipalities, is not yet there.

The pilot-to-scale transition is where most efficiency gains either materialize or evaporate, and the evidence on this is consistent across multiple analyses. Agencies that define success metrics before launch, measure against a documented baseline, and actively manage the workflow changes that AI requires tend to see returns that compound over subsequent deployments. Agencies that deploy a tool and assume the savings will appear on their own tend to generate a line item on a budget report and not much else. The National League of Cities' AI in Action initiative was established specifically to address this gap, helping local leaders build the operational and governance capacity to move from a one-off pilot to something that actually changes how a department works. The technology itself is, at this point, the more tractable part of the problem.

Resident-Facing Tools, Internal Wins, and the Planning Angle

The section heading is a list, but the reality is a spectrum. At one end you have the tools residents actually see and interact with. At the other end you have the infrastructure changes happening inside city hall that residents will never notice directly but will eventually feel in the form of faster service and fewer inexplicable delays. Both matter, and most cities are currently only working on one of them.

What Residents Actually Experience

Atlanta's AI chatbot deployment, recognized in GovTech's 2024 Digital Cities Survey, reduced wait times for residents interacting with city government and improved the overall digital experience. That outcome sounds modest until you consider what it replaced: phone queues, inconsistent answers depending on which staff member picked up, and information buried on a website that no one had updated since the previous administration. A chatbot that gives the same correct answer at 11 PM on a Sunday is not replacing a human relationship. It is replacing a frustrating non-experience that was eroding resident confidence one unanswered question at a time.

The multilingual dimension is underappreciated in most coverage of these tools. A resident navigating permit requirements in a second language, under time pressure, faces a meaningfully different experience than a native English speaker with the same question. AI tools with multilingual capability do not just improve convenience; they change who can actually access government services without needing an intermediary. The National League of Cities' AI in Action initiative frames equitable access as a central design criterion for local government AI, not an optional feature to be added later. Cities that treat it as an afterthought tend to deploy tools that work well for the residents who were already best served and add little for everyone else.

"A chatbot that gives the same correct answer at 11 PM on a Sunday is not replacing a human relationship. It is replacing a frustrating non-experience that was eroding resident confidence one unanswered question at a time."

The Internal Wins Nobody Talks About

Away from the resident-facing layer, AI is doing something quieter and arguably more consequential: giving staff time back. Meeting summaries, report formatting, routine correspondence, eligibility pre-screening: these tasks follow predictable patterns and consume skilled hours every week. The OECD's "Governing with Artificial Intelligence" report is direct about this: AI can free public servants from routine administrative work so they can focus on duties that require human judgment. That is not a vision statement. In departments facing chronic staffing shortages and high turnover, it is a description of how organizations stop drowning.

The compounding effect is what makes this worth taking seriously. A planning department where two staff members spend Monday mornings processing and formatting the previous week's submissions is a planning department with two fewer people available for actual planning on Monday mornings. Recovering that time does not automatically produce better outcomes; it creates the precondition for better outcomes, which is a meaningful distinction when setting expectations for what AI will and will not do on its own.

Forecasting With the Data Cities Already Have

Local governments are sitting on decades of records that most of them have never had the analytical capacity to use properly. Infrastructure maintenance histories, complaint patterns by neighborhood, permit approval timelines, seasonal service demand: this is exactly the kind of data that machine learning tools are designed to find signal in. The OECD report identifies predictive maintenance and anomaly detection as among the more mature government AI use cases, specifically because the feedback loop is clear: a water main that fails at 2 AM is a news story and an emergency budget line, while the same main replaced on a planned schedule because a model flagged its deterioration months earlier is neither. The outcomes are measurable, the data already exists, and the cost of being wrong is immediately visible.

On the planning and policy side, McKinsey's framework for generative AI in government points to scenario modeling for urban design and transport route optimization as near-term applications that expand what planning departments can realistically evaluate before committing to a decision. Los Angeles has built this into its formal AI Roadmap, with workforce training as a central component rather than an afterthought. The cities generating useful results from these tools share a common approach: they started with a specific analytical question they already needed to answer, rather than deploying a tool and hoping it would surface something interesting. That distinction separates a useful planning capability from an expensive experiment that produces a report no one reads.

What Can Go Wrong; and Already Has

Seventy percent of local government respondents in the ICMA's 2024 survey identified AI-generated disinformation as their top concern around adoption. That is not a theoretical worry from people who have never used the technology. It is the dominant anxiety of the people who would actually be responsible for deploying it. They are not wrong to be worried, and the failure modes are specific enough to be worth examining one at a time rather than gesturing at "AI risk" as a general concept.

The Confidence Problem: Wrong Answers That Look Right

Generative AI tools can produce incorrect information with complete conviction. In a consumer context, a hallucinated restaurant recommendation costs you a mediocre dinner. In a government context, an incorrect answer about zoning rules or benefit eligibility can have real financial and legal consequences for the resident who relied on it, and real liability exposure for the city that provided it without adequate human review. The problem is not simply that AI gets things wrong. It is that it gets things wrong in a way that does not look wrong, which means errors travel further before anyone catches them. A chatbot that confidently cites the wrong fee schedule for a business license is not obviously malfunctioning. It just looks like a city employee who gave bad information, except there is no employee to correct and no record of what was said.

This is why implementation guidance consistently points toward starting AI deployments on low-stakes, high-volume tasks where errors surface quickly and can be corrected without serious consequence. The OECD's "Governing with Artificial Intelligence" report is explicit that the absence of clear governance frameworks is one of the primary reasons government AI projects create unintended harm, and the confidence problem is a direct consequence of deploying tools without establishing how errors will be identified and corrected. Deploying AI on enforcement actions or individual eligibility determinations before you have built that feedback loop is how cities end up in the news for reasons that have nothing to do with innovation.

Bias That Hides in the Data

AI systems learn from historical data, and historical data in local government often reflects historical inequities. An algorithm trained on decades of complaint response times will learn, accurately, that certain neighborhoods received slower service. Whether it then perpetuates that pattern or corrects for it depends entirely on how the system was designed and what the training data was prepared to remove. Most off-the-shelf tools have not been audited for the specific historical patterns embedded in your city's records, because your city's records are yours and the vendor has never seen them.

"An algorithm trained on decades of complaint response times will learn, accurately, that certain neighborhoods received slower service. Whether it then perpetuates that pattern depends entirely on how the system was designed; and most off-the-shelf tools have not been audited for the specific historical patterns in your city's data."

The housing and property context offers the clearest documented examples of this failure mode in government-adjacent decisions. Research published in the California Law Review on algorithmic systems in property contexts found that automated tools can reproduce discriminatory patterns from historical data even when race is not an explicit variable, because proxies for race are embedded throughout the underlying data. A city that deploys AI for service prioritization or code enforcement targeting without auditing for these patterns is not taking a neutral position. It is automating its own history and calling it efficiency. The OECD's governance framework addresses this directly, setting out accountability and human-centred values as foundational requirements for public-sector AI, not optional enhancements to be added after deployment if budget allows.

Data Exposure and the Governance Gap

Local governments handle resident addresses, permit histories and complaint records, and sometimes considerably more sensitive personal information about people in vulnerable situations. When vendor tools process or store that data on external infrastructure, cities need explicit contractual clarity on where it goes, who can access it, how long it is retained, and what happens when the contract ends or the vendor is acquired. Vague assurances in a sales conversation are not a data governance policy, and the AI vendor market is consolidating fast enough that a tool procured from a small govtech company may end up operated by a much larger organization with different data practices within the life of the original contract.

The organizational risk is as real as the technical one, and it gets less attention. Staff who do not understand what an AI tool is doing cannot provide effective oversight of it. A chatbot that gives inconsistent answers because no one has maintained its knowledge base since deployment is worse than no chatbot at all, because it erodes the resident trust the tool was supposed to build. The ICMA survey found that 56% of local government respondents cited public perception and trust as a significant concern around AI use, which is the correct instinct. Research published in PNAS Nexus on public attitudes toward algorithmic versus human decision-making found that people apply different standards to AI errors than to equivalent human errors: an AI mistake in a government context tends to generate more distrust than an identical human mistake. Residents will tolerate a slower process more readily than a fast one they do not trust. That asymmetry has real implications for cities that treat transparency as a compliance checkbox rather than a genuine operational priority.

Governance That Works Before Something Breaks

New York City passed Local Law 49 in 2023, requiring city agencies to publish annual reports on their use of automated decision systems. A subsequent Consumer Reports investigation found the law was falling short of its goals: agencies were filing incomplete disclosures, and the public accountability the law intended had not materialized in any meaningful way. That outcome is instructive not because NYC failed, but because it tried something specific and measurable and then discovered the gap between policy intent and operational reality. Writing a transparency law is the beginning of governance, not the end of it. Most cities have not even gotten that far.

The governance gap at the local level is not subtle. The ICMA's 2024 survey found that only 10% of local governments have delegated personnel to oversee AI efforts, and just 9% have an organization-wide policy governing AI use. Those figures mean that in the vast majority of municipalities, AI is being adopted on an ad hoc basis, with individual departments making tool decisions independently and no central function tracking what is being deployed or with what accountability structure. That is not a governance model. It is an absence of one, and it tends to produce exactly the kind of incident that makes elected officials wish they had paid attention earlier.

Define the Decision Before You Deploy the Tool

The most useful governance question is not "what AI tools should we use?" It is "which decisions are appropriate for automation, and which require a person who can take responsibility for the outcome?" Benefits determinations and enforcement actions should require human review, as should anything that materially affects an individual resident's access to services. The reason is straightforward: accountability in government has to attach to a person. Residents have a reasonable expectation of being able to understand and contest decisions that affect them, and an algorithm cannot sit across a desk and explain its reasoning to someone whose application was denied. The OECD's Recommendation on Artificial Intelligence, updated in 2023, sets out accountability and human-centred values as foundational requirements for trustworthy AI in government, not optional features for well-resourced jurisdictions.

Seattle's approach is worth examining in detail because it is replicable at smaller scale. Seattle's Responsible AI program requires that any AI tool used in resident-facing or decision-support contexts be documented with a clear description of its purpose and the human review process that sits above it. The city established explicit policies on data use and vendor access before expanding deployment, rather than retrofitting governance onto tools already in production. Smaller cities without Seattle's resources can still apply the same principle: before deploying any AI tool, identify the specific person whose job it is to review its outputs and correct its errors. If you cannot name that person, the tool is not ready to go live.

"The most useful governance question is not 'what AI tools should we use?' It is 'which decisions are appropriate for automation, and which require a person who can take responsibility for the outcome?' An algorithm cannot sit across a desk and explain its reasoning to someone whose application was denied."

Data Governance Has to Come First

The cities that have avoided the most serious AI missteps share one habit: they defined their data policies before a vendor touched their systems. What data can enter the AI tool? Where is it stored and processed? How long is it retained, and what happens to it if the contract ends or the vendor is acquired? These questions are significantly easier to answer before a deployment than after a data incident. The OECD's 2025 report on governing with AI identifies data governance as a foundational prerequisite for public-sector AI adoption, noting that the absence of clear data frameworks is one of the primary reasons government AI projects underdeliver or create unintended harm. Vague contractual language around data handling is not a minor procurement detail. It is the thing that becomes a headline.

Staff capacity is the other prerequisite that consistently gets treated as an afterthought. An AI tool deployed into an organization where staff do not understand what it does, or why it sometimes produces wrong outputs, will not be effectively overseen regardless of how well the tool itself was designed. This means training staff on the tool's limitations before launch, not after the first error surfaces publicly. It means building a feedback process so that incorrect outputs get reported and corrected rather than quietly ignored or, worse, acted on. The National League of Cities' AI in Action initiative frames this as a leadership responsibility, not a technical one: local leaders need to understand enough about how these tools work to set appropriate expectations for their staff and their residents. A mayor who cannot explain in plain language what their city's AI tools do and do not do is a mayor who cannot credibly defend those tools when something goes wrong. And something, eventually, will go wrong.

How to Buy AI Without Getting Played by the Sales Deck

Every AI vendor proposal aimed at local government currently promises smooth integration and an implementation timeline that somehow fits inside a budget cycle. What the proposals rarely include is a clear description of failure modes, a realistic picture of what your IT staff will need to contribute, or a straight answer about where your residents' data goes after it enters the system. The AI vendor landscape for local government is crowded and the sales pitches are, to put it charitably, optimistic. The antidote is a short list of questions that separate vendors who have thought carefully about government deployment from those who have repackaged a commercial product and pointed it at a municipal budget.

Ask About Failure First

Start with the question vendors least expect: describe, specifically, what happens when the system gives a wrong answer. How does it fail? How often does it fail in comparable deployments? What is the process for identifying and correcting errors once the tool is live? A vendor who cannot answer this clearly has not thought about it carefully enough for a government context. In a consumer setting, a confidently wrong answer is an inconvenience. In a government setting, an incorrect answer about permit requirements or benefit eligibility is a liability, and the resident on the receiving end has every right to expect the city to stand behind what its systems told them.

Auditability is the related question, and it is worth pressing on specifically. For any AI tool that informs or supports decisions affecting residents, you need to be able to reconstruct what the system did and why. If a resident contests an outcome, "the algorithm decided" is not an acceptable explanation in a council chamber or a courtroom. Ask the vendor how outputs are logged, how long those logs are retained, and how a resident or an auditor could access them. If the vendor frames this as a premium feature or an enterprise add-on, that tells you something important about how they think about public-sector accountability. The OECD's Recommendation on Artificial Intelligence treats transparency and accountability as baseline requirements for AI in government contexts, not advanced options. Procurement standards should reflect that.

"Start with the question vendors least expect: describe, specifically, what happens when the system gives a wrong answer. A vendor who cannot answer this clearly has not thought about it carefully enough for a government context."

Data, Contracts, and What Nobody Quotes You

Where does resident data go, and who can access it? This question needs a specific contractual answer, not a reassurance. Get explicit language on data storage location and processing jurisdiction, then nail down exactly what happens to your data if the contract ends or the vendor is acquired, because the second scenario is not hypothetical. The AI vendor market is consolidating rapidly, and a tool you procured from a focused govtech company may end up operated by a much larger organization with different data practices and different incentives within the life of your original agreement. Seattle's Responsible AI program documentation provides a useful template for the kinds of contractual commitments a city should require before any vendor touches resident data. It is publicly available and worth reading before your next procurement conversation.

The cost question deserves the same rigor, because the figure in the initial proposal is rarely the total cost of ownership. Software licensing is the visible part. What often goes unquoted includes content management as the tool's knowledge base ages, integration work with legacy systems, and internal staff time your team will need to contribute across the life of the deployment. Ask for a realistic cost projection over a three-year horizon, then ask for references from municipalities of comparable size and IT capacity, not showcase clients with dedicated digital teams and CIOs who came from Google. The vendors who give you straight answers on both questions before you sign are the ones worth continuing to talk to.

One Check That Costs Nothing

Before signing anything, search the vendor's name alongside the words "government," "error," and "audit." Not to disqualify anyone on the basis of a single bad news story, but because how a vendor responds to a documented public problem tells you considerably more about their accountability culture than anything in their sales materials. A vendor who addressed a failure transparently, communicated clearly with affected clients, and demonstrably improved their product is a more credible partner than one with an unblemished record that has simply never been seriously scrutinized. The OECD's governance framework for public-sector AI emphasizes that accountability mechanisms need to be embedded in procurement, not added after deployment. Asking hard questions before you sign is not due diligence theater. It is the job.

The Next Few Years, Realistically

The near-term trajectory for AI in local government is not a question of whether adoption will accelerate. It already has. The more useful question is what form that acceleration takes, and whether the cities that are still in reactive mode can build the institutional capacity to catch up without making expensive mistakes under pressure. The honest answer is: some will, and some will not, and the difference will mostly come down to whether political and managerial leadership treated this as a real operational priority before a vendor showed up with a proposal.

Where the Technology Is Heading

The current generation of government AI tools, primarily chatbots and document summarization, represents the first wave. The next shift is toward what researchers call agentic AI: systems that can take sequences of actions rather than simply generating a response. In practice, this means tools that could move a permit application through multiple review steps, flag anomalies in procurement records, or coordinate maintenance scheduling across departments without requiring a staff member to manage each handoff manually. The OECD's 2025 report on governing with AI identifies this shift toward more autonomous systems as one of the key governance challenges ahead, precisely because an AI that takes actions rather than producing outputs for human review requires a fundamentally more robust accountability framework than a chatbot that answers FAQs.

Predictive applications are also maturing in ways that are directly relevant to local government priorities. Infrastructure maintenance is the clearest example: machine learning systems that analyze sensor data and maintenance histories to flag components likely to fail before they do. For a city managing aging water systems or road infrastructure on a constrained budget, this is a high-value use case because deferred maintenance is both expensive and politically visible. The OECD report points to predictive maintenance and anomaly detection as among the more mature government AI use cases, specifically because the feedback loop is clear and the outcomes are measurable in ways that softer applications are not. A water main replaced on a planned schedule because a model flagged its deterioration six months earlier is neither a news story nor an emergency budget line. It is just competent infrastructure management, which turns out to be exactly what residents want.

"An AI that takes actions rather than producing outputs for human review requires a fundamentally more robust accountability framework than a chatbot that answers FAQs. Most cities have not built that framework for the tools they already have."

The Trust Problem Will Not Solve Itself

Public confidence in AI-assisted government is genuinely uncertain, and mayors should not assume that efficiency gains will automatically translate into resident acceptance. Research published in PNAS Nexus on public attitudes toward algorithmic versus human decision-making found that people apply different standards to AI errors than to human errors: an AI mistake in a government context tends to generate more distrust than an equivalent human mistake, even when the outcomes are identical. That asymmetry has real implications for how cities communicate about their deployments. Announcing AI adoption as a cost-saving measure is a different political conversation than announcing it as a service improvement, and the framing matters more than most communications teams currently appreciate.

Transparency is the practical response, and several cities are already treating it as an operational priority rather than a compliance burden. Seattle's Responsible AI program makes its governance principles publicly available, which serves a dual purpose: it holds the city accountable to its own standards, and it gives residents a basis for understanding how AI is being used in decisions that affect them. For a mayor weighing how to introduce AI to a skeptical constituency, publishing a clear plain-language policy before a high-profile deployment is considerably better than explaining your governance framework after something goes wrong. The National League of Cities' AI in Action initiative is building resources specifically to help smaller cities do this without needing a dedicated digital policy team to write it from scratch.

What to Actually Do Next

The most practical move available to any city right now is to pick one measurable operational problem and ask honestly whether an AI tool can move a specific number attached to it. A permit backlog averaging 23 days. A call center where the majority of volume is the same handful of questions on repeat. A maintenance queue running months behind schedule. Deploy against that problem with a documented baseline and measure at 90 days. Publish what you found, including where the tool fell short, because that kind of transparent iteration is how cities build the institutional knowledge to use AI well over time, and it gives residents and council members something concrete to evaluate rather than a vendor promise dressed up as a strategic vision.

The ICMA's 2024 survey found that 77% of local government respondents cite lack of AI awareness and understanding as their biggest barrier. That is a solvable problem. It does not require a large budget or a dedicated AI office. It requires someone in city leadership who is willing to assign a real person to own a specific AI question, give that person a defined problem to solve, and hold them accountable for a result rather than a report. Start with the permit backlog. Or the call center queue. Pick the one that is costing you the most visible political pain right now, and find out whether AI can move it. That is a more useful conversation than debating whether AI is the future of government, because in most departments, the future is already several months late.