How cities are scaling AI into operations

We spoke to three members of the City Innovation Network—Boston in North America and Prague and Sunderland in Europe—about how they are deploying AI to implement city-wide impact

The deployment of artificial intelligence is forcing a fundamental rethink of how public services are designed, delivered and improved, particularly as pressure grows to demonstrate tangible outcomes rather than isolated innovation.

Across Europe and North America, local governments have spent several years testing AI through pilots, ranging from chatbots and optimisation tools to predictive analytics and automation. While such projects have demonstrated value in controlled environments, most cities are a long way from bringing AI into day-to-day operations. The opportunity to implement data-driven decision-making, optimised workflows and outcomes at scale is yet to be realised.

Silicon Highway, an AI delivery partner working with cities to take projects from proof of concept to full deployment, says that just 7 percent of pilots reach enterprise scale, with many failing to move beyond early validation stages. That gap is not driven by a lack of technical capability, but by the difficulty of embedding those capabilities into the operational and organisational structures that define how cities function.

“Cities often stall at the point where insight needs to become action,” says Nadeem Ullah, Programme Management Director, Silicon Highway. “The technology works, but the operational model hasn’t changed. Moving from pilots to scale means redesigning workflows, not just automating them. You can prove a concept technically, but unless you change how a service operates, you don’t create real value.”

That distinction is becoming central to how cities approach AI. It is no longer enough to demonstrate that a tool can produce better analysis. The real test is whether it can change how decisions are made, how services are delivered and how outcomes are measured, which requires a level of integration that many pilots are not designed to achieve.

Boston: making data accessible

In the US city of Boston, the transition from pilot to scale begins with a focus on usability, and specifically on removing the barriers that prevent people from working with data in the first place.

“We’re about to test the scaling of a couple of our Gen AI products that we’ve developed to democratise our open data portal,” says Shin-pei Tsay, Chief Research and Data Officer, City of Boston, pointing to a moment where experimentation is beginning to shift towards broader deployment.

The work builds directly on the limitations of earlier open data approaches, where datasets were made available but remained difficult to interpret without specialist skills. Boston’s approach changes that dynamic by allowing users to query data in natural language, removing the need for coding or advanced analytics expertise.

“In the open data movement, you still need a certain kind of sophisticated skill set to make sense of the data. What we’ve done is create an MCP server [Model Context Protocol] for the open data portal, which allows you to ask regular questions without any programming requirements, and it looks through all the different data sets. You don’t need to know where to find the data or how to query it, and that changes the way people can use it,” explains Tsay.

The shift has had immediate implications for how decisions are made within municipal teams. By enabling direct access to insights, it reduces reliance on analysts and allows teams to focus on responding to issues rather than interpreting data.

“When we first show the prototype, it solves a lot of problems immediately,” she says. “You don’t need an analyst to tell you what the priorities are or build a dashboard. The conversation can be more about the delivery of the service based on the information, rather than whether the information itself is correct.”

The practical impact becomes clearer when applied to operational datasets such as 311 service requests, where multiple departments interact and where understanding context is critical. 311 is a telephone number that residents in US cities use to report non-emergency requests, for example, to fix potholes.

“With the MCP, it looks at all the data sets very easily, and you get these other facets of answers. We ask it a question about rodents and it brings in data from public housing and public health. That doesn’t even occur to an analyst. Suddenly you have a more rounded answer because it’s connecting across systems,” Tsay says.

While the interface appears simple, the work required to make these systems scalable is far more complex, particularly in areas that receive less attention.

“The things that prevent scale from happening are often the most unglamorous and uninteresting,” she explains. “A lot of what we’ve been doing is reducing the processing load and making the system more platform agnostic, so that it can actually support more users without breaking. You can’t democratise access if ten people can overload it.”

At the same time, scaling depends on how tools are integrated into existing workflows, which often requires changes to operational processes.

“We don’t want to make these things in a vacuum,” she says. “We want them to be usable and useful, meaning they actually improve public service. That connects directly to operational work like our 311 modernisation, where teams are changing how they prioritise and deploy resources. There’s a whole other part of it which is about operations, and that’s still a very big phase of work.”

This highlights a broader pattern. The success of AI initiatives is shaped less by the sophistication of the technology and more by the extent to which it is aligned with how services are delivered in practice.

Boston

• Natural language interface for open data allows non-technical users to query datasets directly
• 311 data analysis connects insights across departments, improving prioritisation and response
• AI tools support faster decision-making by reducing reliance on specialist analysts.

Prague: embedding AI into real-world operations

In Prague, the path to scaling AI is grounded in a clear understanding that data is only valuable if it can be used effectively, and that use must be tied directly to operations.

“AI agents are only as good as the data they rely on and AI solutions fail without well-structured, accessible data. It is data and documentation together that create value,” says Petr Suška, former city CIO and Board Member of Operátor ICT, which is responsible for digital services for the City of Prague.

Rather than focusing on individual pilots, the city has prioritised building a data foundation that supports multiple use cases, ensuring that systems can be integrated into workflows from the outset. This includes improving how data is structured, documented and accessed, making it easier for teams to apply it in their daily work.

“Start small, learn fast, scale what works,” he says. “You need to test use cases with real users and then expand the ones that actually deliver value. Not everything should scale, and that’s part of the process.”

This approach has led to a set of use cases that have moved beyond pilot stage into operational deployment, with one of the most developed being waste collection (see box below).

Instead of relying on fixed routes and schedules, the Czech capital uses AI to analyse operational data and optimise collection patterns, allowing services to adjust dynamically based on demand and conditions. This results in more efficient routing, reduced duplication and a more responsive service overall.

The significance of this lies not just in efficiency gains, but in how the system changes day-to-day operations. Decisions that were previously based on static plans can now respond to real-time or near real-time insights, allowing teams to prioritise resources more effectively.

“Domain specialists use AI more than data consultants, because they understand the problems they are trying to solve and how the tool fits into their work,” he adds.

Prague

• Waste collection optimisation uses operational data to dynamically adjust routes and improve efficiency
• Chatbots and voicebots handle call centre queries, improving response times
• AI supports procurement and vendor processes, including validation and tender drafting.

Sunderland: designing systems that can scale

While Boston and Prague are scaling from pilots, Sunderland’s approach, in the UK, has been to design for scale from the outset, focusing on infrastructure and long-term capability rather than individual use cases.

This strategy began in 2019 when the north-eastern city in England had to face the harsh reality that the private sector was not coming forward to build the connectivity infrastructure necessary for a 21st-century city.

The council believed that ubiquitous connectivity – a network of networks, focusing on both wired and wireless connectivity – was pivotal for the city to realise individual and collective digital ambitions.

“We used the council as an anchor in terms of our own wired connectivity,” says Liz St. Louis, Sunderland’s Director for Smart Cities. “We released a procurement for our own wide area network needs and that led to a significant investment initially by CityFibre to come in and fulfil the council’s requirement but also to invest £62 million to install fibre connectivity across the city.”

LoRaWAN – a low-power, wide area network – covers the whole 153 square kilometres of Sunderland, and the city boasts one of the most advanced 5G private networks in the UK.

Naomi Hutchinson, Chief Innovation and Growth Officer, Sunderland City Council, says the city’s strategy centres on building a platform that can support multiple services and evolve over time.

“We’ve taken a long-term approach to building the platform and connectivity infrastructure so that services can scale across the city,” she explains. “The aim is not to create isolated pilots, but to make sure that when something works, it can be integrated and expanded without having to rebuild everything.”

This “network of networks” approach brings together data from different systems and partners, creating a shared foundation that supports both experimentation and scaling. Rather than developing standalone projects, the city focuses on ensuring that new capabilities can be integrated into this broader ecosystem.

“We need to balance central governance with the ability for teams and partners to experiment,” says Hutchinson. “That’s where the platform approach helps because it provides a shared foundation while still allowing innovation. It means we can support different use cases without creating fragmentation.”

This model is particularly important for mid-sized cities, where resources are more constrained and where the cost of rebuilding systems for each new project can be prohibitive. By investing in shared infrastructure, Sunderland is reducing the friction that often prevents projects from moving beyond pilot stage.

At the same time, the city’s long-term partnership approach supports continuity, allowing initiatives to develop over time rather than being limited to short-term funding cycles, which is often a critical barrier to scaling innovation in the public sector.

“You can’t build a smart city overnight,” observes St. Louis. “There’s a lot of short-termism in this world where it’s very opportunistic to grab a bit of funding and say: ‘We’ll do a little R&D project over here, or we’ll do a trial and testbed over there’, but it’s important to ask how do you scale and how do you make things sustainable.”

Sunderland

• Waste collection optimisation uses operational data to dynamically adjust routes and improve efficiency
• Chatbots and voicebots handle call centre queries, improving response times
• AI supports procurement and vendor processes, including validation and tender drafting.

From pilots to systems

Across Boston, Prague and Sunderland, a consistent pattern is emerging. Scaling does not happen after deployment. It is shaped much earlier, in how projects are designed, how data is structured and how organisations are set up to support change.

“These are the points where many projects lose momentum,” says Silicon Highway’s Ullah. “You need a structured way to build the business case, assess risk and prioritise opportunities. Without that, even successful pilots can fail to move forward.”

AI is no longer treated as a collection of pilots, but as part of a broader system that underpins multiple services and functions, from frontline operations to internal processes and strategic planning.

In Boston, that shift is already beginning to reshape how analytics are delivered and used, particularly as tools move closer to the people responsible for service delivery rather than remaining within specialist teams.

“I do think it’s going to transform the way we think about analytics and how we set up our teams and hire,” Tsay says. “When we saw what it could do, it became clear this isn’t just another tool. It changes how you approach the work, who needs to do it, and how quickly you can move from understanding something to actually acting on it. This is a big change. It’s not just an evolution.”

Prague: waste collection optimisation

Prague has moved waste collection optimisation into operational use, applying AI as part of a wider effort to embed data-driven decision-making into frontline services rather than treating it as a standalone innovation project.

How it works in practice

• Uses internal operational data alongside general AI capabilities to support routing and service planning
• Integrated into existing service workflows rather than operating as a separate system
• Forms part of a broader set of AI-enabled operational tools across the city

What has changed

• Shifts from static, pre-defined routes to more responsive service planning
• Enables teams to adapt to real conditions rather than relying on fixed schedules
• Supports more consistent use of data in day-to-day operational decisions

“Generally there is low tolerance for uncontrolled risk,” Suška says, pointing to the need to ensure these systems operate within defined parameters even as they are embedded into services.

Why it has scaled

The use case is supported by a broader framework that enables experimentation while maintaining control, ensuring that tools can be adopted across teams without creating fragmentation or risk.

Main image: William Zhang on Unsplash