AI & the future of the smart city

An ordinary traffic light runs on fixed timings or reacts to a loop in the road surface. A smart traffic light looks further. Cameras and sensors count cars, cyclists and pedestrians, and AI works out how long each direction stays green at every junction. Junctions coordinate with each other, so traffic does not grind to a halt at every light. An approaching bus, an ambulance or a group of cyclists in the rain can get green sooner. In a trial in a medium-sized city, travel times across the participating junctions fell by about a quarter. Next comes a network of lights that exchanges data with buses and connected cars, and predicts a jam before it forms.

A digital twin is a virtual model of the city, fed with up-to-date data: buildings, roads, trees, pipes, traffic and weather. Whatever you change in the model can be tested before it happens in the real city. A planner places a new residential tower on a plot and immediately sees the shadow it casts on the square, how the wind moves down the street and where it will get hot in summer. Within the city's requirements, AI drafts dozens of options for a neighbourhood, with more greenery, less heat or more homes. The next step is a digital twin that runs live alongside the city's sensors and explores how the city could develop under different assumptions.

Sensors on lampposts, façades and bridges measure particulate matter, nitrogen dioxide, noise, temperature or the number of people walking past. Together, thousands of these readings show changes street by street and hour by hour, including between larger measurement campaigns. Increasingly, the AI sits inside the device itself. A smart camera can count pedestrians and pass on only the number, without storing any images. The city sees where it is busy without knowing who walked there. The move from measuring to predicting has already happened: AI combines the readings with the weather forecast, traffic data and the events calendar, and predicts which streets will be too hot, too polluted or too crowded tomorrow.

A camera on a refuse truck or delivery van that already drives through the city collects images of roads, pavements, lighting and street furniture between inspection rounds. Image recognition picks out cracks in the asphalt, loose paving stones, faded road markings and damaged lampposts, and plots them on a map with their location. For pipes, AI uses other data: age, material, soil and pressure readings. From these it estimates where a leak is most likely, so the operator can plan maintenance before a pipe fails. The next step closes the loop from spotting to fixing: the images will become the basis for a repair order to a crew or a repair robot.

An AI agent is AI that takes the steps needed to complete a task by itself. When someone reports a fallen tree, an overflowing bin or broken play equipment, it can ask follow-up questions where details are missing, log the report with a photo and location, and pass it to the crew responsible. With access to the city's systems, it keeps the person who reported it up to date. It will not stop there. Image recognition can already flag damage itself and start a repair order through a linked maintenance system. The agent will connect reports, detection and crew planning into one chain. A dangerous situation or an angry neighbour remains a job for a member of staff.

With the right data, crowds can be predicted well. Counts from sensors and cameras, ticket sales, the weather, public transport and past events together form a pattern. AI learns from it how full a square, a shopping street or a station area will be in an hour or a day. Whoever manages visitor flows can then see in advance where space will run short and adjust: reroute people, open an extra entrance or direct visitors to a quieter spot through an app. Anonymous counts are enough for this. Linked to a traffic model, AI also tests choices for the city: how a new events venue or a closed bridge would shift visitor flows in the city centre.

The idea of the 15-minute city is that you can reach what you need every day within a quarter of an hour on foot or by bike: school, GP, supermarket, park, and work or a place to work. That calls for a different way of planning, with the neighbourhood as the starting point. AI helps show where things are missing. It combines data on amenities, population, routes and travel behaviour into a map of which neighbourhoods fall short, and where. Then it calculates where a new amenity, a cycle link or a hub with shared bikes and charging points would help many people. An urban planning ideal becomes something you can measure, neighbourhood by neighbourhood.

In an AI-first city, all repetitive and predictable tasks go to AI and AI agents. They work with data from records, applications and sensors. Much of this happens behind the scenes: processing invoices, sorting applications for permits and events, planning maintenance of roads and pumping stations, and predicting demand for public services. Staff work for residents and neighbourhoods, weigh things up and use their judgement, for example on a neighbourhood's concerns, a dispute between residents or the choice between a parking space and a tree. Many municipalities still use AI for separate tasks. How much time is freed up for residents depends on how the municipality redistributes the work.

Data analysts and privacy advisers already work for cities. In a scenario for 2040, roles that have no name yet could take on a larger share of the city's work. The director of AI agents mediates when the city's agents pursue conflicting goals: the agent for traffic flow wants a junction to stay green longer, while the agent for air quality wants to divert the traffic. The scenario designer builds the futures in the digital twin that the city's leaders choose between, from a heatwave to a new district. The keeper of the live digital twin makes sure traffic, greenery, water and sewers in the model show the same city as the street. City staff shift from carrying out tasks to coordinating the work and weighing up options.

In city management, people and AI increasingly work as one team. The system keeps an eye on thousands of reports, readings and camera images at once and sees what is out of the ordinary: a drain that overflows more often than normal, a square that gets too crowded every Friday evening. The neighbourhood manager and the enforcement officer go and look on site, talk to residents and look for the underlying cause. AI that predicts flood risk street by street is already being tested. The next step is a team in which AI pulls together signals from the whole city and calculates ahead where the pressure points will be. People then gain more time for residents and decisions.

Anyone who leads a city or a city service today is also leading towards the city of the future. That takes initiative: starting with AI and new technology yourself, one street or one service at a time, even before everything is settled. It also takes a picture of the city in twenty years that speaks to staff and residents, so they want to contribute. Leaders give staff and residents room to put forward ideas and help build, and make visible what happens with those contributions. That is how those who govern, staff and residents build tomorrow's city together.

An organisation with its own future academy trains its people alongside regular education, because the knowledge is needed before new graduates can supply it. Current staff learn to work with AI, data, AI agents and other new technology. New colleagues start with the applications the organisation uses. Leaders practise decisions about dividing the work and overseeing AI. The academy changes as the technology does.

An organisation does not get smart on its own. Start-ups build new services, scientists develop models, partners bring complementary knowledge and governments offer room to test. Together they form an ecosystem, and how well it works determines how fast an organisation renews itself. Organisations all over the world are working on the same questions, and some are already exploring how to link their models and digital twins. An approach that works elsewhere can then be modelled first, before it is trialled locally.