Most of what circulates under the banner of artificial intelligence in urban planning is still a procurement slide. A much smaller group of systems has been running in live city networks for over a decade, and since late July 2026 the European rules governing them finally carry firm dates.
What survives scrutiny
- Adaptive signal control has documented field results going back to a 2012 pilot, not a forecast.
- Digital twins pay off on physics simulation, not on the three-dimensional visuals that sell them.
- Regulation (EU) 2026/1744 moved the high-risk compliance date to 2 December 2027, which buys procurement time it would be unwise to waste.
- Governance capacity, not model accuracy, is the constraint most municipalities actually hit.
Adaptive signal control is the oldest working example
The clearest evidence that AI changes how a city moves comes from traffic signals, and it predates the current wave by years. Surtrac, developed at the Robotics Institute at Carnegie Mellon University, was first deployed in 2012 across a nine-intersection network in the East Liberty district of Pittsburgh. Rather than running fixed timing plans, each intersection builds its own schedule from detected vehicle flows and passes that information to its neighbours.
The pilot results reported by Carnegie Mellon are unusually specific for this field: travel time through the network fell by around 25%, waiting at signals dropped by roughly 40%, and projected vehicle emissions fell by about 21%. The technology was subsequently commercialised through Rapid Flow Technologies, acquired by Miovision in 2022, and the city has extended it well beyond the original corridor.
Two caveats belong with those numbers. They describe a specific arterial network with particular flow patterns, so they are not a coefficient to apply to any city. And the gains come from coordination logic, which means the sensing infrastructure has to be in place and maintained before any of it works.
| Application | What is documented | What is still a claim |
|---|---|---|
| Adaptive signal control | Measured travel time and idling reductions on instrumented networks | City-wide congestion elimination |
| Urban digital twin | Wind flow, solar yield and flood modelling at district scale | Real-time control of the physical city |
| Utility demand forecasting | Load prediction feeding network operation | Autonomous grid decisions without human oversight |
| Waste collection routing | Route optimisation against sensor fill data | Structural cuts to municipal waste budgets |
Digital twins earn their keep on physics, not on visuals
An urban digital twin is a continuously updated three-dimensional model of a city fed by survey data, sensors and administrative records. Virtual Singapore, developed under the National Research Foundation with the Singapore Land Authority, remains the reference implementation and shows where the value actually sits.
Its productive uses are simulation tasks that were previously impossible at scale: modelling wind flow between buildings to understand street-level ventilation and heat build-up, estimating rooftop solar potential across the whole housing stock, and testing flood scenarios against terrain. None of that is a rendering exercise. Each requires geometry accurate enough to compute against, which is why the data collection, not the software licence, is the expensive part.
The same logic applies at building portfolio scale, where the value comes from instrumented data rather than visualisation, a pattern we have looked at in the context of AI in real estate planning.
The compliance clock moved in July 2026
Anything a European city deploys on traffic or utility networks sits close to the EU AI Act’s high-risk category. Annex III, point 2 covers AI systems used as safety components in the management and operation of critical digital infrastructure, road traffic, and the supply of water, gas, heating or electricity. Classification is cumulative rather than automatic: the system has to function as a safety component, in one of those listed domains, for an operator designated as a critical entity.
The timetable changed this summer. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It pushes the application date for stand-alone Annex III high-risk systems from 2 August 2026 to 2 December 2027, and to 2 August 2028 for AI embedded in products already covered by harmonised product safety legislation.
For a planning department, that is sixteen additional months to write requirements properly rather than a reason to stop. Documentation, data governance and human oversight are contract clauses that are difficult to retrofit once a system is live.
Governance capacity is the real bottleneck
The gap between what these tools can do and what municipalities can run is administrative. The Global Assessment of Responsible AI in Cities, published by UN-Habitat in 2024 on the basis of a survey of 122 municipalities across five regions, found adoption expanding faster than the governance around it, with half of responding cities in Asia and Latin America reporting no specific AI governance initiative at all.
That points to an unglamorous conclusion. The determining factors are whether a city owns and maintains its own data, whether procurement can specify auditability, and whether anyone in the organisation can challenge a model output. A city without those three things buys a dependency, not a capability.
Questions that come up in committee
Does an AI traffic system need to replace existing signal hardware?
Not usually. Adaptive control runs on top of existing controllers where detection is adequate, and the recurring cost is detection and maintenance rather than the algorithm. Where a city has no reliable vehicle detection, that gap is the first line of the budget.
Is a digital twin worth it for a mid-sized city?
It depends entirely on whether there is a specific question to answer. Twins commissioned to model heat, flooding or solar potential tend to justify themselves. Twins commissioned as a platform, with use cases to be identified later, tend not to.
Who is accountable when a model output turns out to be wrong?
Under the EU framework, obligations fall on both the provider and the deploying authority, and the deployer retains responsibility for human oversight. Contractually, that responsibility needs to be written down before deployment rather than negotiated after an incident.
Do these rules apply outside the European Union?
Not directly, but they apply to systems placed on the EU market regardless of where the supplier sits, and several suppliers are converging on a single compliant product line. Cities outside the EU are likely to be offered the compliance features whether or not they require them.
Start with the network you already run
Before modelling a whole city, the fastest measurable gains usually come from instrumenting energy and utility flows properly.
Published 10 August 2025. Updated 8 August 2026. Sources: Carnegie Mellon University Robotics Institute (Surtrac pilot results); Regulation (EU) 2026/1744, Official Journal of the European Union, 24 July 2026; Regulation (EU) 2024/1689, Annex III; National Research Foundation Singapore and Singapore Land Authority (Virtual Singapore); UN-Habitat, Global Assessment of Responsible AI in Cities, 2024.
