In the current discussion around Agentic AI, the focus is often on the capabilities of models. New systems can plan, call tools, orchestrate workflows, and increasingly take on complex tasks. However, once this discussion leaves the world of demos and meets real operational questions, the perspective shifts. Then the question is no longer whether a KI agent can perform a task. Much more interesting is whether it understands the environment in which a decision must be made. This exact question was at the starting point of an Urban Digital Twin scenario for Frankfurt am Main. The goal was to investigate how spatial analyses, simulations, and KI can be combined to support decision-making processes in a complex urban environment. However, during the work, a much more fundamental insight emerged: Before a KI agent can support meaningful decisions, it must understand the spatial contexts that shape a city. Cities are not a collection of independent systems. Traffic, infrastructure, weather, energy consumption, public transport, construction sites, and human behavior constantly influence each other. Anyone who wants to understand these interactions needs more than data. They need context. And this context is spatial in many cases.
Every City Has a Memory
When Digital Twins are mentioned, many first think of real-time data. Traffic flows, sensors, construction sites, weather information, or current events often determine the perception of what a digital twin should achieve. However, while working on the Frankfurt scenario, it quickly became clear that the present only tells part of the story. Cities also have a memory. Every traffic accident leaves traces in the data. Recurring traffic jams create patterns. Certain intersections or street sections show similar anomalies over years. When these patterns are made visible, an understanding gradually emerges of how risks develop in urban space. For this reason, the analysis did not start with KI but with historical traffic accident data.
The aim was not to examine individual accidents or produce another heatmap. The interesting question was whether stable spatial patterns could be identified over several years that provide clues about how the traffic system functions. Some results confirmed expectations: high concentrations of accidents occurred along major traffic axes, at heavily used intersections, and around central traffic hubs. Other results were surprising: certain areas of the city repeatedly appeared in analyses even though they did not seem to play an outstanding role in the traffic network at first glance.
Only by including additional geodata—such as settlement structure, traffic infrastructure, or public space usage—did a more coherent picture emerge. The real insight was not in individual hotspots but in the persistence of certain patterns. Some areas of the city developed similar risk structures over years while others remained surprisingly stable.
Looking Back Is Not Enough
Historical data helps to understand the past. However, decisions are made for the future. For this reason, the analysis was extended with traffic simulations. While accident analyses answer what happened in past years, simulations allow a look at possible future developments.
How do traffic flows change during major construction sites?
What effects do detours have?
How do certain areas of the city respond to increasing traffic loads?
Which corridors come under pressure early?
Simulations do not provide predictions in the classical sense. Rather, they enable playing through scenarios and make visible how loads can spread through a traffic network. Initially, both analytical worlds existed side by side: historical accident patterns on one side and simulated traffic flows on the other. The real added value only arose when both perspectives were combined.
When Future Pressure Meets Historical Risk
High traffic load is important information. An accident hotspot is also important information. But places where both factors come together become particularly interesting. In the Frankfurt scenario, areas appeared where traffic simulations predicted increasing loads and at the same time historical accident analyses documented elevated risk values over years.
Thus, the analysis no longer focused exclusively on traffic or traffic safety. Suddenly at the center was the question of risk. Here one of the most important insights of the project emerged: The greatest added value was not identifying individual hotspots but understanding how different risk signals interact with each other. Cities rarely come under pressure from a single event. Critical situations often arise when multiple factors occur simultaneously and reinforce each other. From an analytical perspective, this means that spatial decision support has less to do with evaluating individual datasets than with understanding their interactions.
This Is Where KI Became Really Interesting
Only after this analytical foundation was established did the question of using KI agents arise in a new light. Until then, the challenge was not generating reports or summarizing dashboards. The real task was to relate different information sources to each other and derive reliable conclusions from them. Suppose a control center simultaneously receives several pieces of information: increasing traffic loads, an upcoming weather event with heavy rain, a major event downtown, as well as historical indications of increased accident risks in adjacent areas.
Each piece of information alone may seem unspectacular. Together they form a much more complex situational picture. This is exactly where room arises for new concepts like Spatial Decision Intelligence. Within the Spatial Decision Skills project it was therefore investigated how geographic analysis capabilities can be provided as tools for KI agents. The idea behind this is deliberately simple: Instead of assuming that a model already knows all information, it gains access to spatial analysis functions and can determine relevant information itself. An agent could identify accident hotspots, analyze critical infrastructures, evaluate alternative routes or examine spatial impacts of a decision. The strength of such a system lies not in knowing answers by heart but in recognizing contexts and justifying them comprehensibly.
The Most Important Question Was Not "Can KI Do This?"
During work another question kept coming up.
Not:Can KI perform this analysis?
But: Why does KI recommend exactly this action?
This question proved to be much more important than the actual recommendation itself. Trust does not arise through automation; trust arises through comprehensibility. Decision-makers want to see which data were used; they want to understand underlying assumptions and follow how a system arrived at a particular recommendation. The more analyses moved toward decision support, the clearer it became that transparency cannot be an afterthought; it forms the basis for trustworthy GeoAI systems. Therefore, a KI agent analyzing spatial contexts must do more than calculate a result; it must disclose its reasoning process.
From Digital Twin to Spatial Decision Intelligence
In retrospect, the Frankfurt scenario mainly highlighted one insight: The future of urban systems will not be shaped by autonomous KI agents making decisions independently from humans. A much more realistic future appears where spatial analyses, Digital Twins, Data Science and KI interplay to make complex contexts more understandable. Spatial Data Science helps recognize patterns; Digital Twins enable understanding dynamics and future developments; Agentic AI opens new ways to combine different knowledge sources and use them contextually. Together these developments point toward something bigger—a discipline that does not primarily deal with data or algorithms but with better decisions.
Because in the end the most important question is not whether KI can generate an answer; the crucial question is whether it can help better understand consequences of possible decisions before action is taken—and whenever decisions affect people, infrastructure or cities spatial context almost always plays a central role.
Further Links
Discussion
The work around the Frankfurt Urban-Digital-Twin scenario as well as the Spatial Decision Skills project is still in its early stages. That is precisely why we are interested in different perspectives from the community.
Which decisions should an AI agent be allowed to support in an Urban Digital Twin?
Where are the boundaries between intelligent assistance and autonomous decision-making?
And perhaps the most exciting question:
Which spatial skills should every Geospatial AI Agent master in the future?
%2520Office%2520skaliert.png)