Why AI hasn't yet delivered for urban transport

Urban transport systems are being stretched to their limits as cities attempt to manage congestion, electrification, climate targets, and rising passenger demand at the same time. Decisions that once relied on static schedules and long planning cycles are now being made in real time, under uncertainty, with little margin for error. With networks growing more complex, attention is shifting from planning tools to operational control and whether artificial intelligence (AI) can manage trade-offs that humans increasingly struggle to balance.

A new systematic review, Operational AI for Multimodal Urban Transport: A Systematic Literature Review and Deployment Framework for Multi-Objective Control and Electrification, published in the journal Logistics, finds that while AI shows strong potential to improve efficiency, reliability, and sustainability, most applications remain disconnected from real-world operations due to fragmented data, single-objective optimization, and weak governance integration.

AI potential outpaces operational reality in urban mobility

The review maps the rapid growth of AI research across urban transport domains, including traffic signal control, demand-responsive transit, green routing, fleet scheduling, and digital twins. Machine learning, reinforcement learning, and advanced optimization techniques are widely explored as tools to manage congestion, reduce emissions, and improve service regularity. In theory, these approaches allow transport agencies to respond dynamically to disruptions, balance competing objectives, and optimize resources in real time.

In practice, however, the study finds a persistent gap between research outcomes and operational use. Most AI models are designed around narrow objectives, such as minimizing travel time, reducing fuel consumption, or improving punctuality in isolation. Real-world transport agencies, by contrast, must weigh multiple and often conflicting priorities, including cost control, labor constraints, passenger satisfaction, safety, and environmental performance.

Data fragmentation further limits operational deployment. Urban transport systems typically rely on heterogeneous data sources, including vehicle location feeds, ticketing systems, traffic sensors, energy infrastructure, and workforce schedules. Many AI studies assume clean, centralized datasets that do not reflect the messy realities of live operations. As a result, models that perform well in simulation often fail to scale or generalize when exposed to real-time uncertainty and incomplete information.

The review also highlights limited reproducibility and benchmarking across studies. Performance improvements are frequently reported without standardized baselines or transparent evaluation methods, making it difficult for transport authorities to assess whether proposed AI solutions offer reliable gains over existing practices. This lack of comparability undermines confidence and slows institutional adoption.

Electrification exposes new constraints for AI-driven operations

Electrification has emerged as a defining challenge for urban transport operators, particularly as cities commit to replacing diesel bus fleets with electric alternatives. While electric vehicles reduce local emissions, they introduce operational constraints that traditional scheduling and routing systems were not designed to handle.

The study finds that many AI applications treat electrification as a planning variable rather than an operational reality. Charging availability, vehicle state of charge, energy pricing, and charging time are often excluded from real-time decision making or handled through simplified assumptions. This disconnect leads to infeasible schedules, unexpected service disruptions, and underutilized electric assets.

The researchers argue that electrification must be embedded directly into operational AI systems. Electric fleet management requires continuous coordination between vehicle assignments, charging infrastructure, crew schedules, and service requirements. AI models that ignore these interdependencies risk optimizing theoretical performance at the expense of operational robustness.

The review identifies emerging research that integrates energy-aware routing, battery degradation models, and charging optimization into transport operations. However, these approaches remain fragmented and are rarely combined with broader multimodal control strategies. As cities expand electric fleets, the absence of unified operational frameworks becomes a critical barrier to reliable service delivery.

Electrification also amplifies the importance of uncertainty handling. Weather conditions, traffic congestion, and unexpected delays directly affect energy consumption and charging feasibility. AI systems must therefore incorporate probabilistic reasoning and real-time adjustment capabilities, rather than relying on static optimization.

A deployment framework for multi-objective, governable AI

To address these gaps, the study proposes a deployment-oriented operational AI framework designed to support real-world decision making rather than theoretical optimization. The framework is structured around five interoperable layers that reflect the practical requirements of transport agencies.

The first layer focuses on data ingestion and harmonization, enabling diverse data streams to be integrated without forcing full centralization. This approach acknowledges institutional constraints and legacy systems while enabling cross-modal visibility.

The second layer introduces streaming analytics capable of handling uncertainty, disruptions, and incomplete information. Rather than assuming stable conditions, the framework is designed to operate under real-time variability, reflecting the realities of urban transport networks.

Under the hood, the framework is a solver-agnostic optimization layer. Instead of producing a single optimal solution, the system generates a range of feasible alternatives across multiple objectives, including cost, service reliability, and energy or emissions performance. This design allows human decision-makers to evaluate trade-offs explicitly rather than deferring judgment entirely to automated systems.

The fourth layer provides multi-criteria decision evaluation, translating technical outputs into decision-ready insights aligned with policy priorities and operational constraints. This step bridges the gap between algorithmic optimization and institutional accountability.

The final layer focuses on governance, monitoring, and auditability. The authors stress that AI systems must be transparent, explainable, and adaptable to regulatory oversight. Continuous performance monitoring and feedback loops ensure that models evolve alongside operational realities rather than becoming obsolete.

The framework is demonstrated through a multimodal case study in Thessaloniki, Greece, where real-time data streams are used to support service planning and disruption management without increasing operational costs. The case demonstrates how AI can improve regularity and sustainability when embedded within a coherent operational architecture.