Demystifying AI for Transportation

Abhishek Dubey, PhD (CSO)

Mike Wilbur, PhD (CTO)

Overview

Across the United States, public and private transportation agencies serve billions of passenger miles each year. Their shared goal is to optimize utilization, improve efficiency, and enhance service reliability while dealing with unpredictable realities like fluctuating demand, driver availability, and traffic congestion. Traditionally, both industry and academia have approached these challenges through combinatorial optimization, solving mathematically complex problems such as line planning, vehicle routing, and crew scheduling. While effective in controlled conditions, these methods often struggle to respond quickly and efficiently to the dynamic nature of real-world operations.

The rise of Artificial Intelligence (AI) has introduced a new paradigm. By learning from vast operational data (trip histories, call logs, GPS traces, weather) AI systems can anticipate needs and adapt in real time. Yet AI is not a single technology or a one-size-fits-all solution. In transportation, intelligence arises from the integration of multiple AI capabilities, each playing a specific role. A practical way to understand this ecosystem is through four interconnected layers: Predictive AI, Generative AI, Planning and Optimization AI, and Agentic AI. Together, they form the foundation for intelligent, adaptive, and resilient mobility management. The following sections describe each layer with examples from Non-Emergency Medical Transportation (NEMT), where data-driven dispatch and routing can significantly improve efficiency and service quality.

Predictive AI: Turning data into foresight

At the heart of any intelligent transportation system is the ability to anticipate what is likely to happen. Predictive AI uses statistical models and machine learning to forecast ridership demand, vehicle availability, and operational risks such as delays or breakdowns. Researchers have shown that combining passenger counts with external data like weather and traffic incidents can improve disruption forecasts by more than 40 percent compared to traditional averages. 

In NEMT operations, predictive AI can play an especially powerful role in what’s known as will-call trip prediction. Many NEMT passengers request return rides after medical appointments only when they are ready to leave, creating uncertainty for dispatchers. By learning from historical data about appointment durations, clinic schedules, and patient profiles, predictive models can estimate when passengers are likely to call for a ride. With these forecasts, dispatchers can pre-position vehicles near expected pickup points, reserve capacity dynamically, and reduce patient wait times.

Generative AI: Exploring possible futures

Prediction alone is fragile in an uncertain environment. Sudden changes in weather, traffic, or driver availability can make even the best forecasts obsolete. Generative AI helps address this by simulating thousands of plausible scenarios that represent different “futures” of the transportation network.  These models can be used to build digital twins (virtual replicas of real-world systems) allowing planners to test strategies under varying conditions. In NEMT settings, agencies can simulate how a 10 percent increase in late clinic discharges affects return-trip demand or how a snowstorm might reduce available drivers. By comparing performance across these simulated futures, planners can identify strategies that perform reliably under a range of disruptions.

Beyond simulation, generative AI, especially large language models (LLMs), is improving communication and decision-making across human teams. In dispatch centers, LLMs can serve as intelligent intermediaries between drivers, dispatchers, and algorithms. If a predictive model forecasts a surge in will-call requests later in the day, an AI assistant can automatically draft messages to drivers, confirm availability, and feed those responses back into the optimization system. When a driver reports a delay, the assistant can parse the message, update the simulation, and notify dispatchers, all in real time. This human–AI teaming ensures that information flows seamlessly between people and systems, enabling faster, more coordinated decision-making.

Planning and Optimization AI: Deciding under constraints

Once agencies can forecast and simulate multiple futures, the next challenge is determining what to do. Planning and Optimization AI provides this capability by evaluating possible actions under uncertainty and real-world constraints. Traditionally, such problems have been tackled using optimization algorithms such as Mixed-Integer Linear Programming (MILP) solvers and heuristics, which work well for static scenarios but can struggle with real-time adaptation. Modern, AI-driven planning often uses algorithms like Monte Carlo Tree Search or reinforcement learning to balance multiple objectives including minimizing costs, reducing wait times, and improving service reliability.

In NEMT dispatch, for example, they help determine which driver should serve each passenger, how to adjust schedules when appointments run late, and how to reroute vehicles during traffic disruptions. The most effective methods are neurosymbolic - combining the flexibility of machine learning with the structure of rule-based reasoning. Machine learning models might predict travel times or demand surges, while symbolic constraints ensure vehicles do not exceed capacity, ADA compliance is maintained, and passengers are picked up within appointment windows. This hybrid approach makes AI-driven plans both efficient and compliant, balancing adaptability with accountability.

Agentic AI: Closing the loop

The final layer, Agentic AI, marks a shift from systems that merely recommend actions to those capable of executing them autonomously. These agents can perceive their environment, simulate outcomes, act in real time, and escalate only unusual or high-risk cases to human supervisors. In practice, this means specialized agents manage different parts of the operation. Dispatch agents automatically assign and rebalance trips, stationing agents determine where to position vehicles across service zones, and charging agents coordinate electric vehicle charging schedules to minimize downtime.

In NEMT operations, an agentic dispatch layer could autonomously reassign will-call pickups based on predicted call-in times and live GPS updates. If unexpected conditions arise, such as a vehicle breakdown or a clinic running late, the system can resolve most cases automatically while keeping human operators informed. This approach moves transportation management from dashboard monitoring to AI co-management, where intelligent systems and human teams work together to handle complex, real-time mobility networks.

Building the future of intelligent mobility

For transportation agencies, understanding the layered structure of AI is more than a technical exercise, it is a strategic imperative. Each layer, from prediction to action, contributes to a broader system of intelligence. When evaluating solutions, agencies should look for platforms that clearly articulate how these layers interact: how data informs predictions, how scenarios are simulated, how plans are optimized, and how decisions are executed. The most effective systems are clear about their decision logic, incorporate feedback loops, and are designed to evolve with operational realities rather than replace them. 

Equally important are assurance and accountability. AI-driven systems must demonstrate reliability, and compliance with existing transportation and accessibility regulations. Agencies should seek solutions that not only optimize performance metrics but also provide measurable assurances, clear performance baselines, explainable outputs, and robust monitoring frameworks that detect bias or drift over time.  Lastly, it is important to emphasize that the greatest potential of AI in transportation lies not in full automation but in human–AI teaming. When AI systems handle the complexity of data and decision space, human operators can focus on oversight, empathy, and contextual judgment - areas where human experience adds irreplaceable value. A well-designed partnership between people and intelligent systems can deliver a level of service that is faster, fairer, and more responsive to community needs.

To read more about some of these techniques please consider reading the following papers. Further information is available on our research partner website at smarttransit.ai.

• Talusan, J. P., Han, C., Rogers, D., Mukhopadhyay, A., Laszka, A., Freudberg, D., & Dubey, A. (2025). An End-to-End Solution for Public Transit Stationing and Dispatch Problem. ACM Transactions on Cyber-Physical Systems. https://doi.org/10.1145/3754454

• Wilbur, M., Kadir, S. U., Kim, Y., Pettet, G., Mukhopadhyay, A., Pugliese, P., Samaranayake, S., Laszka, A., & Dubey, A. (2022). An Online Approach to Solve the Dynamic Vehicle Routing Problem with Stochastic Trip Requests for Paratransit Services. Proceedings of the 13th ACM/IEEE International Conference on Cyber-Physical Systems (ICCPS). https://arxiv.org/pdf/2203.15127

• Pavia, S., Rogers, D., Sivagnanam, A., Wilbur, M., Edirimanna, D., Kim, Y., Pugliese, P., Samaranayake, S., Laszka, A., Mukhopadhyay, A., & Dubey, A. (2024). Deploying Mobility-On-Demand for All by Optimizing Paratransit Services. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24). https://doi.org/10.24963/ijcai.2024/822

• Sivagnanam, A., Kadir, S. U., Mukhopadhyay, A., Pugliese, P., Dubey, A., Samaranayake, S., & Laszka, A. (2022). Offline Vehicle Routing Problem with Online Bookings: A Novel Problem Formulation with Applications to Paratransit. arXiv preprint. https://arxiv.org/abs/2204.11992

• Wilbur, M., Sivagnanam, A., Ayman, A., Samaranayeke, S., Dubey, A., & Laszka, A. (2023). Artificial Intelligence for Smart Transportation. To appear in Y. Vorobeychik & A. Mukhopadhyay (Eds.), Artificial Intelligence and Society. ACM Press. https://arxiv.org/abs/2308.07457