Few technologies have arrived in transportation with as much promise, or as much noise, as artificial intelligence (AI). Vendors, consultants and agencies describe it as a way to spot incidents sooner, flag safety risks before they lead to crashes and stretch maintenance budgets further. But for many state departments of transportation and toll authorities, the harder question is no longer whether AI works in a demonstration. It is whether it can work every day, across every camera and every weather condition, in systems that agencies will trust with real decisions.

Many agencies have run pilots. Some have moved AI into routine operations. The obstacles are often less about algorithms than about the conditions around them: data that doesn't represent the full range of what happens on the road, legacy roadside equipment, staff who need to learn to work alongside new tools and procedures written before this technology existed. Getting past those obstacles is as much an organizational task as a technical one.

Ahmadreza Azizi works on that problem from the vendor side. He is head of AI at Quarterhill, a tolling and intelligent transportation systems company, where he leads development of AI tools for toll back-office operations and roadside systems, including automated license plate recognition and vehicle identification. He previously served as a founding engineer at two AI startups and holds a Ph.D. in physics and a master's degree in computer science from Virginia Tech University.

In this email Q&A, Azizi discusses where AI is delivering results today, why so many projects stall after the pilot stage and what data and hardware agencies need before scaling up, drawing on a phased Quarterhill project with the Illinois Tollway. He also explains where engineers should keep final authority and how a system's own confidence scores can help decide which tasks get automated and which go to a person for review.

R&B: AI is increasingly being discussed as a tool for transportation agencies, but where is it actually delivering meaningful results today? Which types of applications—whether in traffic operations, roadway safety, infrastructure planning or asset management—are showing the most promise? 

Ahmadreza Azizi: AI is already delivering meaningful results across the transportation ecosystem from tolling and traffic management to roadway safety, infrastructure planning, asset management and maintenance. 

In traffic operations, operators may traditionally need to continuously monitor cameras and multiple data sources to identify incidents. AI can automate much of those processes to bring the most relevant information to the operator’s attention and recommend appropriate response plans while leaving approval and oversight with the operator.

AI can also directly impact congestion and roadway safety. Applications such as adaptive signal management can help optimize traffic flow, while AI can process information from roadside infrastructure cameras, sensors and connected vehicles to identify risks and support faster responses. By analyzing vehicle movements and interactions between road users, AI can help agencies detect near misses, sudden braking and other safety risks before they lead to collisions.

Over time, AI can also identify potential infrastructure or roadway design issues to support proactive safety improvements, predictive maintenance and infrastructure planning. 

The opportunity is to connect these different sources of transportation data and turn them into information agencies can act on to improve safety, mobility and operational efficiency.

R&B: AI is only as useful as the data and infrastructure supporting it. What does an agency need to have in place—both digitally and physically—before it can successfully deploy AI at scale? 

AA: Before an agency deploys AI at scale, it needs a reliable data foundation and the operational infrastructure to keep that data flowing consistently.  

On the digital side, data must represent the environment the system will operate in. Data collected on a single corridor, in good weather and during daytime, will not necessarily generalize to night, glare, rain, snow or heavy occlusion. Agencies benefit from capturing representative data across sites, seasons, lighting conditions and traffic patterns before expecting production-level performance. That means thinking about AI as part of a broader transportation technology ecosystem: the roadside equipment, sensors and digital platforms collecting the data are just as important as the model analyzing it.

Data quality has to be measured, not assumed. For example, within Quarterhill's ALPR pipeline, images are quality-gated before a model runs. An image that is too dark or lacks usable contrast is rejected rather than being used to produce an unreliable answer. 

Labeled ground truth is the other half of that foundation. Agencies need a way to establish the correct answer, so performance can be measured objectively and re-measured over time. Without ground truth, an agency can report that a system is running, but not that it is right. 

On the physical side, agencies need reliable capture hardware such as cameras, LiDAR and loop sensors; power and communications capacity to move that data; and compute capacity sized to the workload. 

R&B: As transportation agencies increasingly use AI to make decisions about traffic, safety and infrastructure, how should they balance the benefits of automation with the need for human oversight? Where should engineers and transportation professionals remain firmly in the loop?

AA: AI should primarily be viewed as a tool that empowers transportation professionals, rather than a replacement for them. Its strength is its ability to continuously process very large amounts of information, recognize patterns and anomalies, prioritize events and recommend actions much faster than would be practical through manual monitoring alone. This is the approach we see as most valuable: using AI to help agencies make faster, better-informed decisions, while keeping people in control where judgment and accountability matter most.

Routine, well-understood processes may be suitable for greater automation once sufficiently validated. However, where an action can have a meaningful impact on public safety or enforcement, engineers and transportation professionals should remain firmly in the loop and retain final decision-making authority.

A well-designed AI system should not just produce an answer; it should indicate how certain it is. Confidence scores can help determine what happens next: results with high confidence can be handled automatically, while lower-confidence results can be flagged for human review.

The goal should be human-AI collaboration: AI provides speed, scale, continuous monitoring and decision support, while people provide engineering judgment, accountability and oversight. Automation can increase as agencies gain experience, but that progression should be based on validated performance and the level of risk involved.