UrusHyby via WikimediaPublic transit is emerging as a new frontier in the application of artificial intelligence, a May report shows.
The American Public Transportation Association’s Artificial Intelligence (AI) and Machine Learning (ML) in Public Transit: A Primer surveyed 32 transit agencies about how they’re using AI tools. Agencies are experimenting or considering the technology across a variety of use cases:
- Back office, 50%
- Operations, 47%
- Customer support, 44%
- Maintenance, 41%
- Safety and security, 38%
- Customer analytics, 34%
- Planning, 22%
- Fares and ticketing, 16%
There are plenty of examples. CapMetro, the transit system serving Austin, Texas, has developed an AI system to allow customers to book and cancel trips and check estimated times. This has allowed CapMetro to handle more calls.
New York City’s Metropolitan Transportation Authority equipped 623 buses on 14 routes with cameras that identify vehicles that are violating bus priority lanes, are double-parked, or are parked in bus stops and tickets the vehicles’ owners. According to the report, “The MTA has found that since implementation that general route reliability has increased, bus lane speeds have increased by 5%, and collisions have decreased by 20% on average for relevant routes. In addition, since implementation bus lane violation recidivism is at 9%.”
Further reading:
- ASCE’s AI RACE positions civil engineers to embrace responsible AI integration
- Hacks of US water systems make it clear: Cybersecurity is a civil engineering problem
- Engineers monitor Florida bridge with digital twin
And there are plenty of projects agencies are running that were not mentioned in the report. The Chicago Transit Authority, for instance, created an innovation challenge to explore a specific safety problem: right-of-way intrusions by pedestrians. CTA selected STV, a New York City-based national infrastructure-focused professional services firm, to develop a $60,000 pilot project, which ran from February 2025 to February 2026 to deploy cameras along the platforms of two CTA rail stations (the Ashland Green Line and Cicero Blue Line).
According to STV, the CTA receives about 2,000 reports annually of people or objects entering the rail network’s right of way, creating safety hazards and causing service delays. The cameras, with the help of AI, can differentiate pedestrians from other moving objects and could, said Mark McElroy, director of digital solutions at STV, “quickly determine if a pedestrian is somewhere that they’re not supposed to be. And the purpose of this pilot is to alert the CTA if an intrusion occurs that they need to be notified about.”
McElroy is seeing a targeted focus of safety initiatives among transportation agencies. AI systems can analyze stations or intersections a lot more quickly than humans, “tracking every single car, every single bicycle, every single pedestrian within an intersection over a long period of time,” McElroy said. This helps transit planners decide when they need to repaint traffic lines or crosswalks or change lighting to develop a better environment for the public.
“I think that is a huge use case that transportation clients are very, very interested in,” he said. The APTA reports other safety projects, including one from the New Jersey Transit Corp., which piloted an AI video analytics tool to identify unsafe crowding, track obstructions, and unattended packages.
McElroy added that the Chicago pilot model eventually functioned with 99% accuracy. But there have been a few unusual situations. During the holidays, for example, the CTA runs some trains with holiday lights. “They had a Santa Claus sitting in the middle of the train cars, and Santa is a person, and it flagged that on the train as an oddity,” he said. “It’s funny that you would never think to program an AI that you don’t need to worry about Santa Claus.”
2 kinds of AI systems
Abhishek Dubey, Ph.D., associate professor of computer science at Vanderbilt University, described two kinds of AI technologies.
“One is automation, like fire and forget it,” he said. “You will transcribe my calls, and you can automate certain steps, and you don’t even need to talk to a human.” This routine automation category, he said, is the easier part and the approach most likely to change how routine work is done.
The more ambitious, and maybe more useful, type of AI is what he refers to as “nonmyopic optimization,” which means the ability of AI systems to anticipate and plan around problems. “Imagine if you can see the disruptions that could happen on a road because you know yesterday there was construction work going on there,” Dubey said. “So rather than reacting to the detour, you know the detour is happening. You know that there is a major event happening in the city. … You react to it, and you plan for it.”
For instance, if a bus breaks down and transit operators have sufficient data available, they can identify where other buses in the system are and what prevailing traffic congestion looks like at the time of the disruption and then move the most appropriate bus into service.
Context is key when analyzing an agency’s operations, Dubey said.
Bruce Englehardt via Wikimedia
“The more and more data of your performance we see,” he said, “the better and better we merge it with all the data of the city that is available: infrastructure, construction, events, road closures, detours, traffic congestion. We can predict better and better what are the possible problems that you are going to have.”
Nonmyopic AI systems could be used to predict demand based on new housing developments or the average appointment times of patients seeking routine medical care, such as dialysis. If it’s the third quarter of a football game and the home team is losing big, AI models could forecast that fans will start leaving the stadium early and deploy transit assets to pick them up.
“So what did you gain?” Dubey said. “You lowered the deadhead, you gave a better service, but what do you need? You need to now anticipate the pattern that more demand will come from certain areas at certain times. … So you can improve the efficiency, lower the fuel usage by distributing your resources, which are the drivers and the vehicles across the city.”
According to the APTA report, California’s Riverside Transit Agency is piloting a disruption-management tool “that lets dispatchers rapidly define dynamic detours and then automatically pushes real-time updates (including detour paths) to riders and operators. The system can push real-time updates to riders through multiple channels, reducing confusion and missed trips. Early trials have shown that the platform can shorten response times and help operators maintain schedule reliability.”
Process and data are key
A supplemental APTA guidance brief listed some challenges transit agencies will face when implementing AI, including timely access to data and better data integration; internal policies and governance structures in place and compliance with state and federal policies; and training staff to use and assess AI tools and write procurements for those tools.
Observers agree that data is key.
“The research showed the transit industry is seeing AI is most effective when the task is repetitive, high volume, and has a clear right answer you can check so you can be confident in the AI,” said Seyedmostafa (Sal) Zahedi, Ph.D., senior transportation planner and data scientist with Foursquare ITP. Zahedi was one of the researchers who prepared the APTA report.
“What’s lacking or maybe stopping AI adoption by transit agencies is not an AI problem; it’s a data problem,” Zahedi said. “The AI is only as good as the data it has access to.”
Data that AI uses is often siloed among different departments: fare, schedule, finance, HR. “An AI, to be really smart, has to have access to all of that data combined,” he added.
“Accuracy, security, privacy, data handling, accessibility, governance, and cost all have to be considered before and after deployment,” said Blythe Nebeker, CapMetro’s public information officer, in an email to Civil Engineering Source. “AI is also not a substitute for well-defined processes, good data, or human judgment. Customer-facing or operational uses need appropriate oversight because an inaccurate response can have a real impact on a rider or an operational decision.”
She added that the “commercial model” is important: “Even when the technology works, the way usage is priced can affect whether a use case is sustainable. More broadly, we try not to assume that adding AI automatically improves a process; we validate the business value, technical fit, risk, and outcome before deciding whether to expand a capability.”
Still, across all of the categories of the APTA report, only a handful of surveyed agencies (five or fewer) are actually using AI tools.
Cecilia Viggiano, Ph.D., vice president at Boston-based EBP, a consulting firm focused on transportation and economic development, was another of the researchers who prepared the APTA report. She suggested that we’ll see more pilot programs unfolding in the coming years, but it’s still the early days. “I guess my hypothesis is it might take a bit longer for anything that’s really changing the way a transit agency runs,” she said.
For all the promise of AI, many observers caution that the future of mass transportation won’t begin with AI.
“At STV, our philosophy is that you don’t start with AI; you start with an operational challenge,” McElroy said. “And if you can’t clearly define that problem that you’re trying to solve, AI is not going to solve it for you.”
Feed Your Brain
Interested in artificial intelligence and other cutting-edge technologies? Registration is open for ASCE2027: The Infrastructure and Engineering Experience, a first-of-its-kind event bringing together subject-matter experts from all across the infrastructure space, March 1-5, 2027, in Philadelphia. So, yes, we know you’re excited about the Philly cheesesteaks, but ASCE2027 also is the perfect place to satisfy your appetite for learning.
Technical deep dives. Infrastructure expertise. Come feast!
Learn more about ASCE2027: The Infrastructure and Engineering Experience.
