Public infrastructure projects aren’t just large-scale, they’re layered with complexities that engineers are expected to solve under enormous pressure. Between tight government budgets, changing site conditions, rigid deadlines, and multiple stakeholders pulling in different directions, the margin for error is razor thin. Most teams are still working with outdated processes, siloed data, and traditional planning tools that can’t keep up with the pace or scale of modern demands. These limitations lead to the usual pain points: delays, budget blowouts, design inconsistencies, and poor coordination. For engineers, that means long hours trying to troubleshoot problems that should’ve been avoided from the start.
This is where artificial intelligence (AI) and machine learning are beginning to reshape what’s possible in public infrastructure. These aren’t just tech trends, they’re becoming real solutions to long-standing industry headaches. AI-driven planning tools can analyze thousands of project variables in seconds, helping teams make smarter decisions early on. Machine learning models can predict cost overruns, safety risks, and scheduling clashes before they happen, giving engineers a chance to adapt before issues spiral. Whether it’s managing traffic flow for a highway expansion or optimizing utility layouts beneath a city street, AI is helping engineers approach infrastructure with more clarity, speed, and accuracy.
And the payoff is real. Teams using AI-enhanced workflows are seeing faster delivery times, fewer change orders, and millions saved in reduced rework and material waste. For engineers, this means more room to focus on quality, innovation, and sustainability, without getting buried in repetitive tasks or fire drills. As more government agencies and infrastructure bodies begin mandating smarter planning processes, the integration of AI isn’t just becoming an advantage, t’s becoming a requirement. In this article, we’ll explore how AI and machine learning are transforming the way we design and deliver public infrastructure, and why now is the time for engineers to lean into this shift. We have the opportunity to work with many public sector agencies trying to help them build differently and bringing a whole host of various consulting services and most recently AI which has been a great learning experience for me. A few months ago some group of engineers conducted a survey and found that 46% of engineering and construction firm leaders believe that they lack the technical skill and training to adopt artificial intelligence, and 40% of construction executives in that survey maybe with some of you, I’m not sure who actually took it, perceived that the reluctance of staff to change was a barrier to adopting AI. Yet at the same time we know that it’s coming. It’s actually already here. AI is already built into so much of what we do every day. 83% of companies consider using AI as part of their corporate strategy. Everyone is talking about it. I know certainly at Del every week if you don’t have at least two or three meetings about AI you have to check in and see what’s going on because you should be. The number of businesses that use AI grew by 300% over the past 5 years and so it’s very challenging especially in an industry like ours where our number one concern is often safety. We need to be slow to change. How are we adopting AI?
Panel Introductions and Strategic Focus
The panel today have leaders in their organizations and have been leading in adoption of AI. We’re going to hear some use cases, stories about how AI is helping their organizations and how it’s being adopted. We’ll go from there. We’ll have a panel discussion. If there are questions, you’ll be able to find the panelists floating around. Before we get started we’ll do a brief introduction. You want to start Brad with an introduction of yourself and your organization? Sure. Good morning everyone. Brad Levens. I’m the Executive Director for Strategy and Organization Resiliency for New York MTA Metro North Railroad. Under my purview I have everything from the strategic planning of the railroad to benchmarking international research to AI governance to communications within the railroad as well as a few other areas under my portfolio. One of the new areas is also standardization of work. It’s a new endeavor sponsored by our chairman Jan Liber at MTA. One of the first agencies, Metro North Railroad, to start down that standardization path. As we talk today on the panel and maybe afterwards AI is going to play a large role hopefully in making sure we’re efficient in that standardization. If you have any questions on what Metro North or who Metro North is or what New York MTA is, if you don’t know already, I’m happy to address that after the panel. Thank you.
Good morning everyone. I’m Faran Abdullah. I’m Director of Information Technology at New York City School Construction Authority. SCA is responsible for design, building, and renovation of New York City public schools across all five boroughs and we’re making sure that the schools are safe, sustainable, and safe for the kids in New York City. Hi. My name is Khan ASB and I’m a little bit of an oddball here. I’m a professor and the director of the Transportation Center at NYU. It’s called C2SMART and it’s funded by the US Department of Transportation. We work very closely with all the agencies dealing with transportation in the region and also in our partner regions which is Washington, Texas, and so on. It’s a national center. Our claim to fame is to use big data sensors combined with predictive technologies and AI and machine learning. We do training, education, research, development, and deployment. In a minute I can talk about some of the projects. We’re in Brooklyn, part of NYU’s Tandon School of Engineering campus. If you want to visit you’re always welcome to come and see what we’ve been doing in terms of AI, machine learning, robotics, and transportation projects.
How AI is Transforming Public Sector Infrastructure and Transportation
AI in Safety and Operational Performance at Metro North
I love that we have a mix of public sector, private sector, and research. I think that’s a good combination for a conversation. Why don’t we start with some examples of use cases that your organizations have been using around AI? Sure. Thank you. So break this down to two phases or sections. We’ve really broken this down into two components. The first is projects or AI use cases that are within the security, safety, and operational performance spectrum. The second phase, which is more long-term, would be things such as operations resiliency, efficiency, and optimization. Those are longer-term projects. From there we move more into perhaps things that are more specific to the standardization work that I mentioned in the beginning. Let me talk to you about a couple of examples in each of those sections and phases. The reason those phases are so important plays off of what Obi said about AI being in every conversation. AI is a lot of sex projects assets. You hear, “Hey let’s throw this widget onto the side of our train. Let’s throw this camera on the side of our right of way. Let’s put cameras inside of our cars. Let’s do X, Y and Z.” It all sounds really great and fascinating, but if you don’t have a phased and logical approach or a strategy to how you’re going to roll out and learn from your rollouts, lessons learned, the good, the bad, the ugly, then you’re setting yourself up for failure. Part of a strategy is a phased approach. That’s why we have these two or three phases we’re going through.
So what are the use cases under phase one, which is really more about ensuring the safety, security, and operational performance of our railroad for our people, our employees, and our customers? Last year alone we had 60 some million passengers on Metro North Railroad and we have 6,300 employees. Safety and security for those passengers and for those people are paramount to everything that we do first and foremost. What are some of the things AI is doing that we put in place? These are pilots we’ve just started. These are not all fully operational. Some are, some are in the pilot stage. The first one I’ll mention is forward-facing trespassing detection. We’re using AI on trains to identify anomalies on tracks right of way across our service operations. Rule 22 compliance. How we communicate across our sectors. We’re using AI to successfully transcribe audio recordings into text and to ensure Rule 22 compliance and optimize language use by our conductors, engineers, and staff. Train door detection safety. We have cameras on our platforms looking at when and the AI algorithms are cueing us. Are trains opening at the exact right point of stop and start? Or do we have door opening a little offset that could perhaps cause a customer or an employee risk on station platforms?
AI-Powered Security and Environmental Monitoring
Intrusion detection using AI and infrared night vision and other alerts for intrusion or trespassing detection on our right of way and in our facilities helps with security. Our security teams are tight. How are we able to really scan a vast and massive territory of Metro North with limited bodies? How can AI help us with those security data points? Grade crossing violation detection analytics. Automated passenger counting. We also have a laser train that uses lidar and is building a neural network to tell us when we are trying to clean fall foliage off of our tracks using a laser versus water which is an environmental savings. How long should we have that laser on the track? The AI is helping us understand where the points of most need on the track are. Where are those deviations? So we are optimizing how we’re using that lidar in order to keep our track safe and clean, and reduce slippage of our trains for operational performance. Then later on I’ll pause here. That’s the first phase for security and safety. In a moment I think in the next question I’ll move a little bit more into those longer term projects which are really more about operational efficiency and resiliency.
Automating Code Comparison and Safety Updates
Inspection checklist that reference specific SE sections in New York City building codes, and as you know these building codes often get updated. What that means is our team needs to update those codes as well in our deficiency database. Just imagine going through thousands of sections and 3,000 plus pages of building codes, fire, mechanical, plumbing codes. An approach that took weeks, not months, for us to do that comparison, review, and update. We delivered a very powerful comparison engine which basically extracts the deficiency list from our database, compares them with the sections across different code books, identifies those changes, highlights those changes, and provides visual indicator for our users to review. I even seen the tool do flag even missing. Yes, we can update our deficiency database in hours as opposed to months now, but more importantly is we want to make sure that we’re keeping up with those critical updates that coming up for the New York City code and that can directly impact our safety for the schools.
Using AI to Enhance Accessibility and ADA Compliance
Another use case worth highlighting is we delivered for our accessibility team. The tool can read very dense architectural drawings in PDF to identify the accessible elements like toilets, the ramps, the shape of the ramps, is it straight or the L-shaped, the U-shaped, and water fountains and so on. It’s making it easier for them to quickly identify those elements and then used for reporting and downstream processes. We’re actually in the process of expanding the capability of the tool to whether do dimensional analysis, whether the side grab bar or the back grab bars or toilet or the toilet seat height from the floor actually comply or meet the specifications set by the American Disability Act 88. There are some other exciting things we’re working on.
Leveraging Complex Facility Data and AI for Long-Term Infrastructure Insights
Some of the, we’re in the scoping phase of reviewing very complex data set. We call it building condition assessment. Survey inspectors go out to all our facilities, school facilities, and they take photographs, data, the comments from custodians, comments from principals, and that assessment is used to prioritize our capital improvement program. We’re looking into how we can look at the data set to make sure our schools are always in the state of good repair. We’re also looking at our closeout documents. We produce thousands of documents whether it’s PDF, walkthrough videos, and so on. One of the challenges we have is generating insights from those documents. Once we deliver the school to DOE, if we need to change the roof now or we need to change the boiler, what is the warranty information, who’s the manufacturer? Or if we need some sort of operating guidelines which may be buried in a walkthrough video, we’re trying to figure out how we can generate this insight and provide those insights to our users. So there’s some exciting things we’re working.
How Public Agencies Are Using AI for Building Code Compliance, Infrastructure Upgrades, and Safety Innovation
So I’ll talk about a couple of projects but first, just going back to what we heard this morning. AI unlike any other technology is moving forward so fast that we don’t have a good understanding especially with our world department of transportation, how it’s being adopted or how it can be adopted. The National Academies a couple of years ago asked us to do a nationwide study to see how DOTs are using AI and how they’re adopting it. As part of that we just finished the project a couple of months ago. We talked to different DOTs and asked about their projects. The difference is amazing. Some DOTs are really ahead and trying to do things, some are not, but there’s a common theme. Again, what we heard this morning, there’s not enough people who know what AI is and feel comfortable about what it can do. The report is out there. I can send it to you. It was a kind of a roadmap of what DOTs can do. We also found out very interesting things about not only DOTs but also their consultants and the lack of understanding of AI.
Just very quickly, one of the largest DOTs in the nation, they were using computer vision to identify road litter and they had a few hours of labeled data. They trained an algorithm to do the litter detection and the algorithm was working correctly 90% of the time and they closed the project. They said this is not good enough. We want 100%. Of course, what they didn’t understand this was only a few hours of training data and this could work 24/7, 365 days and eventually it would get almost perfect. They didn’t want to do that and actually they missed probably an opportunity. It will come back, but again, that shows that there’s a need for understanding the technology and how it’s evolving. On our side, we are doing several interesting projects. One with New York State City DOT on the BQE Caner section. When we were involved in that project as C2 Smart at NYU, we suggested to DOT that you could really extend the lifetime of that structure if you reduce the truck loads. Especially there were a lot of overweight trucks. We installed what’s called We Motion Sensors and then we came up with the idea of using computer vision and automatic ticketing. First time in the nation for ticketing overweight trucks so that they don’t use that part of the roadway. That lower loads will extend the lifetime of the facility, will give enough time to DOT to do repairs until there’s a plan for replacement. I think this was a very successful project that used our engineering expertise using the impact of the loads on this aging infrastructure, AI, computer vision, and so on and then helped DOT to address a real-world problem.
The other one is more fun. With my students during COVID they were bored. They came and asked what can we do to understand if people are listening to the distancing guidelines. We decided to use the public feed from DOT cameras. There’s thousands of them. They’re publicly available. We used computer vision to, without calibrating, without going out, to have an estimate of the density of the people around intersections. We kind of perfected that technology. Now we are working with DOT, with Port Authority, with a couple of other agencies to use existing cameras. This is like a repurposing of the existing infrastructure without using new cameras, new investment, but using the existing infrastructure, repurposing it to detect accidents, to detect conflicts, to detect traffic. Wherever it’s possible to take work zones, for example in New York, there’s no way really knowing exact time of the work zones. It’s like a period of time they can do the work zones, but from these cameras we can identify work zones and many other things. This is really important because the technology is really cheap. You’re using algorithms that we calibrated and then you’re using the existing cameras without investing new dollars. I think this kind of repurposing, reusing existing investment with the use of AI for things that we cannot imagine right now is going to be the next big thing. This will save public dollars and also it will be very exciting for researchers like us to see what are the opportunities.
Addressing Governance and Labor Considerations in AI Implementation
Something that we are definitely leaning into. I know the gentleman from New York New Jersey Port Authority mentioned death by process. We need to do that smartly, but we do need to have process and governance as it’s concerned around AI camera use.
For instance, we have to be very careful with the labor agreements with unions, that we’re not using cameras for punitive action in a way that’s going to contradict the agreements and the labor agreements that are in place. However, we do want to use the cameras and the algorithms to spot and identify conflict, security risk, safety risk, etc., throughout the train internally and on the right-of-way externally. So a piece of the governance.
What are some of the other things that we step back to? How do we get this started, which I think is your question, and what do we go through?
Strategic Planning and Risk Governance for AI Rollouts
The first thing I mentioned in the first question was we paused and said we like all of these really amazing ideas and assets and technologies. But what is the rollout strategy? Strategy is number one. A phased approach, what can you do simple: safety, security, operational performance, and what is going to be more complex in the later phases. That’s the operational efficiency items such as predictive maintenance, track scheduling optimization, those type of things.
The next piece was part of the governance is risk management. We are following the New York State and the MTA policy on IT and AI risk management application. It follows a lot with the National Institute of Science and Technology, the NIST risk management framework, which New York State uses.
All government agencies should be using and MTA is a New York State government agency. All government agencies in New York should be using the NIST policy which aligns with the New York State policy. Risk management allows us to ensure that we’re not slapping a bunch of technology on the side of our assets and our cars because it sounds or it is operationally relevant.
How Public Transit Agencies Are Building Ethical AI Systems in Transportation
Have we spoken with our chiefs of security, our chiefs of safety, our chiefs of personnel? Have we looked at data privacy, data verification and validation? Have we looked at the reputation of the railroad, the customer impact, the employee impact? Are we using data sharing, data analyzing, data reporting on data that could have an adverse effect on the reputation of a state government local agency?
That all is very important that you go through, especially in the government realm, to ensure that you’re not doing something that seems great that’s going to come back to haunt you in the long term. So strategy, risk management, data validation and verification.
I think we’ve hit on this a little bit. Gentlemen, you did for sure in the first round of questions on communications and training, and we’re not there yet because we’re new. So we can put all these governance process phased approaches and strategy in place, but we are struggling.
We all should know we’re in the right spot with common sense, consistent language around what is AI, where are we in the AI adoption life cycle and where are we headed. That takes a lot of consistent communications, simple language, and training at all levels of your workforce.
The Importance of Human Oversight in AI Decision-Making
I’ll close the question by saying that let’s not forget when you’re going through all those things about standup, a lot of people forget to mention the phrase of human in a loop. AI in our agency and it should be in all agencies, as far as a government agency is concerned, we should not be allowing AI to make decisions on behalf of the agency. AI collects data and information that provides better information to inform decisions that a human, a leader, will make on behalf of that institution.
That is very important, the human in the loop. I’ll pause there.
So I’ll pick up on human in the loop conversation. One of our early challenges, and still is a challenge, is to help our users understand the probabilistic nature of AI. I’ll give you a quick example. When we said this computer vision model identifies the accessible toilets with 90% accuracy, the question came back — what do you mean by 90% accuracy? What happened to the 10%?
Clarifying AI Accuracy and User Expectations
We found that reframing that conversation and providing our user with the review screen that consolidates data from multiple places, and quickly for our experts to look into that data before sending it to the downstream processes, was more helpful.
Picking up on that AI, this approach is helping us to make sure that the AI’s capability and this recipe is working for us very well. AI’s capability, combining that with our experts, is actually a great strategy for us moving those AI projects forward.
I think just as a university, obviously you’re very interested in training. Even the report I mentioned made that point that there’s more training needed. The other thing is that human in the loop is great, but also AI is just an algorithm.
Understanding AI Accuracy, Human Oversight, and Workforce Training in Transportation
If I make that statement in my class, if I called artificial intelligence artificial prediction, nobody would have been interested. When you put the intelligence, the human kind of a thing, then people get interested. But it’s not that. It’s just an algorithm. We need to understand that. Expecting 100% from an algorithm is not correct.
Human beings make mistakes too. We collect data with the students. There’s like 25% to 30% error in regular traffic data collection. You collect data with sensors, that’s 15%, so you need to go back and do cleaning and so on. There’s nothing perfect in this world. We need to understand that.
But I think my suggestion is that the technology of AI and machine learning is moving so fast, there’s no time to wait. I think training, understanding, and moving forward is the only solution. There’s nothing else we can do.