Norfolk Southern is investing in artificial intelligence to reshape how it moves freight across its 28,000-mile rail network. And one of the people driving that train is Mabby Amouie, Assistant Vice President of Enterprise Data and Analytics, who oversees the company’s AI strategy — from making use of sensor data to deploying machine learning models that optimize safety and efficiency.
With more than 300 real-time diagnostic sensors on a typical locomotive, Amouie likes to say that the company’s engines were smart “before Tesla was even a thing” — not just relaying information about their own status, but also inspecting the track below them as they were in motion. Today, the company is layering advanced AI atop that data to power everything from predictive maintenance and scheduling to digital train inspections and track monitoring.
Norfolk Southern’s transformation began nearly a decade ago, but Amouie says the pace has accelerated dramatically with the rise of generative AI and new enterprise tools. “The efficiencies, the safety, the productivity—and all the gains we’ve made—are tremendous,” he says. Norfolk Southern now has hundreds of AI-powered solutions in production and is rolling out AI-assisted tools to employees across departments.
Based in Atlanta, Norfolk Southern is one of just a half-dozen Class I freight railroads in North America — the largest networks. It operates in 22 states and the District of Columbia. In Q1 2025, the company reported $3 billion in railway operating revenue, moving goods ranging from vehicles and steel to UPS packages. Its expanding AI capabilities are helping it deliver those goods with greater speed, reliability, and insight than ever before.
• • •
Can you explain where AI fits into Norfolk Southern’s railway business, and your role in designing an AI strategy for the company?
I’ve been around for a little over a decade and I really enjoy making an impact for a company, and an industry, that is virtually moving the US economy. …Class 1 railroads…ship anything from raw materials all the way to finished products, from BMWs and UPS packages all the way to steel and lumber.
Railroads have been around for two centuries. We’ve seen the industrial revolution, and I think the AI revolution is another revolution that a lot of industries quite frankly are going to benefit from, not just railroads. Anything I can do in my capacity, along with my teams, to be able to support this mission and strategy and take it to the next level is [fulfilling] to me. I’m…responsible for anything that touches data, from its ingestion all the way to the consumption side, where traditional dashboards and reports have evolved to more advanced analytics like artificial intelligence, machine learning, and generative AI. There are lots of opportunities to augment our people with additional data and AI and intelligence and basically the power that comes from the responsible use of AI.
What are some of the hypotheses or beliefs behind your AI strategy?
We had a very strong belief that AI could fundamentally transform different aspects of our business and how we run the railroad. …We saw a lot of opportunities to be able to supercharge our safety infrastructure, enhance our customer experience and improve our network fluidity and ultimately bring value for our railroaders, our customers, or the communities that we serve.
…The reality of it is that way before Tesla was even a thing, our locomotives had 300-plus diagnostic sensors…
We really made a bet on AI early on, and the other aspect to this is that we have a ton of assets that are intelligent and that have been generating a ton of data that generally folks really don’t know about. A lot of people view the rail industry, and especially Norfolk Southern, as one of the big classical railroads out there…that haul steel and cars, but the reality of it is that way before Tesla was even a thing, our locomotives had 300-plus diagnostic sensors that basically would tell you about how a locomotive was feeling, and the components of it, whether it was its oil pressure or water temperature you name it. These 300-plus sensors were sending and transmitting data in real time to the back office. Some of our locomotives have additional sensors underneath them to inspect our tracks as they are hauling traffic.
For AI you really need a ton of good quality data, in this case machine-generated data, and we had all of it. We embarked on the journey very early on, almost a decade ago. Ever since, we’ve just been building on top of it and more recently, a few years ago, we decided to turn this into a more comprehensive AI strategy, as opposed to just solving one or two business problems we bring them all together and make it a very robust comprehensive strategy. …We are witnessing how AI is really transforming our industry and in particular Norfolk Southern in the way we operate – the efficiencies, the safety, the productivity and all the gains we’ve made are tremendous.
What has changed? What is different from a decade ago to five years ago to now?
Most prominent to me is what is available to us. We have capabilities that five or 10 years ago, we could not even have conceived… The fact that you have AI right now helping you write code, and the fact that you can give AI hundreds if not thousands of documents [and then] ask questions…got us to think a little differently about our AI strategy.
We think about it in three major buckets. We think of using AI for what we call operational excellence — anything that has to do with running a better railroad, and making sure that we are super charging safety and increasing our network fluidity.
The second bucket is really about our customer experience, so that they would want to come to the railroad and ship with rail and really have a pleasant, Amazon-like experience.
And then the last bucket that has substantially changed since ten years ago is AI for our associates and railroaders. Whether it is productivity tools like a ChatGPT, tools like CoPilot, or whether it’s in the tech division with our developers and the way that they can utilize these productivity tools to develop code and to enhance the way that they test their code or design applications…AI really has reshaped itself over the past few years.
We don’t just want to jump on every shiny object, but at the same time, we want to tap into the opportunities as they present themselves to us…
We are closely watching, because every day there is something new coming out… We don’t just want to jump on every shiny object, but at the same time, we want to tap into the opportunities as they present themselves to us, and we are really focused on what really moves the needle for the three buckets of AI…
You mentioned changes that impact employees’ workflows. How do you help employees with change management, as well as shifts in employees’ skill-set requirements?
That’s a great question… Whenever we are trying to solve business problems, it is a very collaborative approach where the data scientists or AI scientists never approach a problem in their silo. They never come to a meeting and say, “Give me your requirements and I’m just going to go and vanish and after two months I’ll come back with the solution.” The way we have these problem-solving sessions is that the business brings the knowledge about their expertise, whether it’s in the mechanical area or engineering area or the commercial area. Then the data and AI scientists bring their experience in utilizing the AI tools…
One of the things that we have seen is that by including the stakeholders at the same table, and having those conversations early on, we are taking the approach where the change is not all of a sudden, like a switch… It’s very gradual, and it’s with a sort of like a training and understanding of the outcomes that we are going to achieve, and how we are going to achieve them…
The prominent example of that is our digital train inspection (DTI) portal, where we are fundamentally reshaping and adding layers of inspection with this technology. We included the right stakeholders at the right time, they have the proper training and bring that AI literacy for the entire organization… It’s a very collaborative and methodical approach to it, as opposed to, “here is a solution that is going to solve the universal problems that you have,” and all of a sudden creating a shock in the organization.
What approaches are you using to train employees?
I would say it’s a hybrid approach. We do have webinars, we do have events like office hours for a targeted audience. …The need for training is changing from needing to upskill or train a certain group of people, to going more broadly and bringing that awareness and training to the broader organization. …As we are rolling out some of the productivity tools, we’ve been attaching some articles on how best to use these tools to boost your productivity, whether it’s how it is going to compose an email for you, or how you can summarize a document.
How do you prioritize your AI projects?
That’s a great question, because, quite frankly, the opportunities are endless and so we do need to have a rubric to go by, because I don’t envision any company having infinite resources or time to put against every single business problem… Sometimes people forget to start with the most basic question of what business problem are you trying to solve, because in a lot of cases, you need to start with what the business problem is and then figure out what the solution is. It could be a dashboard, it could be a very simple report, it could be that you don’t even need AI to solve that business problem. So we are very methodical when it comes to that.
AI needs good data to begin with, so we are very cognizant of the concept of garbage in and garbage out…
As you know, AI needs good data to begin with, so we are very cognizant of the concept of garbage in and garbage out, because you can have either no data, or subpar quality data, that you put in AI and you might not get a good answer. It could be very convincing, but it could be the wrong answer.
How is the AI function structured within the company? Is it centralized, is it federated or is it hybrid?
We have a center of excellence with our AI scientists as of today. That does not mean that it’s not going to change; certainly at some point, it can become federated. We are empowering our business partners more and more to leverage AI.
You are using AI in many areas of your business, in train inspections, defect detection, proactive maintenance, AI-powered asset mapping, etc., Can you give me a sense of how challenging it is to manage AI across Norfolk Southern?
They are all very different solutions to different problems. To date, we have deployed nearly some 300 to 400 AI solutions in different domains and I’m not just talking about proof-of-concept or things that sit in a research lab. These are real tried-and-true production solutions that we have deployed and I can tell you it takes a lot to run 28,000 miles of track with hundreds of trains that run daily to move the US economy. The scale is massive, and sometimes people underestimate that. How we route our cars, how we plan and build our trains, how we sequence the cars on the train, their destinations, how we plan rail maintenance for that 28,000 miles of network with the terabytes of data that we have, how we plan maintenance on our locomotives, even down to how we inspect our railcars these days with some of the DTI portals, all of these are solutions that we have deployed for different domains, not only to solve business problems but also take it to another level of intelligence and make the best use of data to enhance the data-driven decision making and data driven operations…
Back to the Digital Train Inspection portal. Can you explain how that works?
For almost two centuries, the inspection of the trains and railcars have been done the same way across the globe, which is two people are on each side of a particular train walking along this train that could be three miles long. At every railcar, they visually look for different types of issues or defects that could be going on. It’s a very cumbersome process, it’s very time-consuming, and it is vulnerable to some human limitations.
You think of winter in Chicago at 2 AM in the morning, with fog and heavy snow. You are not going to be able to see everything that’s going on in that particular weather condition. So, there was definitely an opportunity for us to add an additional layer of safety. What we did is we have a mini tunnel that these running trains go through. These tunnels have a multitude of cameras and stadium lights. There are some 38 cameras with 24 megapixel resolution for more detailed images. Those cameras will come on along with like 42 stadium lights, and as the train is passing by at a speed up to 60 or 70 miles per hour, it can image our railcars from a wide variety of different angles, and they give us this 360-degree view of every railcar. On average, we take about 1,000 pictures of every single railcar that passes by, which is crazy to me when I think about it.
So if I have a 100-car train, I have 100,000 pictures on average…
Then, we have 75-plus AI algorithms that go through all these images and very accurately, autonomously, identify particular issues, and they automatically generate a work order for that particular railcar and also send an email to our network operation center…back in Atlanta at our headquarters, where we have a human who looks at that issue to confirm it and depending on the criticality of the issue, depending on the response protocols that they have, they take corresponding action. That could be something like bringing the train to a safe stop and addressing the issue right there, or it could be reducing the train speed for the safety of our operations, or it could be something that can be addressed at the next location. All of this is happening in real time. It’s been a safety boost, it’s been an efficiency boost, and it’s definitely been a network fluidity boost…
The first portal was deployed in Ohio in October 2023. Can you tell me how many DTI portals you have now, and what your plans are to implement more?
We have eight portals at seven locations fully operational, and by the end of this year, we are going to have 10 portals at nine locations fully operational. We are seeing tremendous value out of these portals so we are going blazing fast deploying these portals across our network and tapping into all kinds of benefits, with safety being at its core.
What’s the return on investment here?
ROI is something that it doesn’t always need to be in formal dollars. When you look at our stats last year, we’ve identified over 25,000 defects, and that was with a limited deployment of these portals. We also detected 85 tier 1 defects. These are some of those critical defects that [cause us to] bring the train to a safe stop and address the issue, so the safety benefits are tremendous. The efficiency gains and the network fluidity gains that we are seeing have been fantastic.
…We stop our trains less, because now we are proactively finding things before they find us, and we are giving this additional data and insights to our railroaders. Between 2022 and 2024, we had a 60 percent reduction in train accident rates. Now there were a wide variety of programs that contributed to that success and the metrics that we are witnessing, but the impact of these DTI portals cannot be neglected, and I would say the stats speak for themselves.
Earlier this year, one of your DTI portals discovered a cracked wheel while the train was in motion. Can you explain why this was such a significant event?
It was the first-ever instance of an autonomous detection of a cracked wheel by a system using ultra-high-resolution cameras and advanced AI algorithms, to the best of my knowledge, in the railroad industry. …Identifying such a minute defect in real-world conditions is an immense challenge, and this success underscores the transformative power of the portals and the vigilance of our teams. As soon as the crack was detected, our railroaders acted swiftly, removing the affected car from service before the issue could escalate — preventing a potential accident and avoiding network disruptions.
You also use AI in your Automated Track Geometry Measurement System. Can you tell me more?
There’s a lot that it takes to run this network, and one of the aspects is that we maintain our tracks. It’s not maintained by the federal government. …We spend roughly a billion dollars every year to maintain our tracks and roadbed.
Now, with the advent of sensors, what our engineering team did is they basically put a collection of sensors and lasers underneath our locomotives, and as the locomotives travel, the technology continuously measures track geometry. …So as you are hauling traffic, you are autonomously inspecting the track by gathering all the sensor readings and then we have processing that happens with AI and other technologies that basically analyze this data and points us to the problem areas, areas of interest that we need to address.
Who do you report to?
I report to the EVP and chief information and digital officer, that’s Anil Bhatt.
And who does he report to?
The CEO, Mark George.
How often do you meet with the board?
The technology team meets with the board frequently, and I would say on the AI initiatives, from time to time, we provide updates to the board of directors if they have interest as well.
What do you think is the next chapter for AI at Norfolk Southern?
I always get that question from the community, and I’m going to give you the answer that I give the community. Whatever answer I tell you now, a year from now I’m going to be wrong, because the advent of AI innovation is so fast, the pace at which we are innovating is so fast, that quite frankly every day there is a new large language model, every day there is a new thing that’s coming up that we would not be able to even predict a day before.
You must be logged in to post a comment.