AI In Manufacturing Is Hard Says A CEO Actually Doing It

Editor’s note: Scott Carlton will share his story at Chief Executive’s Manufacturing Leaders Summit in St. Louis on May 5. Join us!

Scott Carlton will tell you upfront: He doesn’t know his AI-ROI yet. Ask him in 2028.

Carlton, president of Tokai Carbon U.S., a $300 million annual revs maker of graphite electrodes for the steelmaking industry, is now 20 months into what he calls a full-blooded AI and data analytics transformation, without a penny of return in sight—just yet.

But that doesn’t mean he’s going to stop grinding (and it absolutely is a grind, he says).

The reason is simple: His effort is unlocking the biggest opportunities (and potentially the only big opportunities) in his company’s pitched battle with rivals, a battle where one or two percent of cost savings, margin or speed to market make all the difference. It may be even more—the creation of unassailable, perhaps permanent competitive advantages that cannot be mimicked by anyone else in his industry.

That said, he’s not here to sell you on the hype.

“It’s a lot,” he says of the project. “There’s no way you can sit there and say, by the end of the year I’m going to accomplish this. A CEO cannot set an erroneous goal of 2 percent times X. You’re just gonna lose everyone.”

His big advice for manufacturers approaching AI with trepidation: Get the right people on the bus, clean your data before you do anything else and get curious, get into the details and be honest about the timeline. “It’s a journey,” he chuckles. “It just kind of is.”

Start with the data problem you already have. Carlton’s AI journey didn’t start with AI. It started in 2021 with a mundane infrastructure decision: The company’s manufacturing execution systems were 15 to 20 years old, running on aging servers and completely disconnected from each other. Engineers were spending three to four hours every week just doing “pivot table gymnastics” to consolidate basic production data. The question was whether to simply replace like-for-like—or go bigger.

Carlton went bigger. “I think we need to look higher than just replacing the systems,” he told his team. Tokai Carbon had accumulated 15 to 20 years of rich technical and operational data, but it was all scattered across disconnected systems, effectively sitting in the garage. Before any analytics or AI could happen, the garage had to be cleaned out. That data cleansing process took two years—nearly double the six months their implementation partner had projected. It wasn’t glamorous, and it wasn’t fast. But Carlton is convinced it was the only path. “If we did this and didn’t do that, our future is useless. You can’t do anything.”

Find the unicorn—or build one. The biggest unlock in Carlton’s journey wasn’t a software platform. It was a person. He recruited a data analytics specialist away from a competitor—someone who was simultaneously completing a master’s degree in analytics and, critically, already knew the industry cold. “You have to find someone who knows the business and who has an analytics background,” Carlton said. “You gotta find the unicorn.”

When the unicorn is hard to find, you build the team around the gap. Carlton paired his industry expert with a dedicated data engineer and let them work in parallel, largely left alone. When colleagues asked what the new hire was doing, Carlton’s answer: “I don’t know. Just let him do his thing.” The combination—deep domain knowledge on one side, technical data skills on the other—is what made the data cleansing possible. Without someone who understood what extrusion pressure and milling time actually mean in a graphite electrode operation, no data engineer in the world could have made sense of the correlations.

Don’t expect off-the-shelf AI to solve a niche problem. Carlton’s team has experimented with commercial AI tools—ChatGPT, Perplexity, others. The results have been, in his word, useless. “We put it in there and we get the most obscure strange answers that make no sense.” The reason is simple: Tokai Carbon operates in a highly specialized industrial niche for which essentially no public training data exists. The large language models have never heard of his industry.

The implication is significant: For manufacturers in specialized or proprietary processes, generative AI is largely a dead end—at least for now. What Carlton actually needs isn’t a chatbot. It’s high-speed pattern detection across massive proprietary datasets. “Tell me what the anomalies are. That’s really what we’re looking for.” The roadmap points toward building a custom LLM trained on Tokai Carbon’s own data—but that’s a future chapter. Right now, the priority is getting the data clean and the test cases running.

Define the win for year one—and keep it humble. Carlton is clear-eyed about what success looks like in the near term. It’s not a financial return. It’s organizational muscle. “I want to build the methodology into the organization to where it becomes commonplace. That’ll be the win this year.” Specifically, he wants his teams running the problem-solving methodology—hypothesis, data, statistics, test, iterate—with real data backing every step, not the informal “good enough” approach that passed for rigor before the system existed. “In the past, we were kind of bullshitting ourselves.” Financial benefits, he believes, will follow in 2027. A hard ROI number? Check back in 2028.

The CEO has to be in the details. This is perhaps Carlton’s most emphatic point—and the one most at odds with how senior executives typically engage with transformation initiatives. He’s not talking about receiving a dashboard with green and yellow dots. He’s talking about understanding mixing times, extrusion pressures and how a 32-inch electrode performs in a DC furnace versus an AC furnace. “You have to be willing to get into the details. If you’re going to deal with a PowerPoint that has a green dot or a red dot, you may miss out.”

The reason isn’t micromanagement. It’s direction-setting. Without deep engagement, a CEO can’t meaningfully guide what gets measured, what correlations matter or whether the data scientists are solving the right problem. “Everyone’s got different perspectives of what the end in mind is. You gotta make sure your data correlates well—otherwise all this fancy stuff, it’s just not gonna work.”

The competitive moat is the journey, not the output. Carlton is thoughtful about IP protection—in a niche industry, a competitor who learned exactly which process variables Tokai Carbon had optimized could theoretically replicate the result quickly. But he’s also philosophical about it. The real competitive advantage isn’t any single insight the system produces. It’s the two-year data infrastructure build, the institutional methodology and the accumulated learning that come from actually doing the work. “They could steal today’s trade secret,” he says. “But I’m probably going to have moved on to the next one.”