What Titan Cement got right about industrial AI

In April 2026, HBR published a Cold Call podcast episode and case study, Transforming a Titan, profiling Titan Cement International — a 120-year-old Greek family-owned cement business that went through a triple crisis (the 2008 Florida subprime collapse, Egypt's revolution, and Greece's bankruptcy) and emerged with one of the first cement plants in the world running fully on AI in closed-loop control.

The story is unusual because the AI build started in the worst possible economic conditions for the company. In the depths of the Greek bankruptcy, around 2013-14, Titan had three of its largest markets in simultaneous freefall — Greek cement volumes were down 82% from peak and stayed there for ten years, Florida volumes were down 75%, and Egypt was politically unstable. Capital expenditure was effectively impossible.

That constraint is what produced the AI plant optimizer. Dimitri Papalexopoulos, who was CEO at the time, describes the logic on the podcast: "ours is a capital-intensive business. If you want to grow, if you want to really improve, you have to invest serious capital. We didn't have any capital to invest, yet we had good people who were sitting around and dealing with crisis year after year after year." So the question they asked was specific: what can we do that's people-intensive, not capital-intensive? The real-time plant optimizer was the answer.

The case is the best public example I've seen of how to do industrial AI with discipline. There are four things to take from it.

One use case, measurable, in the core

Professor George Serafeim, who wrote the case, describes a pattern he sees constantly with the executives he teaches: organizations show up with 267 AI pilot projects and no idea which one moved the financials. "Which one actually really makes a difference?" he asks on the podcast. "And how much has it improved the financials of the organization?" The answer is usually a story, not a number.

Titan went the other way. They picked one use case — closed-loop AI control of a cement plant's production process — and went big in the core of the operation, not the periphery. They picked it specifically because the result was measurable. Plant output, energy consumption, yield, and quality variance are all hard numbers. Either the AI system improved them or it didn't.

That measurability matters for two reasons. The first is straightforward: it forces accountability. You can't tell a nice story about a closed-loop optimizer; either the plant runs better or it doesn't. The second is cultural. Serafeim describes a flywheel effect: most plant managers, when first approached, said "not in my plant — don't break this, this is working." Once the first installation was running and the numbers proved out, the same managers were calling and asking when the team would arrive at their plant.

If you're building AI agents in a non-tech business, this is the order of operations: pick one use case in the core where the metric is unambiguous. Don't run 200 pilots in marginal areas hoping one breaks through. The breakthrough comes from one measurable win in the core.

Cement heads have to talk to digiheads

The technical breakthrough wasn't the AI model. Papalexopoulos is direct about this: "the key thing was getting what we call the cement heads to talk to the digiheads." In most organizations those two groups talk past each other. Domain experts say "this is what we have to do, you don't understand the process." Digital teams say "this is what we have to do, you don't understand the model." Neither side ships anything.

What worked at Titan was finding integrators who could speak both languages — people who understood enough kiln chemistry to make the right model trade-offs and enough machine learning to know what was actually feasible. That's the role most enterprise AI org charts are missing. There's an ML team, there's an operations team, and there's no one whose job is to make them ship together.

This is the same gap we run into when we build AI automation for clients in regulated or process-heavy industries. The model is the easy part. The real engineering is in the data integration, the operator workflow, the plant control layer, and the change-management work to get the people running the line to trust the system. None of that lives in any single team's backlog by default.

Innovating under constraint, not under abundance

There's a teaching point Serafeim returns to: most case studies on innovation are about Google or Apple, where the problem is "what would you do with $20 billion?" Those cases miss the more useful question — what do you do when you have no capital and you still need to transform the business?

Titan's answer was that scarcity forced focus. With unlimited capital, you can run 267 pilots. With no capital, you run one — and you'd better pick the right one. The forced selection became a feature. The team had to be specific about what business outcome the AI was responsible for, who would own it, and how it would be measured. That kind of clarity is hard to get when budgets are loose.

Most enterprise AI roadmaps in 2026 have the opposite problem. There's plenty of capital, plenty of pilots, plenty of board-deck slides, and very little operating impact. The Titan playbook — pick one core use case, measure the financial outcome, integrate domain and digital — works under abundance too. It just doesn't get applied because nothing forces the discipline.

Three transformations on the same body

The other thing the case makes clear is that Titan wasn't just doing digital. They were running three transformations simultaneously, on the same operating business, with no capital cushion.

Digitalization — the plant optimizer plus a small set of other AI applications, building toward what Papalexopoulos describes as the "first digital cement company."

Decarbonization — rethinking the energy mix, reducing the clinker ratio (clinker production is the major source of cement's CO₂), substituting alternative materials, planning for carbon capture and storage. This wasn't optional environmentalism; it was a long-horizon balance-sheet calculation. Cement plants run for decades, and carbon will at minimum be an expense and likely a liability over that horizon.

Decommoditization — moving from "we produce cement, here's the price" to customer-centric solutions: niche products like ocean-resistant concrete to defend Florida coastlines, low-carbon variants for builders who want them, application-specific formulations.

These three are linked. The digital layer is what makes decommoditization possible — once the plant is run by an AI controller, you can fine-tune it for a niche product without losing volume on the base. And the decarbonization investments compound when the optimizer is squeezing every gram of clinker for efficiency. None of the three would have worked alone. Multi-axis transformations are the norm in industries surviving structural change, not the exception.

What this means for your AI roadmap

If you take Titan's experience and pull it into a 2026 AI roadmap, four things follow.

• Find your core use case before you build a platform. Most enterprise AI programs build the platform first and hope use cases emerge. Reverse the order. Pick the one core production decision that, if AI got better at it, would meaningfully change the P&L. Build the platform around that, not around imagined future workloads.
• Hire integrators, not just modellers. The bottleneck is people who can hold both sides of the conversation — domain process and ML capability. They're rare and worth paying for. If your team is all data scientists and no process engineers, your model will technically work and operationally fail.
• Make the metric the product spec. The Titan optimizer wasn't shipped when the team thought it was good enough; it was shipped when it could be measured against ground truth and beat the legacy controller. The eval harness is the contract. Build it before scaling user count, the same way we approach this in our InsightAX deployments.
• Plan for multi-axis transformation, not point solutions. If your business is going through structural change — regulatory, competitive, climate, or otherwise — your AI work is one axis of three or four. Sequencing matters. The digital layer usually has to come first because it makes everything else cheaper to run.

Serafeim ends his case discussions with one question: "Are you leading in a way that the organization becomes more dependent on you over time, or less dependent on you over time?" The same question applies to AI. Are you building AI capability that makes the organization more dependent on a single vendor, a single model, or a single outsourced team — or less? Titan's answer was the latter. They built the muscle in-house, with integrators, and the optimizer is now a part of how they operate, not a service they rent. That's a useful bar.