In a recent Roland Berger and Roebling executive webinar, Brentan Alexander, CTO and co-founder at Roebling, joined Pedro Caruso, Senior Partner at Roland Berger, to share how to leverage AI in industrial capital project planning. This recap covers why CapEx projects fail early, why general AI isn’t enough, what a live deterministic planning model actually produces, and the change in how teams work makes any of it pay off. 

Top 4 takeaways:

1. Most large capital projects fail in the early planning stages. 

Over 90% of large capital projects fail, and more than 60% run over budget or significantly late. Some of the earliest decisions carry the most leverage and get the least attention.

2. You can't safely vibe code a factory. 

General AI works where domain expertise is everywhere and mistakes are cheap. Infrastructure is the opposite, where one mistake could cost you millions.

3. Traceability is the precondition for trust. 

Every number in Roebling traces back to the document and line it came from. That's what makes AI usable on irreversible decisions.

4. The hardest part is people and process, not the tool.

Keep your old meeting cadence and you get none of the value.

Why do most capital projects fail?

Over 90% of large capital projects fail, and early stage planning is where we need to make smarter decisions. If you get technology selection, siting, configuration, or the basis of design wrong, nothing downstream can save your project. 

But that's also the stage with the least money and attention on it. Your spend happens later, when you're placing orders and breaking ground. So the highest-leverage decisions get made fast, on limited information, by teams who are out of time.

Pedro shared that EPCs and engineering firms make most of their money when dirt is moving. The planning stage is a bit of a distraction for their teams. The incentives just don't point at the front of the project, and the tools made today were built to replicate the drafting table from a hundred years ago.

Why can't you just use ChatGPT for capital project planning?

You just can’t vibe code a factory safely. Generally, AI works great where the domain expertise is everywhere and the mistakes are cheap. For example, software can be exactly like that. You’ll typically find a huge open library to train, run the code to check it, and if you find a bug you can patch it in the same afternoon. 

You can’t do that in industrial infrastructure. The expertise is locked in proprietary data and the heads of a few SMEs are those who are mostly near retirement. And you can't check a bug when you’ve already put steel in the ground.

That gap is why a general AI model isn't enough. You need physics, the domain knowledge, and you need to see where every number came from. Brentan called it a gray box. The AI drives the planning, a deterministic engine does the actual thermodynamics, and every decision traces back to the document and line it came from.

What did the live Roebling demo show?

Brentan prompted the platform by outlining a 1 million metric ton methanol plant on the U.S. Gulf Coast and let it start building the process diagram in real time. Then he switched to one that showed a biomass-to-power facility where wood waste was in and electricity came out. The system generated the full block flow diagram, an equipment list, cost every line, and ran the economics.

Brentan also showed that the answer wasn’t a single number. Roebling was able to run a Monte Carlo analysis and returned a range, roughly 140 to 215 million with a long tail toward 300 million in CapEx, plus a clear read on what was driving the uncertainty. At an early planning stage, it gave an honest output, and distribution and a list of what to do next. 

Is there a bottleneck to using AI on capital projects?

You can have the best data and the best AI and still get nothing if your meeting cadence stays the same. Pedro has seen examples where a question that comes up during a planning meeting kicks off a six-week loop. You give a new scope to the engineering partner, wait for information, a couple of meetings, then six weeks later you finally see how that one change moved your numbers.

By leveraging a tool that answers in minutes, you can set the follow-up meeting for after lunch.

That only pays off if you change the way you work. If you keep the same cadence, you won’t get the same value. The work only expands to fill the time you give it. We find comfort in wrestling with a problem until we run out of runway. The tool shrinks the time and your job is to let the decision happen faster and trust it.

That means sending the senior person who is the decision maker to the meeting, not the junior person taking notes. It means convincing executives that a recommendation that came back in three weeks is as good as the one that used to take a year. 

Does this mean handing over your competitive edge when using AI?

A lot of the real knowledge in this industry isn't in a database. It's in the heads of people who've run these projects for thirty years and are close to retiring.

Your data is yours. Roebling doesn't train models on it,  your data doesn't use one client's information to inform another's, and each customer is walled off in their own org. The whole thing only works if you trust the AI.

Have a project you want to test in Roebling? 

In Roebling, you describe a project, and the platform builds the process diagram, the equipment list, and the costed economics, with an uncertainty range and a read on what's driving it. The work that used to take weeks with third-party firms happens in an afternoon, and every number traces back to where it came from. That lets you run a thousand scenarios instead of three and still trust the answer.

You don't have to start big. Pick a few high-value questions you currently answer slowly, and answer them fast. Then change one meeting to match the new speed.

Bring us a project that you’re thinking of running and we'll show you what the deterministic model does with it. Get in touch.