Nemetschek’s AI guru Julian Geiger: construction needs industry-specific AI tools

But mostly, Google’s in a digital world. That makes things not really easy, but certainly easier than in our industry and in the physical world.

BMW is a little bit closer to AEC, but sits maybe in the middle. Basically, you have the physical world involved, not just digital, but it’s structured, controlled data. The factory is a closed environment. You have repeatable processes. A million cars are being made in the same way on the same assembly line, and this is quite helpful for AI adoption, especially pre-generative AI.

If you think about machine learning, you want to have a million samples of the same kind in the same controlled environment. You train a model on it; it predicts how good the paint quality is when you run the car through the line, for example. And that makes a lot of use cases possible. And it’s why manufacturing was ahead of the AI adoption curve compared to AEC.

Coming to the AEC environment, I think this is pretty much the inverse of the other two: every project we do is a one-off prototype, every single construction site is different, and it’s for that reason the industry was never really suitable for classic machine learning, because it doesn’t really scale. You train a model on construction site A, you deploy it on construction site B, and it knows nothing about what’s going on. It’s totally impossible for most of the use cases.

So we have a structural blocker to AI adoption in the industry, but then we also have a huge fragmentation issue. We have, in any kind of project, many dozens at minimum, if not hundreds, of participants working on the construction site, and the data is trapped and fragmented across all of those silos. There is no kind of assembly line whatsoever, no structure, no standardisation.

To summarise it, at Google, the data comes to you automatically. In construction, you inherit a fragmented mess from one-off prototypes for every single building. The first job I think for AI in AEC is to create that structured data layer that other industries started with. That’s where BIM comes in. That’s where AI can help structure data into BIM.

In the past, the Nemetschek Group has been a bit of a conglomerate financial holding company – very capable but quite isolated, with individual companies doing their own thing. I think a key learning I’m bringing here from Google is a data- and platform-first mindset, which is going to be super important for AI.

With AI, you want value to compound, not just be sitting in every single product individually. You want the insights that you create in product A to translate to product B to product C, and so forth to build the one-plus-one-equals-three game: ie a unified data foundation that sits across the Nemetschek Group, consolidating some of our AI capabilities into a cloud-native platform, and then run that AI and data platform across the group instead of launching scattered, isolated features. That that’s one key takeaway.

The second takeaway, from BMW, is that AI earns its trust by solving measurable problems in the real workflow, not just demos. At BMW, we had millions of showcases and prototypes. But getting people excited works only for so long with so many demos: after a year, it becomes dull. So you need to really get into the ROI and the measurable problem-solving: that kind of thing is really strong in automotive when it comes to quality inspection.

At Google, we worked with another carmaker. Their car painting line was a super energy-intensive process – and that was during the height of an energy spike. We used AI to optimise that process and save that company probably a two-digit million euro annual electricity bill. And that’s an amazing use case that everybody loves: it’s sustainable, and it’s reducing cost. And nobody lost their jobs: there’s nothing bad about that situation. I think creating such use cases is very important: where the ROI is not just net positive, but positive all around.

Coming in as an industry outsider, the stats about the AEC sector – 90% over budget, over time, 40% of global CO2 emissions coming from the sector and 20% of material wasted – sounded incredible to me. But I think AI can really fix that.

That’s a great question. I think the answer is a resounding yes, with the caveat that you need to get the right AI for the right job, and that you know what you’re doing.

The industry needs AI tools that are AEC-specific, not just tools that sit on a horizontal platform like ChatGPT or Claude. AEC needs output that is verified, that is tested, and that is ideally underwritten by the software provider – output that you can actually trust. You know, here’s the building design that’s actually constructable, here’s the structural outline that actually doesn’t collapse – we verified it with AI and non-AI measures. That’s the combination of probabilistic AI output with deterministic verification.

AI is a huge opportunity for our industry, especially the latest generation of AI. The previous machine learning era was structurally unfit for what the industry needs, given that every project is a prototype. Now we have that opportunity to scale, and I think the industry will see its biggest changes over the next couple of years. I’m so excited for that!

This is a very well-known theme and hotly played on the stock market as we all see it unfold in real-time. At first, the narrative was around thin, horizontal SaaS tools being at risk. So stocks dropped for pretty much everyone in the software market, especially horizontal tools that aren’t as vertically specialised.

We can all develop our own customised workflow tool this afternoon using maybe two hours of our time and $50 in tokens, right? That’s fine. But do you really want to do that on production systems? Do you want to keep that alive? Do you want to deal with all the security things that come up now, basically daily?

If you want to use that as a company, you basically need a full-time person or multiple people to look at all those things that come up: the support requests, the security holes, and the continuous patching and maintenance of that tool. Looking at most AEC companies, I think they don’t want to do that for their business operations tools. They want somebody else to do that. It’s not typically their core business.

In the vertical-specific space, look at Allplan, Archicad, Vectorworks, Bluebeam: they all have decades of industry expertise behind them. And yes, you could crawl through the product and basically let the AI write the description of that product and then build it. But some of that functionality is really, really complex to build. It took years of engineering to get that right and get it reliable. I don’t really see anyone having the appetite to build anything remotely like that, especially now with AI [embedded in the software].

We’ve seen decades of assets getting built using our products. The building projects that run through our tools are an extremely important signal for our product improvements, especially for AI services. There’s a major amount of waste in the industry, a major amount of time that architects, engineers and contractors sink into meaningless, low-value, low-end tasks that they should otherwise spend on higher-value tasks.

You don’t become an architect to triage RFI issues for six hours per week. That’s something I think needs to go away so that an architect can actually focus on why they became an architect in the first place. That kind of thing is the opportunity, and I think the value we create as a software company for customers using those exact kinds of use cases is going to be tremendous.

And that will expand our potential market size from just providing tools to delivering outcomes for our customers that move the needle. And I think those outcomes can be monetised; they can be priced and packaged in a way that customers will happily pay for. For us, it’s a huge opportunity.

Keep up to date with DC+: sign up for the midweek newsletter.