AI in Construction is no Longer a Category—it's Infrastructure

“AI is becoming more like computing in general, just a mean and not an absolute thing.”

That observation from Sylvain Hellin, Managing Director of the Leonard Obermeyer Centre at the Technical University of Munich, captures a shift happening across construction technology. AI isn’t a separate product category anymore. It’s the substrate on which the next generation of construction tools is being built.

The Data Gap That Defines Construction

Before understanding what’s being built, it helps to understand what’s missing.

Manufacturing has doubled its productivity over the past 22 years. Construction has managed 10%. The compound annual growth rate: 0.4% for construction versus 3% for manufacturing.

The explanation isn’t attributable to one factor because construction is uniquely complex. But a core problem is lack of data.

“If you compare to manufacturing where they have tens of sensors everywhere, or software-as-a-service applications that can track everything—every mouse movement, how long you spend doing what—we can improve products,” Hellin notes. “Unfortunately, we don’t have this in construction. At best, we get a daily report or most likely a weekly report written on an email explaining on a high level what kind of progress happened.” That’s all the information you have.

Every technology highlighted here solves some piece of this data problem—creating the foundation for what comes next.

From Site Selection to As-Built Models: AI Across the Construction Lifecycle

Early stage: What should I build here?

Matthias Zühlke spent 15 years as an architect answering this question for developers. Every time: gather zoning data, engage architects, hire energy consultants, run cost calculations. Weeks of work before knowing if a site was worth pursuing.

syte, the platform Zühlke co-founded, maintains digital twins of every building in Germany—over 25 million properties. Unlike most AI construction startups, syte built its own foundational model, winning Germany’s AI award. “We don’t offer a wrapper on ChatGPT. It’s our own AI.”

The result transforms that weeks-long feasibility process into something almost instant. Enter an address, and syte returns building potential, energy analysis, market data, renovation simulations, and ROI calculations—all immediately. Portfolio managers now prioritise across thousands of properties for renovation potential, a task that would have been unthinkable at scale just years ago.

“The capacity of LLMs is about 10x since three or four years ago,” Zühlke observes. “This helps us make simulations over portfolios of one million buildings.”

Design stage: Can non-architects explore spatial ideas?

TUM’s research here is perhaps the most striking. They set out to answer a provocative question: what if a property developer with no architectural training could explore building designs without hiring an architect? To test this, he built multi-agent systems that generate complete BIM models from natural language prompts.

“I want a two-storey hotel with eight rooms on each floor. Lay out the rooms so that four rooms are placed on each side separated by a 4-metre-wide corridor.”

The LLM translates this to programmatic commands, controlling Vectorworks BIM software. Solibri model-checking validates the output, feeding errors back to the agent for correction. The loop continues until the model passes compliance checks.

The team has since pushed even further with what’s called “computer use”—giving AI agents direct control of mouse and keyboard, just as a human would operate the software. No APIs, no custom integrations. The agent simply sees screenshots of its work, decides where to click, types commands, and iterates. The result: roughly ten minutes to generate an editable BIM model from a hand-drawn sketch on paper.

“LLMs have a lot of potential to assist specifically non-designers—like property developers—to play with design ideas in the early design stage,” Hellin notes. But architects shouldn’t worry: “We are very far away from having any kind of this system taking away the real job of an architect.”

The current output reflects this reality: the buildings look like buildings, but they’re nowhere near construction-ready. What matters, though, is what this capability enables. For the first time, non-architects can explore spatial configurations, test layouts, and iterate on ideas—all before investing in professional design services.

Contracting: What does construction actually cost?

“Nobody really knows what construction costs,” says Christoph Berner, founder and managing director at Cosuno.

The issue isn’t missing data—Cosuno has processed €50 billion in construction volume, with every line item priced and billions of data points accumulated.

The issue is that this data speaks different languages. One contractor writes “three red windows.” Another writes “windows, three pieces, red.” A human eye can match these instantly, but scaling that matching across millions of line items? That was impossible—until Cosuno spent three years training AI on their proprietary dataset.

The result: 85% accuracy in matching equivalent positions across non-standardised formats. In practice, this means uploading a bill of quantities and receiving benchmark pricing at the position level within seconds. Users can search any line item against historical bids, or generate and optimise entire bills of quantities based on patterns the AI has learned.

Most recently, Cosuno launched an AI agent that works directly within the bill of quantities—reviewing documents, suggesting improvements, and applying changes with a single click. The data that always existed but remained trapped in incompatible formats is finally becoming usable intelligence.

The recently launched AI agent reviews documents directly within the bill of quantity and improvements are accepted with one click.

Construction: What’s actually happening on site?

Two very different approaches are tackling this question, but both share a common goal: generating structured data where none existed before.

TUM’s process monitoring takes the visual approach. Fabian Pfitzner installed 27 cameras across five Munich construction sites, each capturing images every 30 seconds. The result: 4 million images documenting construction sites progressing through time. But raw images accomplish nothing on their own.

Pfitzner’s breakthrough came from layering AI on top: computer vision identifies resources—formwork, silos, workers, equipment—and maps each to the BIM model. Each image becomes a node in a knowledge graph connecting building components, workers, equipment, and zones. Graph neural networks then analyse this structured data to recognise processes: loading concrete, moving materials, pouring.

What emerges are activity plots showing when sites start and stop—but more importantly, they reveal outliers. For example, on one day there was zero activity on site. Correlating that with weather data - it showed that due to high winds, the cranes were grounded. Combined with complexity metrics, weather data, and schedules, this feeds predictive models for task duration.

Versatile’s crane intelligence takes a different approach: IoT devices with load cells and sensors on cranes, generating 32,000 data points per day per job site.

“If you are in business of converting cash into assets, you probably want to use cranes,” says Nikolai Suvorov, Versatile’s Global Go-to-Market Leader.

So Versatile are using crane sensors to track the critical path progress and therefore to protect the tight margins of steel erectors and general contractors.

Versatile has been in production since 2019—“doing AI in construction before AI in construction was really hot”—serving Vinci, Eiffage, and Bouygues among many others.

The convergence pattern - digital twins are finally unleashed

What connects these technologies isn’t the AI techniques. It’s the data infrastructure they create.

This convergence also transforms what “digital twin” actually means in practice. For years, the concept promised much but delivered frustration. Owners pushed hard for digital twins; everyone else in the industry pushed back. “The data entry and data aggregation and update was really a fairly heavy effort,” observes Suvorov, recalling his experience in the US market. “At some point this effort yielded negative returns. Not everybody was very disciplined in updating the information.”

“One of the best advantages of AI is really to kill at some point this manual data entry and keeping this single source of truth up to date,” Hellin explains. “It’s very tedious to do so. When you have it, it’s very nice—you can do a lot of automation, have a lot of workflows that work very well. But to keep it up to date and make sure you can really rely on all of the information stored in your digital twin as a building’s state changes quite often is a very time-consuming and tedious process.”

The technologies profiled here attack this problem from multiple angles. Versatile’s crane sensors automatically update as-built status. TUM’s camera-based monitoring captures progress without manual input. syte maintains property twins at national scale precisely because AI handles the data aggregation that would be impossible manually.

Digital twins become practical when the data flows automatically rather than requiring human transcription.

There’s also the output problem. Even when digital twins existed, surfacing useful information required custom dashboards—and everyone wanted their own version. “It felt like you needed 15, 20, 30 different dashboards,” notes Nikolai Suvorov. “With the introduction of LLMs and other means of getting information from a huge data set exactly the way you want it—whether it’s numbers, a statement, or a custom dashboard—that is way simpler now.” AI doesn’t just make digital twins easier to maintain; it makes them easier to query.

The parallel to manufacturing’s productivity gains isn’t AI algorithms. It’s sensors everywhere. Structured data constantly flowing. Systems that improve from every interaction.

Construction is building that infrastructure now. AI is the means. The end is closing a 90-percentage-point productivity gap.

What these technologies mean for construction professionals

For project developers: Early-stage feasibility analysis can now happen in hours rather than weeks. sytes platform and TUM’s text-to-BIM research represent the near-term: faster site evaluation, spatial exploration before engaging architects.

For general contractors: Procurement intelligence is becoming a competitive advantage. Cosuno’s position-matching and pricing benchmarks create pricing transparency that previously required extensive experience and historical records.

For project managers: Field data capture is transitioning from manual to automatic. Whether through camera-based monitoring or IoT sensors, the trajectory is towards continuous, structured data about what’s actually happening on site.

For technology strategists: The interoperability question matters more than individual tools. As these systems mature, the ability to flow data between property analysis, design, procurement, financial management and controlling, and field monitoring will determine which organisations capture the full productivity potential.

Key statistics

• 0.4% — Construction’s compound annual productivity growth rate over 22 years
• 3% — Manufacturing’s compound annual productivity growth rate over the same period
• 10% vs ~100% — Net productivity gain in construction versus manufacturing
• 4 million — Images captured across 5 construction sites for TUM’s process monitoring research
• 27 — Cameras deployed per TUM research installation
• 32,000 — Daily data points generated per crane by Versatile’s IoT system
• 25+ million — German properties in Site’s digital twin database
• 85% — Position-matching accuracy of Cosuno’s tendering AI
• €50B+ — Construction volume processed through Cosuno’s platform

Expert Contributors

Which of these AI construction technologies are commercially available today?

All three are fully commercial: syte and Cosuno operate as SaaS platforms serving developers, banks, and contractors across Europe. Versatile has been in production since 2019, deployed with major general contractors including Vinci, Eiffage, and Bouygues. TUM’s research—process monitoring and text-to-BIM—represents frontier capabilities that will eventually be productised but currently require research partnerships.

Will AI replace architects and construction professionals?

Not according to the researchers building these systems. “We are very far away from having any kind of this system taking away the real job of an architect,” notes TUM’s Sylvain Hellin. Current AI tools are best suited for early-stage exploration by non-designers—enabling property developers to test spatial ideas before engaging professional architects—not replacing architectural expertise.

How long does it take to implement these technologies?

Implementation varies significantly by technology. syte and Cosuno are cloud platforms requiring only data integration— operational within weeks. Versatile requires IoT device installation on cranes but generates data automatically once deployed. TUM’s research represents the timeline challenge: “Getting a demo up and running is pretty fast,” Hellin explains. “Where it becomes much more tedious is to bring it at scale and take care of all the edge cases.” He draws the parallel to autonomous driving: “The first demo was 15 years ago. We still don’t have this in mass production.”

What makes European AI research in construction different from US approaches?

Several innovations highlighted—syte’s foundational model, TUM’s process twinning—represent European-developed AI systems rather than applications built on American foundation models like GPT-4 or Claude. Site specifically emphasises building proprietary AI: “We don’t wrap ChatGPT or something. It’s our own AI.” This approach won Germany’s AI award two years ago and enables processing at national scale across 25+ million properties.

What’s the business case for construction AI adoption?

The core business case is closing construction’s productivity gap with manufacturing—10% gains over 22 years versus nearly 100%. Each technology addresses a specific value driver: Site compresses weeks of feasibility analysis to one hour. Cosuno’s 85% position-matching accuracy eliminates manual bid comparison across thousands of line items. Versatile’s crane data enables contractors to build more realistic financial models by connecting design intent to field reality. The common thread: structured data where only manual reports existed before, enabling analysis, prediction, and optimisation that was previously impossible.