I sat down with Andrea on the Digital Construction Podcast for an episode he called 5 Bold Predictions for the Construction Industry. He’d had them in advance, which meant he came with counter-arguments rather than agreement. That made it a better conversation.
The five:
- The app store is going to die. Not the software — the model where every project team logs into a different point solution. Agentic desktops become the single entry point, and the value moves from the feature to the orchestration.
- AI increases collaboration rather than reducing headcount. The bottleneck in construction has never been how many people you have. It has always been coordination between them.
- Projects are graphs, not lists. We manage them as flat waterfalls and Kanban columns while the actual dependencies form a network. Mapping that network is where ontology — and usable data — starts.
- A new internal role is forming. Call it the forward deployed engineer: part translator, part HR manager for a team that happens to be made of agents.
- The architectural flip. If internal teams can build the vertical thing themselves, SaaS vendors get pushed down the stack toward generic horizontal infrastructure — and per-seat pricing stops making sense when the seats are agents.
Read back like that, they look like five separate bets. They’re not. They’re the same argument told five times.
Every one of them is about the process we refuse to look at, not the output we keep optimising.
Take the automation point. The industry gets paid for drawings, models and reports, so that’s where every efficiency effort lands. Faster labels. Automatic dimensions. Better sheet generation. Fine. But the drawing is the last twenty metres of a very long run, and nobody is looking at the first two kilometres. I used an analogy from a Trimble executive on the podcast: we’ve trained every player on the team to peak athletic condition, and the team still doesn’t play together. We have excellent automation at every individual step and no view of the workflow in its entirety.
The interesting thing about lowering the coding barrier isn’t that everyone can now write automation. It’s what the automation tells you. If you capture why all these small scripts get written, you stop seeing scripts and start seeing a map of where your process leaks. The automation is a symptom. Most firms treat it as the cure.
Andrea pushed back hard in two places, and both were fair.
First, the incumbents. Autodesk and Procore are not going to sit still while an agentic desktop routes around them. They own the system of record, and construction data is not desktop-friendly — the models are heavy, and everything is coupled to everything else. His counter was that this might only resolve if local compute gets powerful enough to orchestrate without the cloud layer.
Second, the risk. When a structural engineer or a fire engineer hands judgement to a generative tool, the failure mode isn’t an embarrassing output. It’s a slab. The risk aversion in this industry is structural, not cultural — which is why the answer isn’t more autonomy; it’s determinism. Construction is prescriptive enough that you can build the check into the loop: the agent iterates until a deterministic gate says the answer is correct. That capability exists now. Most of the industry is still evaluating AI against a chatbot they tried two years ago.
The data conversation follows the same shape. Every conference has the same people saying we need standards, better data, cleaner data. Nobody shows what they’re actually capturing or where they started. Teams already doing this — the ones on site building their own ontologies — didn’t wait for an industry standard. They mapped what they knew, named the fields, and started collecting. The gap isn’t a data standard. It’s that nobody publishes their starting point.
Which brings me back to the platform question, and the distinction I want to keep sharp. When I say everyone can now automate, I mean everyone can now build a prototype. Not a product. A prototype that solves one project’s problem, gets retweaked on the next one, and gets retweaked again. That repetition is the signal worth reading. If the same automation keeps getting rebuilt, the value isn’t in productising it — it’s in the workflow that keeps demanding it.
Five predictions, one underlying claim: the industry has spent a decade layering tools on top of a process it has never examined. AI is the first technology that makes examining it unavoidable, because it needs the process legible before it can do anything useful with it.
That’s not a prediction. That’s the bill coming due.