Construction AI: a global consultancy buys its way into an AI platform, and the MEP design bottleneck gets $50m of attention

Here's the one that matters this week. On 3 August, Arcadis, one of the biggest engineering and design consultancies in the world with a heavy UK footprint, announced it had taken a strategic stake in an AI platform called Nomic and signed a long-term commercial deal (AEC Magazine, 3 August). Terms weren't disclosed. What was disclosed is the trial behind it: six months, around 150 engineers, 12 countries, running Nomic's agents on real delivery work rather than a sandbox.

What Nomic does is sit inside the coordination end of engineering. Its agents handle drawing review against a firm's own standards, code compliance checks, submittal review, RFI research and BIM coordination, and they plug into tools Arcadis teams already run, including Autodesk Forma and Bentley ProjectWise (Arcadis news, 30 July). Arcadis says 86 per cent of the engineers who took part reckoned it changed how they approach their work, with over a quarter calling the shift fundamental, and one workflow dropping several days of document cataloguing down to hours. That's a consultancy-reported figure from people trialling a tool their employer went on to buy, so treat it as directional, not gospel. I'm not sure "86 per cent" survives contact with a wet Tuesday on a live job, but the direction is clearly right: the review-and-coordination layer is where AI earns its keep.

The bit an editor might cut, so I'll keep it in. When a firm the size of Arcadis takes a roadmap seat, the platform starts optimising for Arcadis. Nomic's CEO, Andriy Mulyar, put it well, that the aim is to get AI "inside real engineering work, not next to it", and that's the right target. But the smaller consultancy or contractor buying the same tool later inherits a product shaped around a giant's workflows and client base. That's the case the broad, AI-first challengers keep making from the other side: buy something you can shape and leave, not something already bent to someone else's shape. The comparison only goes so far, but it's the difference between renting a suit and buying one off a peg cut for a much bigger man.

The procurement filter: before you buy any "AI inside your engineering" tool, ask two things, what it does with your project data, and whether you can take that data and your configuration out if you leave. If the answer's vague, that's your answer.

The other side of the same coin. Endra, a Stockholm start-up building AI that generates coordinated MEP designs, calculations and documentation, confirmed a $50m Series A led by Andreessen Horowitz and, on 31 July, opened offices in New York, San Francisco and London (Construction Management, 31 July; AEC Magazine, 3 August). The London base is pitched as the hub for enterprise customers across Europe and the Middle East, and the company says it's already talking to some of the largest engineering firms. Total raised now sits around $75m, with roughly 30 enterprise clients on the books (vendor-reported).

What Endra is going after is the grind that Revit and AutoCAD leave you with: the weeks, sometimes months, of coordinating building-services layouts, running the calcs and producing the documentation. The claim is that its software does a coordinated first pass in minutes. If you've ever waited on an M&E package while the programme slips around you, you know exactly which bottleneck they mean. I'd not take "minutes" as your whole design cycle, more as a fast first draft that a competent engineer still has to check and sign, but building services is genuinely one of the slowest, most coordination-heavy parts of design, so it's a sensible place to aim. Andreessen Horowitz putting $50m into what one write-up cheerfully called the most boring part of building design tells you where smart money now thinks the returns are.

Two things to hold in your head. First, this is enterprise sales aimed at large firms, not a self-serve tool a small M&E contractor can switch on tomorrow. Second, a launch event in September, with a chess grandmaster on the bill, is marketing, not a delivered product, so judge it on what a real engineer gets out of it on a real job. For the person doing the MEP coordination, the honest question isn't whether the demo dazzles. It's whether the drawings still hold up on addendum three.

Today's action: if you run M&E design, line up one recently completed job you coordinated the old way and use it as a benchmark, so any tool you trial has a real number to beat rather than a slick demo to hide behind.

Put the two together and the week tells one story. The AI money in construction has stopped chasing the flashy edges, the site photos, the chatbot, the dashboard, and moved into the middle of the engineering: review, coordination, compliance, the documentation nobody enjoys. Arcadis bought into it. Andreessen Horowitz funded it. That's the same conclusion reached from two directions, and it's the right one, because the middle is where the hours and the risk actually live.

But there's a catch worth naming. The middle of the work is also the middle of your data. As these agents plug into Forma, ProjectWise, Revit and the rest, the question of who holds the project record, and whether you can take it with you, stops being an IT footnote and becomes a commercial one. This is exactly the gap the challengers, the broad AI-first platforms taking on the big incumbents, are built to fill: published pricing, self-serve start, your data walks out with you. Neither of this week's deals is one of those. Both are the incumbent-and-consultancy end of the market putting AI where the work is, which is good news, so long as you go in with eyes open about the lock-in.

That's what it's about. Not the percentage in the press release, but whether the person doing the coordination is better off, and still free to leave. Pick one painful workflow, pilot it against a job you already know the answer to, and check who owns the output before you scale it.

For your board pack: one line, "which of our AI tools can we exit inside 30 days with our data intact?" If nobody can answer it, that's the next thing to fix.