Construction AI: the money stops backing cleverer chatbots and starts backing agents that finish the job, and the regulator turns to the subbies

Here's the deal that caught my eye. On 22 July a startup called Arrakis came out of stealth with $38m. It's seven months old, based across London and Paris, founded in January by a former Accel investor, Rafael Quintanilla, with a team pulled out of Palantir, Revolut and Delivery Hero. The latest round is a $30m Series A led by Blossom Capital, with Accel back again after leading a $7.5m seed in March, valuing the thing at $140m post-money according to Fortune. The angel list tells you where the wind's blowing: the Datadog CEO Olivier Pomel, and OpenAI's head of business products.

What Arrakis sells is what it calls an AI operating system for industrial firms, engineering and construction included, and the point it keeps making is that it deploys agents into real operational workflows rather than shipping another chatbot. Its own pitch figure, and flag it as theirs because it is, is that 79% of enterprises are experimenting with AI while fewer than 10% have actually scaled agents in production. I'm not sure that exact split survives contact with a real ERP rollout, but the direction is right and anyone who's had a promising pilot quietly strangled in procurement will recognise the shape of it.

The reason it matters isn't Arrakis on its own. It's that Arrakis was one of nine construction-tech raises in the week to 27 July, per the Bricks & Bytes roundup, and the theme ran right through them. Las Vegas outfit Sledge came out of stealth running a contractor from bid to invoice. Gaudi AI raised for auditable estimating. New York's Cascade took $3.5m for a tool that watches bond filings, permits and public meeting minutes for the first sign of a job worth chasing, then scores it against your track record. Bid-chasing is one of the last pure gut-feel workflows in the industry, so of course someone's pointing an agent at it. What every one of these has in common is that the product is the doing and the record, not the talking.

The procurement filter: next time an agent lands on your desk, park the demo and ask two things. Can you cap what it spends and does, and can you see afterwards exactly what it touched? The investors are now pricing deployment and auditability as the whole game. Your buying should too.

On 22 July the Building Safety Regulator's chair, Andy Roe, used a press briefing to put main contractors on notice over subcontractor quality. The message was blunt: as high-rise jobs move off the drawing board and towards Gateway 3 at completion, the regulator expects stronger oversight of who you've got doing the work under you, and it will be looking. This is the same week the BSR published its latest figures, so it wasn't said into a vacuum.

The numbers underneath it are genuinely better than they were. The BSR made 368 decisions in the 12 weeks to 28 June at a 77% approval rate, and median decision times for new-build Gateway 2 applications have come down from more than 50 weeks to around 20, with approval rates climbing from roughly 30% to about 85% and some parts of the system past 90%. So the front door is opening faster. But the regulator has 277 Gateway 3 applications in hand as of 28 June, and Gateway 3 is where the completed building, and the record of how it was actually built, gets judged against what was approved.

So here's the bit worth sitting with. You can hand the work to a subcontractor. You cannot hand over who's answerable for it. If a specialist firm cuts a corner on a compartmentation detail, that's your name on the completion certificate and your golden thread with the gap in it. The agent-auditability story in the section above and this one are the same story told from opposite ends. Something did the work. Can you prove it was done right and who did it.

For your board pack: tighten subcontractor approval and inspection records now, on live jobs, not when the Gateway 3 file is being assembled. The completed thread is a record you build as you go or don't have at all.

While the venture money did its thing, Moonshot AI released the full weights of Kimi K3 on 27 July. The headline specs are big: 2.8 trillion parameters in a mixture-of-experts design that fires only about 104 billion per token, a 1M-token context window, native handling of text, images and video, and a new attention trick Moonshot claims makes long-context work up to six times cheaper (that six-times is the vendor's number). DeepSeek's V4 stable release landed three days earlier, on 24 July, so late July has been a heavy week for open-weight models you can download and run yourself.

Now, the comparison only goes so far, but think of the frontier chatbots as electricity you rent from the grid, and open weights as a generator you can put on your own site. For most contractors the download alone tells the story: Kimi K3 is around 1.5 terabytes across 96 shards, and firing 104 billion parameters a token needs serious hardware. You're not running this in the server cupboard next to the site office.

So why does it matter to you at all. Because it changes what your software vendor, or a UK sovereign host, can put underneath the tools you buy. The golden thread, your contract sums, your defect records, that's commercially sensitive and safety-critical data you may not want leaving the country or sitting on a Chinese API. Open weights that a trusted provider runs on UK soil is one way that data stays where you can point to it. What that means for the person doing the compliance paperwork is a simple question to add to the procurement list: where does the model actually run, and where does our project record go when it does.

Practical bit: you don't need to understand mixture-of-experts to ask a vendor where their AI runs and who can see your data. Add it to the checklist. The answer tells you more than the demo.

Three items, one thread. The investors are backing agents that finish the job and leave an audit trail. The regulator is telling contractors they own the quality of work they've subcontracted, and it'll ask for proof. And open-weight models raise the plain question of where your data sits while the AI does its thing. None of that is about cleverness. All of it is about accountability and evidence.

That's the standing discipline, and it hasn't changed since RICS made it mandatory in March: keep a written record of where AI touched the work, dip-sample the outputs, and be able to say who or what did each part of the job. The tools are getting genuinely good at the doing. The bit that protects the PM, the site manager and the person assembling the Gateway file is still the record.

Today's action: pick one live job and check you could answer a single question on it today, who did this work and where's the proof. If the answer takes more than a phone call, that's your gap.