Construction AI: the regulator got quicker, and the housing starts followed

Here's a headline you don't see often at the moment. New-build starts in England came in at 35,910 in the second quarter of 2026, up 20 per cent on the same quarter last year and 6 per cent on the previous three months. Those are the government's April to June housing supply figures, reported across the trade press on 25 September. Over the rolling year to June, starts totalled 136,330, up about 15 per cent. Completions, at roughly 143,770, haven't moved much, so the pipeline is filling faster than it's emptying, which is the right way round after two flat years.

Now the bit you need before you repeat the number to anyone. A real part of that 20 per cent is the Building Safety Regulator changing how higher-risk building starts get into the statistics. The BSR reported 5,758 higher-risk starts in the quarter, against a previous quarterly peak of around 1,206, and those now make up roughly 16 per cent of all starts counted. So the jump is partly a genuine recovery and partly starts that simply weren't captured this way before. I'm not going to pretend the whole 20 per cent is fresh spades in the ground, because it isn't. But the direction is right, and the mechanism behind it is one we flagged only last week: the BSR's innovation unit brought the median Gateway 2 approval for new higher-risk buildings down to 22 weeks and lifted the pass rate to around 90 per cent. When the approvals move, the starts follow. You can now see that in two different datasets pointing the same way.

The comparison only goes so far, but it's a bit like a blocked drain finally clearing. Some of what comes through is the backlog that was stuck, and some of it is genuinely new flow, and from where you're standing it's hard to tell which is which on the day. What matters is that it's moving. And worth keeping honest: starts are still about 47 per cent below the exceptional peak of the second quarter of 2023, so this is a recovery off a low base, not a boom.

What that means for the person planning a residential programme is practical. The higher-risk approval timeline is no longer the automatic reason to sit a scheme on the shelf, and the numbers now back that up. Plan to the real 22-week median, watch your own caseworker, and if a client waves the headline growth figure at you, be the one in the room who knows the 20 per cent has two stories inside it. The people who get houses built are the ones who read the statistics properly, not the ones who quote them loudest.

For your board pack: put the starts figure in with one line of context, that BSR reporting changes flatter the year-on-year jump, so nobody builds a forecast on a number that's part real growth and part accounting.

UK Construction Week opens at the Birmingham NEC today, 29 September, and runs to 1 October. The reason it's in this brief rather than the events diary is the Digitalisation and AI Stage, curated this year by Build in Digital and sponsored by Zutec, one of five stages carrying more than 150 hours of CPD-accredited content across the three days.

Look at what's actually on it and you can see how far the conversation has moved. Tuesday runs "AI in Construction: Hype, Reality and the Next Five Years" and "From Adoption to Acceleration: Helping Supply Chains Embrace AI Together". Wednesday has "Collaboration by Default: Digital Twins, the Golden Thread and Gateway 3" and, the one I'd not miss, "The Proof Clients Want: Turning Impact into Contracts". The speaker list is a proper cross-section of the industry rather than a vendor parade: Colin Bell from Kier, Doug Johnson who co-founded Mesh, Dr Bola Abisogun OBE of the Digital Twin Skills Academy, Robert Barbour from Travis Perkins, Stephanie Finley from Procurement Hub, Kathryn Donald from Max Fordham, and Paul Indinger, the Building Radar chief executive.

What that programme tells you is the mood has changed. A year ago the AI sessions at these events were about what the technology might do. This year the titles are about proof and contracts, about the golden thread and Gateway 3, about embedding AI into a supply chain rather than admiring it. That's the sign of a sector past the demo stage and into the awkward middle bit, where you have to show a client the thing actually saved time or money before they'll pay for it. The RICS numbers from last month said the same, that firms are good at pilots and bad at making them routine. The stage titles have caught up with the problem.

So if you're going, and a lot of you will be, spend your two hours on the proof and the golden-thread sessions, not the crystal-ball ones. The person who comes back from Birmingham with a way to show a client the impact is worth more to the firm than the one who comes back with a tote bag and a headline stat. That's what it's about.

While Birmingham fills up, OpenAI holds its DevDay in San Francisco today, 29 September. It's not a construction event and most people reading this won't watch a line of it. But it belongs here for the same reason the Anthropic and Grok releases did over the past fortnight: DevDay is where OpenAI sets out its direction on agents, its developer tools and its pricing, and those decisions turn up inside the construction software your team already pays for, usually a few months later and without anyone telling you.

I'll be straight about what's known and what isn't. The date and the fact of the event are confirmed. What actually gets announced isn't, and I'm not going to guess. The pattern from the last two weeks is the useful guide: models have kept getting cheaper and better at the agent work, the reading, checking and chasing, that's slowly landing in the tools a quantity surveyor or project manager opens. If DevDay carries on that trend, the read-through for construction isn't a new gadget, it's that the AI features in your existing platforms get cheaper to run and a bit more capable, quietly.

The bit worth doing sits with whoever owns your software decisions, not your site teams. When the model underneath a tool can be swapped for a newer one with a couple of clicks, the question of which model sees your project data stops being theoretical. Decide your line on model choice and data residency, write it down, and make sure the people building on top of these platforms know it. That's the governance job hiding inside a developer conference you'll never attend.

The takeaway: watch what ships out of DevDay, not what gets promised on stage, and ask your software lead one question afterwards, whether anything announced changes which model touches your project data.

Put today's three items next to each other and they rhyme. The housing figures show a regime that got quicker and a supply chain that responded, once you strip out the counting change. The Birmingham stage has swapped "what AI can do" for "the proof clients want". And DevDay is a reminder that the tools keep improving underneath you whether you're paying attention or not. The common thread is proof. Not the promise of what a technology or a policy might deliver, but the evidence that it did.

That's the discipline to carry into the last quarter of the year. When a launch or a headline number lands on your desk, the useful question is the same every time: what actually changed on a real job, for a real person, that you can point to. The 20 per cent starts figure passes that test in part and fails it in part, which is exactly why you read the small print. A conference session called "turning impact into contracts" passes it by design. A model release passes it only when the tool built on top of it does something your team couldn't do last month.

A practical step: before the quarter closes, pick one AI thing your firm adopted this year and write down, in one line, the evidence it earned its place. If you can't, that's your answer, and it's a cheaper lesson to learn now than at renewal.