Construction AI: OpenAI's agents got cheaper and learned to use software, and the clients started asking for proof

Here's the release that actually mattered this week, and like last week's it wasn't a construction product at all. On 29 September OpenAI held its DevDay in San Francisco and made more than twenty announcements. Most of them you can skip. Two you can't.

The first is that its Agents API can now use software directly. Up to now an agent could read, answer and draft. Now it can operate an application, clicking through the screens to finish a task the way a person would. OpenAI paired that with GPT-6.1 Sol, a new model built for exactly this kind of agent work, and put it at roughly a fifth of the price of GPT-6 Astra (that pricing is OpenAI's own figure, so treat it as direction). The second is a deal with Amazon called Bedrock Managed Agents, which lets those agents run inside a customer's own AWS environment rather than on OpenAI's. That one sounds dull and it's the important one, because it's about where the work happens and whose machine holds your data while it does.

So why put a developer conference in a brief for builders. Because the agents that check a drawing against a spec, review a compliance pack or chase an RFI all sit on top of models like these, and almost none of them arrive as a thing you buy directly. They turn up inside the platforms you already open. When the model gets cheaper and learns to drive the software itself, the tool gets cheaper to run and reaches further into the job. It's plumbing, but it's the plumbing under a lot of what's landing on your desk. And there's a quieter point worth making plainly: the cheaper an agent gets, the more it rewards the AI-first platforms built around agents doing the actual work, the broad challenger construction software taking on the big incumbents, over a legacy product with a chatbot bolted on the side. What we've found is that the underlying price falling is rarely neutral. It helps whoever built for it.

Now the bit an editor might cut, because it's the bit that bites. When an agent can operate software on its own and can run in one of several clouds, "which model answered" and "whose computer ran it" stop being questions for the IT team and become questions for whoever signs the contract. It's a bit like sending a subbie onto site who can now let themselves in and use your tools. Usually fine, and you still want to know who it was and which gate they came through. So if your firm is starting to lean on agents inside its project tools, write down which model does which job and whose cloud it runs in, and make both things you can check rather than take on trust. The person who'll thank you is whoever has to explain, a year from now, why two checks on the same building were done by two different machines.

Down the road from all that, UK Construction Week was on at the NEC in Birmingham from 29 September to 1 October, and the Digitalisation and AI Stage is the part I'd have spent my time on. It's curated by Build in Digital and sponsored by Zutec, sitting inside more than 150 hours of CPD content across five stages. The line-up is a decent read on where the industry's head actually is this autumn.

What jumps out is the change in the questions. A year or two ago the sessions were about what AI could do. This week one of Wednesday's was called, without any hedging, "The Proof Clients Want: Turning Impact into Contracts", and another paired digital twins with the golden thread and Gateway 3 under the heading "Collaboration by Default". The speakers weren't vendors doing a pitch either. Colin Bell, the Digital and Data Director at Kier, was on the stage, as was Kathryn Donald, a partner and director at Max Fordham, alongside Dr Bola Abisogun OBE and Robert Barbour from Travis Perkins. That's people who carry the risk, not people who sell the licence.

Here's the aside worth a line. Pairing digital twins with Gateway 3 in the same session is a quiet admission of where the next jam is. We've spent recent briefs noting that the Building Safety Regulator finally sped Gateway 2 up, and that's real. But Gateway 3, the completion certificate at the end, is the gate nobody's solved yet, and a digital twin is only worth building if the record it holds is the one that gets you through that last door. What that means on site is that the golden thread stops being a filing exercise and becomes the thing that proves the building is what the drawings said it was. Get that right and the person keeping the record is no longer the one chasing paper at handover. They're the one holding the evidence when the client, or the regulator, finally asks for it. That's what it's about.

Put the week's two threads next to each other and they turn into one. The tools underneath are getting cheaper and are learning to work software on their own, and the clients paying the bills have stopped wanting a demo and started wanting proof written into the contract. Those pull in the same direction. An agent that can do more of the work only earns its keep if you can show what it did, to a standard that holds up when someone leans on it.

The professional bodies have already moved to that ground. The RICS professional standard that came into force on 9 March 2026 requires members using AI in surveying to keep human judgement and accountability in the loop, and the big contractors are framing their own rollouts the same way. Kier has said around 40 per cent of its people now use Microsoft 365 Copilot, under a banner it calls "Human at the Helm" (both reported on 11 September). Read those together and the operating model is clear enough: let the agent do the reading and the chasing, keep a named human answerable for the decision, and keep the record that proves who did what. None of that is exciting. All of it is what turns a clever tool into evidence a client will pay for.

For your board pack: pick one process where an AI agent already does real work in your business, and write the one-line answer to three questions: which model does it, whose cloud it runs in, and where the record of what it did is kept. If you can't answer all three, that's this quarter's job.