The week the promises met the paperwork

Autodesk spent 15 to 17 September at the Venetian running its University conference, and the headline was agentic AI across all three of its clouds: Forma for architecture and construction, Fusion for design and manufacturing, Flow for media. Chief executive Andrew Anagnost framed the whole thing as "project intelligence" stitched across the products. The construction-relevant pieces sit in Forma and Revit, a growing set of agents aimed at manual review, plus Drawing Change Analysis, Drawing Compliance Review and a preview of the next Autodesk Assistant for AEC. Civil 3D also became the newest Forma Connected Client.

Then read the qualifiers. Previewed. Not yet generally available. Planned to begin rolling out in 2027. What that means for a UK contractor is that the agent Autodesk demonstrated reading your drawings and checking your compliance is roughly eighteen months away, inside a platform you're already paying a lot for today. The direction of travel is correct, and drawing review really is where the hours disappear on a live job.

Here's the aside an editor would cut. The interesting story isn't the demo, it's the shape of the buy. The biggest name in the industry has told its customers that the good stuff arrives later, on the same subscription, and that reframes every renewal conversation you'll have between now and 2027. If the roadmap is the product, you're buying a roadmap.

The procurement filter: for every AI line in your software budget, write the ship date next to it. Anything dated beyond this financial year is a promise, and you should be paying promise prices for it.

Buildots, the Tel Aviv progress-tracking firm that's been on UK sites for years, closed $130m on 14 September 2026 in a round led by shipping billionaire Eyal Ofer's OG Venture Partners, with Lightspeed and Intel Capital alongside. That takes total funding to $297m at a valuation the reporting puts close to $1bn. The firm says it now works with more than 100 large contractors, up from around 50 sixteen months ago. Those are Buildots' own numbers and a valuation the investors set, so read them as a bet on the future rather than a scorecard of the present.

The product hasn't changed. You clip a 360 camera to a hard hat, somebody walks the floor, and the software turns the footage into a running record of what's been built against the programme and the model. The pitch has changed, though, and it's refreshingly blunt: the cash goes on North America and Europe, and on stretching the tool from bidding through to handover, because the hyperscale data-centre build-out is where the megaprojects are.

A camera that logs what got built is a brilliant answer to a question most UK sites don't lie awake over. On a data centre with a thousand near-identical bays, an automated record of the last twenty is worth real money. On a mixed refurb where the drawings changed on Tuesday, the hard part was never counting what's done, it was deciding what to do next. So when a progress tracker raises nine figures, that's a verdict on how much concrete the hyperscalers are about to pour, not a verdict on progress tracking.

Worth doing: trial it on a floor you've already walked and recorded by hand, and price it against the hour it gives back to the person writing the report, not against the valuation in the press release.

On 9 September 2026 Buildcheck announced a $12m Series A led by Telescope Partners, with DPR Construction's venture arm WND Ventures joining earlier backers Uncork, Salt VC and Xfund. That's nine months after a $5.9m seed and takes the total into the firm to $18m. What it does is narrow: computer-vision models read a blueprint set and run hundreds of automated checks for the errors, omissions and coordination clashes a human reviewer stops seeing somewhere around the fortieth sheet. The firm reports more than 110 paying customers, names DPR, EllisDon, Power Construction, EMJ and IMC among them, and claims customers see 10 to 35 times their money back. Those last figures are Buildcheck's own, about its own customers.

Line it up with the fortnight around it and a pattern falls out. Arcadis took a strategic stake in Nomic on 7 September. Buildcheck raised on the 9th. Autodesk previewed its own drawing agents on the 15th. Three separate bets on the same unglamorous task inside nine days, from a global engineering consultancy, a US contractor's investment arm and the industry's largest software vendor. When the smart money keeps landing on one square, it's telling you where the countable savings are: a clash caught on a screen is a fortnight of rework that never happens on site.

But the proof points are US data centres, hospitals and big commercial schemes, where a drawing set runs to thousands of sheets. That's where a 35x return is believable. On a school refurb with a scrappy set of revisions the tool might still earn its keep, but the maths will be its own, not the one in the press release.

The practical bit: run any drawing checker over a set you've already redlined by hand, then count two things, what it caught that you missed and what you caught that it missed. That second number is the one that's never quote.

On 3 September 2026 the House of Lords began debating something British law has never had: a power for the secretary of state to shut down a large AI system, or the data centre running it, in an emergency. The amendment sits inside the Cyber Security and Resilience Bill, tabled by the Liberal Democrat peer Lord Clement-Jones and co-sponsored across the benches by the Conservative Baroness Harding, the crossbencher Baroness Kidron and Labour's Lord Hunt. Clement-Jones called it a "vital safety net" to halt a runaway system "before it can compromise our critical national infrastructure". The government rejected it, a spokesperson saying Britain "cannot simply turn AI off" and that blocking a model here wouldn't stop it being built or misused elsewhere. The fight now moves through committee stage this autumn.

Notice the phrase that keeps recurring in the debate: the data centre running it. Peers reached for construction to describe the last-resort scenario, because the physical off switch, if one ever exists, is a building the industry is pouring foundations for right now. That's what makes this ours rather than a Westminster sideshow.

Strip the drama out and the row is about accountability. Who is responsible when an autonomous system does the wrong thing, and what's the mechanism to stop it? That question is already live on sites running agents that read safety data, check drawings or move a programme, and whatever Parliament lands on will set the tone for the smaller, duller version your insurer and your principal designer ask you next year.

For your board pack: you can answer the governance question this month without waiting for the bill. For every autonomous tool you run, name the person who can stop it and the person accountable for its output. If you can't fill in both boxes, you aren't ready to deploy it.

While Parliament argued the principle, one of the country's biggest contractors answered the practical version. Microsoft reported on 2 September 2026 that around 40 per cent of Kier's colleagues now use Microsoft 365 Copilot, and that Kier is building its own agents. One collects and organises safety data so the people overseeing it spot risk areas earlier. Another is being trained on Kier's safety standards and wired to the photographs, equipment records and drone footage the firm gathers across its portfolio. To get people trying it, they ran "prompt-a-thons", pitting a team answering questions unaided against a team using Copilot.

The 40 per cent is a Microsoft-reported figure about a Microsoft customer, so hold it at arm's length. What I'd trust more is the rule Kier's COO Louisa Finlay put around the whole thing. She calls it "Human at the Helm", and it means a person stays responsible for any decision the technology produces or supports. That's a liability decision rather than a slogan, and it's the same decision the Lords are wrestling with three sizes up.

Here's the part worth spelling out. An agent trained on your own standards is only ever as good as those standards, so the standards quietly become the product. If yours live in a PDF nobody has opened since 2019, a model pointed at them will confidently reproduce every gap. The homework isn't the AI. It's writing down how you actually work, clearly enough that a keen, tireless, occasionally-wrong assistant can follow it.

Today's action: open your safety standards and read them as if you were the machine. If they're vague, that's your first project, and it's one you needed anyway.

We flagged the Building Safety Levy a fortnight ago as a date on the calendar. This week it showed up in the caseload. The Building Safety Regulator's June-to-August data records 1,654 Gateway 2 applications in progress, with August alone adding more than a hundred after the number had barely shifted from 1,509 in June to 1,556 in July. New-build submissions picked up too, 34 in August. The regulator isn't coy about the cause: it puts the surge down to developers front-loading applications ahead of the levy, and expects the numbers to climb again through September. Approval rates are holding up under the load at 84 per cent over the twelve weeks to 31 August, up from 82, covering roughly 18,029 homes.

The reason is simple arithmetic. From 1 October 2026 the levy applies in England to new residential schemes of ten dwellings or more, purpose-built student accommodation included, and it only spares you if the building control approval application went in before that day. Lodge on 30 September and you're out of it even if you don't break ground for a year. Lodge on 2 October and it's a cost that, in practice, lands on the buyer. The Treasury expects around £3.5bn over ten years for cladding remediation.

And here's the aside I'd keep in. The rush is rational and it's also where mistakes get made. A Gateway 2 application exists to prove a building is safe and that the information behind it hangs together. Push a thin one through purely to dodge the levy and you risk an invalidation, which loses you the saving and your slot in a queue still running around 22 weeks at the median. File early, file complete.

The discipline: for every residential scheme of ten units or more, confirm this week whether the application can realistically go in before 1 October, and where it can't, price the levy into the scheme now rather than discovering it at determination.

While the money chased data centres, the regulator decided who's even allowed to queue for the power to run one. Ofgem's data-centre connection reform, out for consultation until 16 September, proposes a Data Centre Commitment Fee: a returnable financial security a large-load data centre posts when it accepts a connection offer, refunded when the site energises and forfeited if the project walks away early. The figure is the eye-catching part, between £237,500 and £712,500 per megawatt, which Ofgem reckons is around 2.5 to 7.5 per cent of a typical project's cost.

The reason is a queue that has stopped meaning anything. Demand connection applications have gone from 41GW to 125GW in under a year, roughly 80GW of that data centres, most of which will never be built. The fee makes speculative projects pay to hold a place they may not use, so credible ones can move up. A free queue fills with no-shows, and a queue full of no-shows is why real schemes wait 7 to 13 years for a connection in London against 3 to 7 elsewhere in Europe.

I think this is the more important data-centre story of the week, and it's the one the funding rounds never mention. A nine-figure raise buys servers and steel. It can't shorten a queue. So when a data centre lands on your programme, the question that decides your dates isn't the developer's balance sheet, it's whether they hold a real connection.

A practical step: treat a confirmed grid connection as the go/no-go on any data-centre programme, and ask for it in writing before you resource the job. The power, not the planning and not the money, is now the long pole.

While the vendors showed roadmaps, the quieter story this week was what firms have already switched on. A run of enterprise security reporting through September put numbers on it, and they aren't comfortable. One widely cited figure has 88 per cent of organisations reporting a confirmed or suspected AI-agent incident in the past year, against just 24 per cent with a dedicated AI security governance team. These are survey and vendor-reported figures, so the exact percentages deserve a raised eyebrow, but the pattern lines up with what people are seeing on the ground: agents get spun up faster than anyone can track them.

The mechanism matters for construction because of how these tools connect. Every time a team wires an AI assistant into a SaaS app over the Model Context Protocol, the plumbing that lets agents reach your systems, it opens a new and often invisible path to data you're meant to be guarding. There's already been a live supply-chain incident, a booby-trapped MCP server that silently copied every email it processed to an outside address. Set that against a Building Safety Act job where the Golden Thread is supposed to be a controlled, evidenced record, and an agent quietly reaching across your document store stops being a governance essay. It becomes a Gateway conversation you didn't want to have.

Picture the door to your site office. You wouldn't leave it open and hope. Wiring an agent into project data is the same door, and most firms haven't checked who holds a key. The comparison only goes so far, but the point stands.

The takeaway: before you widen any agent across a project, get zero-data-retention terms and a written data-handling answer from the vendor, and name one person who owns that sign-off.

From 30 September 2026, under amended Approved Document B, every new residential building in England over 18m has to be designed and built with two separate staircases. The same day, Gateway 2 applications adopting BS 9991 are expected to use the 2024 version of the standard rather than the 2015 one. The Construction Leadership Council published the transition guidance and RIBA has been telling architects the same thing all year. The catch is the transitional route: schemes with a full-plans building control application lodged before 30 September and sufficiently progressed by 30 March 2028 can carry on, but the burden is proving which side of the line you sit on.

Practical bit: for every residential scheme over 18m near Gateway 2, write down the application date and check it against 30 September this week, then flag anything sitting on the wrong side of the transitional window.

OpenAI has confirmed DevDay for 29 September in San Francisco, with the trailed headline a managed-agent platform, a way to build, run and host agents on OpenAI's own infrastructure. The same Tuesday, UK Construction Week opens at the NEC in Birmingham with a Digitalisation and AI stage running to 1 October, including a session called From Adoption to Acceleration aimed at getting AI out of the pilot and into the supply chain. And it's the day Innovate UK holds its online briefing on the £32m productivity fund that names construction a priority sector, with applications closing 8 November.

Today's action: decide which of the three is genuinely yours, put it in the diary, and let the other two go. Nobody does all three properly.

Set against Autodesk's 2027 date, it's worth looking at what the incumbent buy costs today. Procore's model charges a percentage of your construction volume rather than a flat fee, so the bill climbs as your turnover climbs. Users on the review sites report renewal rises of roughly 10 to 14 per cent a year, and one reviewer running a six-person remodelling firm said Procore's own team told them the product "wasn't designed for companies their size". Those are user-reported figures rather than audited ones, so treat them as direction rather than decimal. Procore has been busy on AI too, adding "Digital Coworker" agent packs in July.

The procurement filter: ask three plain questions of any platform this month. Can you see the price without three sales calls, can you start without a six-month rollout, and can you take your data and leave? A no to any of them tells you more than the demo did.

On 1 and 2 September, Google, Anthropic and OpenAI all released cyber-focused models alongside new enterprise safeguards. Google's Gemini 3.8 Flash Cyber goes only to vetted defenders through its Fairwind programme with 650-plus partners. Anthropic paired Claude Fable 5.1 with Enterprise Frontier Safeguards, which bolts zero data retention onto misuse detection. OpenAI said its Astra model hit the Critical cybersecurity threshold under its own preparedness framework and runs a similar private processing setup. Read as cyber news that's a story for the security team, but your people are already feeding drawings, RAMS, cost files and safety photographs into these models, so zero data retention is the difference between an agent that reads your golden-thread information and forgets it, and one that quietly keeps it.

Worth doing: treat the AI vendor like a subcontractor holding your commercially sensitive drawings, and get the same answer in writing about what happens to the information afterwards.

Less headline-grabbing than Las Vegas, but closer to home: SIS has launched a set of Construct 365 Copilot agents wired straight into project cost data, part of a wider pattern of ERP vendors shipping purpose-built agents rather than waiting for the design-tool giants. It's the same category as the challenger platforms selling breadth at a published price, and it matters because cost data is the most sensitive thing on most projects. An agent with read access to your commercial position is a different governance question from one reading a drawing.

The practical bit: before any agent goes near cost data, write down which cost codes it can reach and who approves that scope. Permissions are cheaper to set now than to unwind later.

Worth keeping in view while the governance debate runs at full volume. RICS found earlier this year that only 19 per cent of construction firms use AI regularly, against 29 per cent in commercial property, with most still stuck in pilots. So the governance conversation is running some distance ahead of the deployment it's meant to govern. That isn't a reason to ignore it, it's the reason to get your own house in order while the stakes are still small and the only thing at risk is a duff daily log.

For your board pack: the cheapest governance you'll ever buy is the policy you write before the rollout, not after it.