The week the checking moved and the liability stayed put

On 7 September 2026, Arcadis announced a strategic financial stake in Nomic, a New York AI firm building agents that work across drawings, specifications, standards and a practice's own accumulated knowledge, together with a long-term commercial partnership. The stake followed a six-month trial. Around 150 Arcadis engineers, spread across disciplines, sectors and countries, ran the agents on the unglamorous core of delivery: drawing review against firm standards, code compliance, submittal review, RFI research and BIM coordination. Arcadis reports that 86 per cent of those engineers said it changed how they approached their work, and over a quarter called the shift fundamental. Those are Arcadis-reported figures, from its own staff, about a tool it now holds equity in, so treat them as direction rather than evidence.

The same week, Aurecon said it's scaling Nomic across its 6,700-strong workforce, after a four-year partnership that helped shape the platform in the first place. So this isn't one firm's punt. It's two of the largest engineering consultancies in the world committing inside a month of each other, with Arcadis having led the funding round last month and Aurecon following.

The detail that decides what this means for you is where the agent lives. Nomic doesn't set out to replace Autodesk Forma or Bentley ProjectWise. It runs inside them, and inside the other tools the teams already open every morning. So the money here is backing a checking layer that sits on top of the incumbents rather than a challenger to them. The most telling part, in my reading, is that both firms bought the layer rather than building it, after four years and six months of looking at it respectively. And it lands a question on the contractor's desk as much as the consultant's, because the drawing arrives on your site either way.

The procurement filter: next time a consultant in your chain tells you an AI agent reviewed the drawings, ask whose standards it checked against and who put their name to the output. The liability didn't move just because the checking did.

A quiet one slipped out under the noise of the autumn deadlines. From 1 September 2026, a slice of building work stops passing through Gateway 2 at all. The government confirmed it in Circular 02/2026, made using section 11 of the Building Act 1984, and it covers telecoms work: the installation of fibre optic cabling, and building work to mobile communications masts on rooftops.

In practice it works like this. For fibre optic cabling, the procedural requirement to obtain building control approval before starting work is dispensed with, and it applies to buildings inside the higher-risk regime and outside it. For rooftop masts the relief runs at Gateway 2 only for higher-risk buildings, and Gateway 3 still bites at the other end. The fibre carve-out is time-limited to three years, so it's a trial with a clock on it rather than a permanent hole in the regime.

I think this is the right sort of pruning, and I'd rather say so than hedge. Gateway 2's backlog has been the sector's slow puncture all year, and running a cable pull through the full higher-risk process helped nobody. But less red tape only helps the programme if the record still holds, and the dispensation is from the check, not from the golden thread. The catch is that a relaxed procedural gate tends to relax the paperwork habit behind it, which is exactly the wrong lesson to take from a three-year trial that somebody will review.

Practical bit: if you've telecoms work queued behind a Gateway 2 submission, check whether it now sits outside the door entirely from 1 September, and diary the three-year fibre window so nobody's caught out when the check returns.

The Building Safety Regulator's latest figures cover the rolling 12-week window to 31 August 2026, and the top line reads well. Eighty-four per cent of Gateway 2 decisions approved, up from 82 in the previous period. In numbers, 287 of 340 building control determinations passed, across 18,029 residential units. New higher-risk buildings and conversions did better still at 92 per cent, covering 9,362 units, and London took 31 of the 50 national new-build decisions and approved 28 of them across 4,477 units.

Then you reach the median determination time, and it's 22 weeks. The statutory milestone is eight. So the pass rate climbed and the wait didn't, which matters because on a live scheme the number that hurts was never whether you'd clear the gate, it was how many months you'd sit not knowing. Twenty-two weeks is the better part of half a year to find out you can start.

The regulator isn't pretending otherwise, and the honesty is useful. Since April it's run a dedicated external remediation team with dates attached: live caseload down to between 80 and 100 applications by 30 September, individual caseloads cut from around 25 to about 10, and average decision times under 12 weeks by December, with a 12-week response on non-complex cases by the end of March 2027. I'm not certain every one of those dates survives contact with the real caseload. The 30 September caseload figure is the one I'd watch, because it's the first to fall due and the hardest to spin.

Today's action: rebuild the pre-construction programme on your live higher-risk schemes around a 22-week Gateway 2 decision. If that breaks your dates, better to know now than in month five.

Here's the number I keep turning over, from McKinsey's State of AI 2026 survey reported in late August. Roughly a third of organisations, 32 per cent, said they'd decided against buying an off-the-shelf software product this year and built their own using agentic coding tools. Among the high performers, the firms attributing real profit to AI, it's nearly half. McKinsey also notes that the actual EBIT impact stayed flat at 37 per cent, so nobody should read this as a productivity miracle landing next quarter.

The direction matters more than the ROI line, because it's the mechanism behind the whole challenger category. What's happened is that the cost of building software has collapsed. A small team can now ship the breadth that used to take an incumbent's full product suite and the best part of a decade. That's why incumbent pricing is starting to look exposed. Procore's own filings show net revenue retention of around 114 per cent, which is a polite way of saying the average existing customer paid about 14 per cent more year on year, and contractors this renewal season are reporting rises of 12 to 15 per cent for software they use exactly as they did last year. Autodesk lifted list prices 6 to 9 per cent in January and began trimming renewal discounts. Those are as-reported figures, mostly customer and analyst accounts rather than vendor disclosure, but any commercial director who's signed a renewal lately will recognise the shape.

I'm not going to pretend a mid-size contractor is about to code its own project platform over a wet weekend. The analogy only stretches so far. But the same force that let a third of firms skip a purchase is the force letting founder-led teams cover a whole project at a published price you can walk away from. When the renewal lands 14 per cent higher, you've somewhere else to look, and that's new.

For your board pack: put your two biggest software renewals on one slide with their year-on-year increase, and name a credible AI-first alternative next to each. The exercise alone changes the next negotiation.

The 7 September ConTech round-up carried a story that isn't a product at all. Three people who built real audiences talking straight about construction technology have moved in-house. AJ Waters, who ran TheEngiNerdLife on the blunt question of why tech rollouts actually fail, joined Procore. Victor Muchiri of The Next Build went to Buildvision. Ala Abdelhamid, who runs Behind the Hard Hat, joined Ingenious Build and then moved to Outbuild. The reporting is Bhragan Paramanantham's in Last Week in ConTech, and he frames it as a funnel rather than three separate hires: a startup sponsors a creator, watches whether the audience actually buys, offers equity to keep them posting, and when the fit runs deep enough, offers a desk.

I don't think this is cynical, and I don't think it's clean either. A construction audience trusts a practitioner's take over a vendor's pitch every time, and the vendors have worked out you can't buy that trust one sponsored post at a time. So they're hiring it. Trust is the scarce material in this trade, the one thing nobody can A/B test, and the money has found it. What that means for you, reading LinkedIn on a Sunday night, is small but real. Keep following the people whose judgement you rate. Just carry the knowledge that the person telling you a tool is good might now be paid by the people who make it.

The discipline: when a source you rely on changes employer, don't drop them. Write the affiliation next to their name in your own head and weigh what they say the way you'd weigh a supplier's case study.

One for the diary. Innovate UK is opening a £32m competition, "Collaborative AI Solutions to improve productivity in key sectors", and construction is named as one of just four priority areas alongside transport, agriculture and the creative industries. The window opens on 27 September 2026, there's an online briefing on 29 September, and applications close on 8 November. It's open to firms of any size, universities, charities and the public sector, alone or as an R&D collaboration, and it sits under the government's BridgeAI programme.

So the money's real, and construction being singled out matters, because for years these productivity pots went to sectors that shout louder. What it means in practice is that you're bidding against transport and agriculture teams who've done this before. A competitive grant is work to win, not a windfall, and the firms that land them tend to be the ones with a bid writer already on the books, which is exactly the firms that need the money least. If you're a smaller contractor or consultant, collaborate: find a partner with the writing muscle and bring the site problem worth solving.

The quiet detail worth reading is that they want a productivity case, not an AI showcase. A clear before-and-after number on a real workflow, the sort a QS or a site manager would recognise, will read better than anything with the word "transformative" in it.

Worth doing: decide by 27 September whether you're bidding, and if you are, write the metric first and the technology second.

STACK shipped STACK IQ on 1 September 2026. What it does is let an estimator run the software by describing the job in plain English, building a takeoff from a spreadsheet, auditing an estimate, drafting a proposal, instead of clicking through screen after screen. It's wired to frontier models, with Claude and ChatGPT named, and it's available to every STACK customer on every subscription tier at no extra charge from launch. That's the vendor's own description, so treat the "does it all" gloss with the usual caution until you've run it over your own drawings.

The interesting bit isn't the chat box, because plenty of products bolted one of those on this year. What's different is that the conversation is pointed at real project data and carries out the task rather than answering a question about it. On a live estimate that's the difference between a demo and a tool. I doubt "just tell it what you want" holds up the first time a drawing set arrives half-labelled and out of scale, because a wall is never only its length. But the pricing choice is the tell. Free at every tier is a decision about where the pressure goes next, and it lands on every estimating product still charging a premium for the AI.

The practical bit: run it against a bid you've already priced by hand and compare the two line by line before you trust it on anything new.

The same 7 September ConTech round-up carried two smaller launches aimed squarely at the paperwork. Groundbook AI, out of California, shipped an AI plan reviewer that reads a PDF drawing set and checks it for conflicts against checklists, reports and building codes. Masaro, a Canadian outfit, put out a voice agent that takes a field update spoken aloud and adjusts the programme in Primavera P6 automatically.

Neither is a platform, and I'd not oversell them. Anyone who's kept a P6 programme honest through a wet October knows the hard part was never the typing, it was the truth of what people told you. A voice agent doesn't fix that. It removes the excuse for the update not making it in, which is worth something on its own.

A practical step: trial either one on a set you've already been through by hand, so you can count what it misses rather than what it catches.

In the 7 September contech funding round-up, SmartCraft, the Nordic construction software group, bought the Irish live cost-tracking tool LiveCosts for €8.45m. It's a small deal by the standards of the week, but it's the one that touches how you buy software, because LiveCosts is a job-cost tracker used across UK and Irish contractors and it's just become part of a bigger group's suite.

A tool that started life doing one job well now sits inside somebody else's roadmap. That's the ordinary rhythm of this market and it's neither good nor bad on its own. But the product you trialled on its own merits is now subject to a bundle, a new price list and a set of priorities that aren't yours.

For your board pack: keep a one-line note against every tool you depend on saying who owns it and what your exit looks like, and update it the week a deal like this lands.

The same week, Reframe Systems raised $40m for robotic homebuilding and Odyssey Energy Solutions closed a $74m round, both reported in the 7 September round-up. Neither lands on a UK site this year, and I'd not build a strategy on either. But the robotic homebuilding money is worth tracking against the UK's housing numbers, because the factories chasing that work will eventually quote against your traditional build.

The takeaway: file these as weather rather than news, and check back when one of them announces a UK facility.

RICS published its 2026 report on AI in commercial property and construction, and the finding is unfashionable and probably right. Optimism is high, but skills shortages and the slog of integrating tools into how a project actually runs are the real brakes on adoption, rather than enthusiasm or money.

That's worth holding next to the £32m grant window and the free STACK tier, because both of those are supply. A grant you win in November and a free feature you switch on tomorrow land in the same place, on a team that's already stretched. Money buys the software. It can't buy the fortnight a QS needs to trust it.

Worth doing: before you sign or apply for anything this autumn, name who owns the trial and block the hours out in their week. A tool with no owner is a tool nobody learns.

An update on a story we ran in week 35. Ofgem's consultation on a data centre commitment fee closes on 16 September 2026, so if you're anywhere near that pipeline this is the last fortnight to respond. The proposal would have a large-load data centre pay £237,500 to £712,500 per megawatt when it accepts a connection offer, refunded on energisation and forfeited if it exits the queue early. Ofgem reckons that's 2.5 to 7.5 per cent of scheme cost, which is the regulator's own framing rather than an independent figure. The number underneath is the one that explains it: connection applications jumped from 41GW to 125GW in under a year, at least 80GW of it data centres.

The procurement filter: before you commit bid resource to a data centre framework, ask whether the developer has paid to stay in the queue. A fee they'd forfeit is a cheaper signal of intent than any prospectus.