The week the calendar started charging

The Building Safety Levy comes into operation in England on 1 October 2026. It's a charge on new residential buildings, calculated per square metre of internal floor area, collected by local authorities, with the proceeds going towards fixing building safety defects across the country. The Building Safety Levy (England) Regulations 2025 that set out how it works were made on 19 November 2025, so nobody can claim it arrived without notice. The calendar has simply run out.

What makes it bite is the spread. The rate is weighted by local house prices and set authority by authority, and the gap between them is wide. Council-published rates run from roughly £13 to £16 a square metre somewhere like Burnley up to £100.35 a square metre in Kensington and Chelsea, with a 50 per cent discount for qualifying previously developed land. So the same block of flats carries a very different bill depending on the postcode, and on a decent-sized scheme that isn't a rounding error. Those are published council figures rather than gospel, and I'd check your own authority's schedule before you put a number in front of a board.

The cliff-edge underneath is the practical part. The levy doesn't apply to building control applications or initial notices submitted before 1 October 2026, including where they're later varied. The most telling detail, in my reading, is what that does to a piece of administrative history: the date an application went in stops being a diary note and becomes a priced fact, and proving it is a thing the industry has historically been poor at. The person who carries this is the commercial lead or the QS, the one who has to explain to a board why a scheme that looked fine in August costs more in October.

Worth doing: Pull every live residential scheme this week, mark the ones you can realistically lodge before 1 October, and put a levy line into the appraisal on the rest at your own authority's rate.

The amendment to Approved Document B that requires a second common staircase in blocks of flats with a top storey of 18m or more comes into force on 30 September 2026. It was published on 29 March 2024 with a deliberately long run-in so schemes already in flight had somewhere to go, and that run-in ends this month. The headline is straightforward and it's had plenty of coverage. The transition is where the money sits, and it's had far less.

To keep a scheme on the old single-stair rules, an application (a building notice, an initial notice or a full-plans building control application) has to be in before 30 September. And the work has to be sufficiently progressed, either before that day or within 18 months of it, so before 30 March 2028. What "sufficiently progressed" means on site is the bit people gloss over: permanent foundation pours or piling actually started, not a cleared plot and a programme that says you meant to. So there's a real scramble on to bank applications before the 30th, and a second judgement afterwards about whether the groundworks count. Get either wrong and the stair count changes, which changes the core, which changes the floorplate, which changes the appraisal.

Where AI earns its place here is unglamorous. Reading a drawing set against the new provisions to flag the misses before a building control officer does, and keeping a clean, dated record of which rule set a building sits under and why. Think of it like the service history on a car. Nobody enjoys keeping it, and the day you sell, the folder that proves what was done and when is worth more than the polish. That's what the golden thread is actually for, and this deadline is the moment it stops being a slide in a presentation and starts being the thing that decides whether a single-stair design is defensible two years from now.

For your board pack: Name one person who owns the "which rule set, proved how" record for every live tall-residential scheme, before 30 September rather than after it.

Construction News reported on 27 August 2026 that more than £100bn of data centre construction is working its way towards UK contractors. The detail underneath the headline is the part that should shape a bid strategy. More than a third of that total already holds planning permission. 55 schemes, worth about £12.7bn between them, have full detailed approval and have not broken ground. Another 22 sit on outline consent, adding roughly £24.5bn. That's a lot of consented work standing still.

For years we talked about planning as the blocker, the objection, the committee that says no. That number tells you the blocker moved. A data centre can clear planning and still wait, because what decides whether it reaches site is the grid: the connection offer, the substation, the transformer, the cooling and water, and the specialist electrical crews to fit it all out. The Electrical Contractors' Association made exactly this point on 13 August, welcoming another west London approval while warning that expansion mustn't outrun grid capacity or the skilled workforce to deliver it. The consent is the invitation. The connection is the door, and the queue for the door is where the programme really lives.

The government's answer is the Nationally Significant Infrastructure Projects route, which data centres formally entered in January 2026, with ministers directing three schemes into it so far at Wapseys Wood in Buckinghamshire, Ampthill Road in Bedford and New Barn Lane in Dartford. Pre-application consultation was stripped out in July, and the claim is up to 12 months off the timeline. I'd take a faster route over a slower one every time. But be clear-eyed about what it does: NSIP speeds up the deciding, not the connecting. It's a bit like getting a quicker answer on your mortgage while the house you want still hasn't been built. The comparison only goes so far, and the point survives it, a quicker yes doesn't move you in.

The procurement filter: On any data centre pursuit, put the grid connection date and its milestones at the top of the risk register, above planning, with a name against them.

The first week of September gave us four frontier model launches inside 72 hours, and the interesting part isn't which one topped a leaderboard. It's what happened to the price of running one. On 1 September 2026 Anthropic shipped Claude Fable 5.1 and cut the cost of cached-input reads by 75 per cent, from $1.00 to $0.25 per million tokens, while leaving base input and output pricing untouched at $10 and $50. The day after, Google put out Gemini 3.8 Flash, its third Flash model in six weeks, at exactly the same $0.75 and $3.75 per million tokens as the version three weeks before it. Two labs, the same move: hold the headline price, drop the cost of the repetitive work underneath.

That repetitive work is the bit that matters on a project. Fable 5.1's biggest jump wasn't in the flashy reasoning scores, it was in long-running, tool-using tasks: Terminal-Bench-Science went from 24.7 per cent to 52.6 per cent. In plain terms that's the drudge rather than the genius. Reading a full drawing set. Cross-checking a revision against the last issue. Chasing an RFI that's gone cold. A cheaper cache read makes it cheaper for the tool on your desk to re-read the same job over and over without the bill climbing, which is roughly your QS costing a quarter as much every time they open the same file for the fifth time.

Here's the catch, and it's the reason this belongs in a procurement conversation rather than a tech one. Google's Flash price is an introductory rate, and the small print says it doubles on 1 January 2027. So the cheap number you're quoted this autumn has an expiry printed on it, and the vendor selling you an annual licence may or may not have read it. These figures are the labs' own published prices, which is about as good as pricing data gets, but the pass-through to your invoice is entirely a matter of what your contract says.

The renewal filter: When your AI-first tool comes up for an annual deal, ask the vendor straight what happens to your price when their model provider's introductory window closes, and get the answer in writing rather than in a demo.

Both labs paired their general release with a cyber-capable model, and both put it behind a locked door. Anthropic's Mythos 5.1, announced on 1 September 2026, is the same underlying model as Fable 5.1 with safeguards loosened for cyber defence and life-sciences work, and access runs through trusted-access programmes with the US government, restricted to US organisations. Google did the parallel thing on 2 September with Gemini 3.8 Flash Cyber, handed to trusted defenders through its Fairwind Programme, which Google's own security team says wrote 2.6 times more correct Chrome vulnerability patches than rival models. Treat that 2.6x as a vendor's own number until someone independent runs the test.

What that means for a UK construction firm is a bit awkward. The most capable tool for defending the systems that hold your project information, your golden thread and your commercial data currently won't answer a British phone. You're not shut out of good security, and the defensive gains do filter through into the products you already run. But the frontier of it is gated by country, and I wouldn't expect that to change quickly. The plainer point sits underneath: the same capability that patches a vulnerability can find one, which is precisely why these got gated in the first place.

On a construction system that now has to hold a defensible record from Gateway 2 through to handover, knowing who and what can touch that record has stopped being an IT footnote. It's part of the safety case. And it's a question a main contractor's client will start asking at pre-qualification, whether or not anyone has an answer ready.

The takeaway: Stop asking "are we using AI" at board level and start asking "which model reads our project data, where does it run, and who vetted it".

Y Combinator's 2026 real estate and construction cohort runs to around 126 companies across the year, and its summer batch has Demo Day on 10 September 2026. Two clusters stand out. AI tools that strip time out of pre-construction, and agent platforms trying to run property operations without a human in the loop. Inside the first cluster, at least three separate teams are building AI estimating and takeoff, which is worth pausing on.

Names, so this isn't hand-waving. Rudus is an AI takeoff and estimating tool aimed at concrete contractors, claiming a 70 per cent cut in estimating time and three times as many projects won a year, both vendor figures I'd want to see survive a real tender before believing them. Foreman turns uploaded plans into takeoffs, estimates and proposals, replacing the spreadsheet-and-email patchwork. PLAN0 AI wants to be the Bloomberg of construction, reading drawings with vision models to price them. FlowManual bills itself as an all-in-one AI back office for the parts of a contractor's week nobody enjoys.

Now the precision, because the category only holds if we keep it honest. Most of these are narrow point tools right now, one trade or one task, and a single-trade takeoff isn't challenger construction software, the broad AI-first platforms taking on the big incumbents the way the challenger banks took the high street. FlowManual's all-in-one pitch leans closer to broad, and it's early enough that I wouldn't pin the label on any of them. What matters here is the pattern rather than any one name. When three funded teams pile into the same estimate in a single batch, it's because the incumbents' pricing and lock-in left a gap wide enough to walk through. That's the challenger effect arriving before any single challenger does, and it's why there's more room to push back at renewal than there was a year ago.

A practical step: Before you re-sign with an incumbent, name two AI-first alternatives out loud and book a demo of one. Even if you stay, you'll negotiate better knowing what else exists.

On 2 September 2026, UK Construction Week revealed the Digitalisation and AI Stage agenda for its Birmingham show at the NEC from 29 September to 1 October, curated by Build in Digital and sponsored by Zutec. The session titles tell you where the mood has moved. One is called "AI in Construction: Hype, Reality and the Next Five Years". Another, on the Tuesday, is "From Adoption to Acceleration: Helping Supply Chains Embrace AI Together". Read those two together and the shift is clear: the question has stopped being whether to use AI and become where it earns its keep, and where somebody is selling you a demo.

The Wednesday session is the quieter and more useful one. "Collaboration by Default: Digital Twins, the Golden Thread and Gateway 3" puts the data plumbing on the main stage rather than tucked into a side room. That matters, because a digital twin is only ever as good as the information feeding it, and most project information still sits scattered across drives nobody fully trusts. The line-up backs up the seriousness. Kier's digital and data director Colin Bell, Travis Perkins' data integration and insights director Robert Barbour, and Dr Bola Abisogun OBE are among the speakers. Not a robot-head keynote in sight, which is progress on its own.

There's a third one I'd point a commercial lead at, "The Proof Clients Want: Turning Impact into Contracts", because that's the session that decides budgets. It's about turning what you did into evidence a client will pay for. If you're going, don't wander the floor collecting branded pens. Pick three sessions, take your own project numbers, and make someone on a stand walk you through the actual workflow rather than the sizzle reel.

The practical bit: Book three UKCW sessions now and put your own project figures in a notebook to argue with, because the show sits on the same week as the 30 September and 1 October deadlines and your diary will fill.

CivilGrid raised a $26m Series A on 27 August 2026, led by Spark Capital, to build what its founder Josh Mackanic (a former PG&E engineer) calls a Google Maps for the underground. It pulls the scattered records together, utility assets, land ownership, environmental constraints, other people's works, into one map you can scope a job against at the desk before anyone is on site. The claim doing the rounds is that on one project it headed off around $6m of rework, which is a vendor-reported figure on a single case and worth a fistful of salt.

The direction is right even if the number isn't provable. Anyone who's had a service strike or turned up a Victorian sewer nobody mapped knows the cost isn't the survey, it's the day the whole gang stands idle while someone rings the statutory undertaker. It's a US product built on US data, so this isn't a tool you'll buy in Basingstoke next week. The shape of it, pulling buried-services risk forward to the desk stage, is where the groundworks headache lives over here too.

A practical step: On your current job, pull buried-services risk forward into the desk stage of the next package and see what you already own but haven't looked at.

ConstructConnect is adding Auto Scale and an AI-powered Auto Name to On-Screen Takeoff this month, finishing off the Takeoff Boost set it built on Google Cloud and launched in April 2026. The existing tools already do the measuring: auto takeoff, auto count and auto link, reading a drawing with computer vision and pulling off quantities. What's new is the grunt work that comes before any of that, setting the correct scale across a whole plan set and naming every sheet so a human can find things.

If you've never done a take-off that sounds trivial. It isn't. Anyone who's opened a fat PDF plan set knows the first job is the fiddly one, checking each sheet's scale and renaming forty drawings from "A-101" into something readable, before measuring a single wall. Computer vision reads a clean drawing well and a scruffy scan less well, so the time saved depends on what you feed it. It's a narrow point tool doing one job, and it's a job estimators genuinely hate, which is usually where a tool earns its place.

Practical bit: Run it on one real plan set this fortnight, a messy one, and time the pre-takeoff setup with and without it. That's the test that tells you whether it's worth the licence.

Togal.AI launched Assemblies on 29 August 2026. As an estimator takes off a drawing, the software builds the bill of materials alongside the measurement, material, labour and equipment together, rather than leaving someone to rebuild it in a spreadsheet afterwards. Drywall is the starting trade, and you can use Togal's templates or build your own assemblies to match how your firm actually prices.

Drywall first is a deliberate choice and a sensible one. A wall is never just its length. It's studs at a spacing, track top and bottom, one or two or three layers of board, a fire rating, sometimes insulation, sometimes a service void, and the labour to hang and finish all of it. Embed that logic and the takeoff stops being a measurement and becomes a priced scope. Like the ConstructConnect release, this is a point solution rather than challenger construction software, and the category only means something if we don't pin it on everything that ships with an estimate attached.

The discipline: When you trial it, price a job you already know the answer to, and check whether the assembly logic is yours to edit. If the templates can't bend to how you estimate, the saving evaporates on the second job.

The same Approved Document B package that brings the second staircase also carries design provisions for evacuation lifts, tighter travel-distance limits, and evacuation alert systems in residential buildings over 18m. That's a set of changes that touch the core, the services and the fire strategy together, not a single rule you can bolt on late.

The reason it's worth flagging separately is that the second-stair headline has taken all the attention, and a design team that has solved the stair and not the lift or the alert system is still going to have a conversation with building control. Read the RICS summary rather than a trade headline.

Today's action: Ask your fire engineer to confirm, in writing, which of the four changes each live scheme above 18m has been designed against.

Data centres formally entered the Nationally Significant Infrastructure Projects regime in January 2026, and ministers have since directed three schemes down that route: Wapseys Wood in Buckinghamshire, Ampthill Road in Bedford and New Barn Lane in Dartford. Pre-application consultation was stripped out in July, with the government claiming the wider reforms could cut planning timelines by up to 12 months and save developers £1bn this Parliament. Those are government-reported figures, not independently audited.

What it changes for a contractor is where the fight happens. NSIP moves the argument out of the local planning committee and into a national process, which suits a developer with the resources to run it and does nothing at all about the connection queue. Know which of the schemes you're chasing has been directed in before you bank a timeline off the back of it.

The procurement filter: For each data centre pursuit, record the consenting route alongside the connection status, because the two now tell you different things about whether the job is real.

The FMB and CIOB State of Trade survey, the twice-yearly read on SME builders, had 72 per cent of firms hit by skilled labour shortages in its most recent edition, up from 61 per cent six months earlier, with mechanical and electrical trades among the tightest. Those are exactly the crews that £100bn of data centre fit-out is going to want.

So the useful question for a stretched contractor isn't whether AI builds the shed. It's whether AI can take enough of the admin off your scarce, skilled people that they spend their week building rather than filling in forms. The grid will clear in its own time. The labour won't get less scarce.

The takeaway: Pick the one weekly task that eats the most of your best estimator's or PM's time and isn't judgement, just admin, and trial a tool against that on a real job this week.