Construction AI: the Regulator lets you split the job in two, and the money buys the excavator not the paperwork

Here's the change most likely to move a programme this month, and it arrived without a press launch. The Building Safety Regulator has extended staged Gateway 2 applications to single-tower higher-risk buildings, the ones at least seven storeys or 18 metres tall. Until now that staged route was reserved for complex, multi-tower schemes. Reported across Building, Housing Today and Building Design in the week to 12 August 2026, the new guidance means a developer can separate groundworks and foundations from the rest of the build, get the substructure approved on its own, and start piling while the superstructure design is still being finalised and checked.

What that does on site is break the all-or-nothing approval that's been holding higher-risk work behind one enormous submission. Under the old practice the BSR effectively made you prove a full design couldn't be provided upfront before it would let you stage, a viability test that was never in the legislation and, as the lawyers have pointed out for a year, acted as a de facto bar on staged delivery. The 2023 Higher-Risk Buildings Procedures Regulations always allowed staged applications. The Regulator just wasn't taking them for single towers. Now it is. Charlie Pugsley, the BSR's interim chief executive, called the staged-applications guidance potentially "key" to how quickly building control approvals get processed.

I'd read this alongside the numbers from a fortnight ago, approvals up to 82 per cent in the twelve weeks to 1 August, because the two point the same way. The door's opening and the Regulator's now handing you a way to get through it in stages rather than waiting for one perfect pack. The caveat's worth stating plainly: each stage still needs its own Gateway 2 approval before that stage can start, so staging isn't a shortcut past scrutiny, it's a way to sequence it sensibly. Get the substructure evidence right and you're moving; get sloppy and you've just multiplied the number of times the Regulator says no. That's what it's about, the record still has to be clean, you just don't have to hold the whole thing back waiting on it.

Today's action: If you've a higher-risk scheme parked behind a full Gateway 2 submission, get your building control route to price a staged application on the groundworks this week. Weeks out of the ground earlier is the biggest single programme win on the table right now.

For a year the AI money in construction has gone into the office. Estimating, scope review, drawing checks, the paperwork about the digging. This week it pointed at the digging itself. Bloomberg reported in early August 2026 that SoftBank is weighing a deal that could value Gravis Robotics, a Zurich firm spun out of ETH, at more than $500m. Gravis doesn't build machines. It retrofits the ones you already run, bolting on a kit that fuses LiDAR, cameras, GNSS positioning and hydraulic sensors so a standard excavator can trench, grade soil and manage stockpiles on its own. SoftBank would fold the stake into Roze, the new AI-and-robotics vehicle it's assembling alongside its $5.4bn purchase of ABB's robotics arm.

Two caveats before anyone gets carried away. It's a rumour, unconfirmed, with the size and structure not settled, so treat the number as a marker not a fact. And an autonomous excavator on a live UK site runs into banksmen, exclusion zones, temporary works and a CDM regime that doesn't much care how clever the hydraulics are. The comparison only goes so far, but think of it like cruise control on a motorway: genuinely useful on the long straight repetitive stretch, still needs a human for the roundabout. Trenching a service run across an open plot is the motorway. Threading a grab between live services on a tight urban job is the roundabout.

Still, the signal's worth reading. When an investor of SoftBank's size starts buying the autonomy layer for plant rather than the admin layer for the office, it's saying the labour shortage on the tools is now a bigger prize than the labour shortage at the desk. What that means for the person running the machine is a slower, quieter version of what the estimators have already had: the AI does the repetitive grind, and the operator's job shifts towards supervising and handling the exceptions. Worth watching whose kit ends up bolted to your fleet, and what it records about your site while it works.

The procurement filter: Same question you'd ask of any software, asked of a machine. Whose sensors are on your excavator, who holds the data they capture, and can you take that machine and that record to another job or another supplier. Autonomy you can't unbolt is lock-in with tracks on.

Sometimes the best argument for challenger construction software, the broad AI-first platforms taking on the big incumbent suites, isn't a launch. It's an invoice. Two standing developments this year keep making the point from the other side. Autodesk moved its multi-user subscriptions to named-user pricing and scrapped renewal discounts from 7 January 2026, so the old network-licensing flexibility, where a seat could float between whoever needed CAD that afternoon, is gone and everything's tied to a named individual. And on the construction-management side, contractors keep reporting Procore renewals climbing 10 to 14 per cent a year. One firm, fifteen years a customer, put it bluntly: the first years were affordable, the last couple of multi-year deals were "outrageous". Analysts reckon a mid-market contractor pays $10,000 to $60,000 a year and uses maybe 30 per cent of what they're paying for.

Report that straight and it's just pricing. But read it as a builder deciding where next year's software budget goes, and it's the whole challenger thesis in one line: you're paying enterprise money for a suite you use a third of, on terms that get less flexible and more expensive every renewal. That's the gap the challengers are built into. The five tests are simple enough, AI doing real work at the core, broad by design rather than one task, usable by the person on site, published pricing you can see, and free to leave with your data. The incumbents fail the last two on purpose, because lock-in is the business model.

I'm not going to pretend a two-year-old platform matches a Procore feature list, it doesn't, and anyone who tells you otherwise is selling. But most firms don't need the feature list. They need the ten things they actually do, done well, at a price that doesn't ambush them. That's the essence of it. The renewal letter is the challenger's best salesman, and it turns up once a year whether you asked for it or not.

For your board pack: Before the next incumbent renewal, get someone to price two of the broad AI-first alternatives on a like-for-like basis, published pricing, data-export terms, notice period. Even if you stay, you'll renew from a stronger position knowing what leaving actually costs.

Put the three together and there's a through-line. The Regulator's just widened a door, staged Gateway 2 for single towers, so work gets moving sooner. SoftBank's cheque says the automation frontier is shifting from the office to the plant yard. And the incumbent renewal is quietly pushing firms to ask what their software actually costs and whether they could walk. All three land on the same person: the one deciding, this quarter, where to spend a stretched budget and where to place a bet on tools that'll still make sense in three years.

The standing discipline holds under all of it. Whether it's a staged evidence pack, a sensor kit on an excavator or a construction platform, the question that protects you is the same. Who holds the record, and can you take it with you. Ask it before you sign, not at practical completion.

A practical step: Pick your single biggest software or plant renewal in the next quarter and write down, today, what it would take to leave it. If you can't answer in a sentence, that's the lock-in, and it's worth knowing the size of before the renewal lands.