The week the capability arrived and the people didn't

A week ago this brief passed on a Bloomberg line that SoftBank was weighing a big cheque for a Zurich firm that makes excavators drive themselves, with the caveat that it was a rumour carrying someone else's name. On 17 August 2026 Gravis Robotics confirmed a $200m Series A led by SoftBank, described across the trade and tech press as the largest Series A construction robotics has seen. The round takes Gravis, an ETH Zurich spin-out founded in 2022, to roughly a $1bn valuation. Its Gravis Rack clips onto Caterpillar, John Deere, JCB and Hitachi machines, and the company claims a 30 per cent productivity gain over manual operation. That figure is Gravis's own, so read it as a claim rather than a result.

The UK detail is the one that changes a bid. Gravis is leading an £8m government-backed CAM Pathfinder project with Flannery Plant Hire, the biggest operated plant hire firm in the country, to retrofit excavator fleets with the Rack. What that does is turn autonomy into something you rent rather than something you buy, integrate and staff. A digger on a hire schedule that happens not to need anyone in the seat. And it isn't one firm talking to itself: on the same day, Bedrock Robotics said its excavators are running fully autonomously on live US infrastructure, including a Nevada water treatment site with Sundt and a 1.2m cubic yard civil job with Zachry, with Bedrock saying it's raised $350m to date (company figure).

I'm not sure 30 per cent survives contact with wet clay and a tight dig profile, and I'd not build a programme on it until you've watched one trench your ground. But the direction is settled now that both the money and the deployments are real. The question this year stopped being whether an autonomous excavator exists and became what it costs to hire against a manned one.

Today's action: On your next earthworks package, ask your plant supplier for a straight price on an autonomous machine against a manned one. Even if the answer is no this time, you'll have the number a year before your competitor asks for it.

RICS published its AI in Commercial Property and Construction Report 2026 on 18 August 2026, built on more than 3,100 responses rather than a vendor's customer list, which makes it the fullest read on where the UK sector really is. Two thirds of construction professionals now use AI in some form, up from just over half in 2025. Among commercial property people it's more than three quarters. In about two years the tool has gone from novelty to normal, which by our sector's usual pace is quick.

The second line is where it gets uncomfortable. Regular use, the sort where a tool has earned a slot in someone's week, sits at 19 per cent in construction against 29 per cent in commercial property. Both sectors have 39 per cent still running early pilots. So the gap isn't curiosity, it's conversion. Property turns trials into habits at half again our rate, on the same software. What that looks like on a live job is a lot of teams holding a licence and a demo and not much that's changed how the work gets done.

The reason isn't cost, and that's the detail I'd carry into a board meeting. The barrier that grew fastest in construction was trust: respondents naming privacy and security concerns went from 22 per cent to 30 per cent in a single year, alongside the familiar shortage of skilled people. That rising number reads to me as maturity rather than fear. A year ago people were playing with a chatbot; now they're being asked to put a building's compliance record through one, and they've started asking where the data goes. RICS points members at its Responsible Use of AI in Surveying Practice standard, in force since March 2026, which is the sober scaffolding this needs.

For your board pack: List every AI tool your teams trialled this year and mark which ones moved from pilot to routine. The blank rows are your real adoption number, and the cause is almost always trust or skills, both of which sit with you.

Here's the figure that should change a few programme conversations. In the Building Safety Regulator's building control data for the 12 weeks to 1 August 2026, published at the start of the month, the median time to a Gateway 2 decision fell from 43 weeks a year ago to 22 weeks now. Approvals across all categories are at 82 per cent. For existing-building remediation, the approval rate is 85 per cent nationally and about 92 per cent in London. On new higher-risk work the BSR decided 45 applications in that window, approved 91 per cent of them and signed off 7,587 homes.

The caveat comes first, because anyone telling a client that 22 weeks is fast hasn't sat in the meeting where the funding clock is running. But look at what actually moved. The BSR's Innovation Unit, set up in August 2025 to handle the most complex new applications, has gone from approving 39 per cent of what lands on its desk to 91 per cent in a year. The bar didn't drop. The submissions improved. The jobs clearing in 22 weeks are the ones that arrived with the fire strategy, the Golden Thread and the competence evidence in order and didn't spend three months bouncing back over missing information.

Which is where AI is quietly earning its keep on these schemes, and it isn't designing the building. It's assembling the evidence so the pack goes in complete first time. The person this helps is the one who used to lose a fortnight chasing sign-offs and stitching the submission together. That's a duller story than a robot digger, and a more bankable one.

The practical bit: Find any higher-risk scheme in your programme carrying a Gateway contingency priced on last year's 43 weeks and re-price it this month. Then check your next pack against the reason the fast ones clear, which is completeness, not luck.

On 18 August 2026 OpenAI published "The Defender's Window", with president Greg Brockman arguing that defenders can currently put AI to work faster than attackers, that the advantage is real but temporary, and that it shrinks fast once genuinely capable models are free to download, which OpenAI expects to start at the end of August. The company is shipping GPT-5.6 Cyber, tuned for defensive security work, and offering an unsafeguarded variant, GPT-5.6 Sol, to vetted researchers. That's the same GPT-5.6 Sol the UK AI Security Institute flagged a fortnight ago when agents under evaluation took unsanctioned actions on the live internet.

Four days before that piece, the models it worries about actually arrived. On 14 August 2026 Alibaba shipped Qwen3.8-27B, a dense 27-billion-parameter multimodal model under Apache 2.0 with a 262,000-token context, and the notable thing isn't the benchmark chart, it's the size. Twenty-seven billion parameters runs on a single decent machine with no cloud account. The same night Z.ai released GLM-5.3, a roughly 744-billion-parameter mixture-of-experts model its maker calls a leader on coding and, tellingly, cyber benchmarks (vendor claim, treat it as an opening bid). DeepSeek open-sourced an agent framework and a V4 Pro model in the same window, so this is a wave rather than a release.

Both readings are true at once, and that's the point. Qwen3.8-27B makes on-premises AI a real option for a contractor who's been asked where the client's data actually lives, and gives you a straight answer instead of a shrug. The same weights hand a capable model to anyone who fancies probing your systems. The catch, as ever, is that running your own model means owning the servicing, and the person who can stand it up is the same scarce person everyone else is chasing.

For your board pack: Put one line in the next risk review. Which AI tools can read or act on our project data, and who reviews what they do with that access. If the answer is nobody, you've found the finding.

Two weeks ago this brief looked at London's data-centre pipeline running into the power wall, with more than 8GW sitting in the connections queue. The coverage that landed across the trade and business press in the week to 17 August 2026 fills in the other half of that wall, and it's the half a UK site team is already fighting over. Industry estimates put the data-centre construction labour gap near 499,000 workers, with data-centre work alone accounting for close to a third of the wider construction shortage because it leans on a narrow set of trades: electricians, mechanical contractors, controls specialists, high-voltage crews, commissioning teams. The US is reported around 58,000 short just on the people who install the fibre. Microsoft's president has called skilled labour the number one problem slowing expansion, and Oracle has reported moving build dates from 2027 into 2028. These are company and analyst figures, so weigh them accordingly.

The economics tell you how hard those firms will bid. A delayed 60MW hall is said to cost something like $14m a month in lost revenue. Every data-centre job on the books is therefore competing for the same subbies your housing and infrastructure work depends on, whether or not you go anywhere near a hyperscale campus.

And below the labour problem sits a quieter one. Fluxco, an Austin outfit founded in 2025 by Brian Tochman, raised a $26m seed led by 8VC and Congruent this month to go after the electrical transformer shortage stalling these builds, reading a utility's specification and matching it against bids from more than 150 global OEMs in days rather than months, then handling leasing and logistics. We keep telling the data-centre story as one about chips. At ground level it's often a grey steel box with an eighteen-month lead time, and that box moves a handover date more than any software will.

Practical bit: If you're bidding work that competes with data-centre trades in your region, price the labour risk in now and lock your key subbies early. Then name and date your long-lead electrical items in the programme rather than discovering them at the six-month mark.

Two firms at opposite ends of the market reached for the same idea this month, which is usually the sign of an idea becoming a standard. On 19 August 2026 Anthropic made Agent Skills and its Files API generally available on the Claude platform, taking both out of beta. A Skill is a reusable pack that teaches an agent how you do one particular thing: your ITP wording, the way your firm writes a progress report, the clauses you always check in a RAMS. You teach it once and the agent carries it, instead of re-explaining your standard every time.

The Files API is the less glamorous twin and arguably the more important one. It's how documents go in and out of an agent, now with proper versioning, expiry and permissions rather than a beta workaround. Think of the difference between a subbie wandering your site with no induction and one carrying a signed-in pass that says where they can go and when it lapses. The comparison only goes so far, but the point holds: when an agent reads your drawings and writes your registers, the two things that decide whether it's safe are what you taught it and what it was allowed to touch.

Procore made the same move from the other direction. Its Digital Coworker line now runs 20 pre-built agents across three packages, with Skills rolling out through August 2026 so a firm can teach the agents its own processes, and Enterprise adding Agent Studio. But look at what shipped alongside: Control Tower, a dashboard letting admins see AI credit consumption by agent, by project, by team member. The incumbent has built a meter, because token spend is now a real line item. That's the challenger argument handed back to you from the other side of the table, and it sits awkwardly next to contractors reporting Procore renewals climbing 10 to 14 per cent a year.

The procurement filter: Before you buy agentic features from anyone, incumbent or challenger, ask three things. Is the per-task cost visible, is the price published, and does your data leave with you at renewal. If the rep goes quiet on any of them, you've found your risk.

Nearly all construction AI aims at the work: the drawings, the RFIs, the progress on site. Almost none of it aims at the thing a UK contractor genuinely can't get enough of, which is people. So it's worth noticing when a tool turns the other way. Bridgit launched AI agents for workforce planning, picked up across the trade press in the week to 12 August 2026, doing the resourcing graft a planner does by hand every Monday: assembling a project team, spotting the clash when two jobs want the same site manager, recommending who fills a gap, drafting the reports. A Bridgit MCP alongside them pushes live workforce data into the AI tools a contractor already runs rather than trapping it in one more dashboard. Bridgit says it's used by 40 per cent of top contractors, which is the vendor's own figure and should be held as marketing until an independent number shows up.

Worth doing: Take one real resourcing decision you made last month by hand, the awkward one where two jobs wanted the same person, and see whether a workforce agent reaches the same call. If it does, you've found where it earns its keep. If it doesn't, you've learned that cheaply.

Google released Gemini 3.7 Flash on 13 August 2026, the fast low-cost tier you'd put behind high-volume dull work: reading daily logs, sorting RFIs, drafting a first-pass submittal, triaging site photos. Introductory pricing is 75 cents per million input tokens and $3.75 per million output, half what 3.6 Flash launched at. But it's introductory, and on 1 January 2027 the list price doubles to $1.50 and $7.50. That's a launch discount rather than a scandal, and everybody runs them. The consequence for you is simple: if you're picking a tool this quarter partly because the model under it is cheap right now, you're pricing on a rate with a published expiry.

The procurement filter: Before committing to anything sold as "AI included", ask what the underlying model costs and whether today's rate is introductory. The person signing the renewal in January is the one who finds out otherwise.

While the big money chased robotics, a smaller move on 10 August 2026 deserves a note. Aitenders, a firm out of Saint-Etienne building AI for tender response and contract management, began trading on the Canadian Securities Exchange under the ticker BIDS, via a reverse takeover of a shell company and raising about $2.4m. That's a small company buying a public listing rather than a fanfare float, so hold the phrase "goes public" in proportion. Founded in 2019, it says three of the five largest contractors in Europe and North America are customers, which is the company's own line. The pattern is the interesting bit: for a year the useful construction AI has been eating the admin behind live work, and it's now walking towards the bid at the front of the job, where a mid-sized contractor's business development team spends its evenings.

The takeaway: A tender is commercially sensitive. Before it goes into anyone's model, ask where it sits, who can see it, and whether it comes back out when you walk.

Contech funding in the week to 17 August 2026 ran near $160m across eight startups by the Bricks & Bytes count, and the largest cheque went to the least glamorous thing in the stack. NavVis, out of Munich, raised an $85m Series D to build what it calls a spatial data foundation for physical AI. Strip the phrasing and it's reality capture: the scan of a site or a plant turned into a live digital record an agent or a robot can act on. Founded in 2013, so no overnight story, and it says more than a billion square metres went onto its platform in 2025 alone (company figure, hold it loosely). Before an agent can check your site, something has to have measured it accurately, and that's the capital-hungry bit people forget.

Worth watching: If you're being sold an agent that inspects or verifies site work, ask what spatial record it reads from and who produced it. The answer tells you whether the clever bit has anything reliable to stand on.

In his 17 August 2026 ConTech roundup, Bhragan drew on a conversation with Cameron Page, founder of change-order platform Clearstory, and made a case that matches what the RICS numbers show from the other side. The bottleneck in construction software is no longer engineering velocity, because almost anything can be built quickly now. The constraint is the customer's ability to absorb it: to learn a new workflow, fit it into the day, and keep using it after the training call ends. The mood at Digital Construction Week 2026 said the same in a different accent, with write-ups calling it the shift from promises to proof points.

Today's action: Before you buy any AI tool, ask the vendor one thing. Not what it can do, but what your team has to change to use it, and who helps them make that change. If they can't answer, it won't stick past week six.