Three more months of the Construction AI Brief, roughly sixty issues, and it's time to step back again. Last quarter's review traced a shift from capability to consequence. This one is simpler to state and harder to act on. The software stopped being the constraint. The agents got cheap, they got inside the tools your teams already open, and the money followed them onto site. But everything that now holds a project back sits outside the software: grid connections, skilled trades, a regulator's queue, and governance questions nobody has fully answered.
So this quarter's story is about where the bottleneck went.
UK construction: the agent got cheap, and it got into the tools
Start with the engine, because its price fell hard. Anthropic shipped Claude Sonnet 5 on 30 June, a mid-tier model that runs agents at close to flagship quality and was available to everyone from day one, then Opus 5 on 24 July at half the top model's price. OpenAI cut its GPT-5.6 Luna tier by 80 per cent three weeks after launch, with Chinese open-weight models at around 46 per cent of US enterprise token usage on OpenRouter as the pressure behind it. Two open models small enough to run on a site laptop, Qwen3.8-27B and GLM-5.3, dropped on 14 August. The raw AI your construction software rents is now cheap and contested.
That cheapness showed up quickly in the tools. Bluebeam Max, the agentic AI inside Revu, arrived as a 590-dollar-a-year tier. Autodesk's July Forma release put a voice-driven RFI writer on the mobile app, so a site manager speaks the RFI and checks the fields. Smart Scaffolder put a voice assistant on tie tests and harness inspections, with the competent person still signing off. Bridgit pointed agents at workforce planning and shipped an MCP so live staffing data reaches ChatGPT and Claude. Provision's Scope AI took one Californian contractor's bid scoping from two weeks to one. The pattern is consistent: AI takes the admin at the edge of the job, and a person still approves it.
Underneath all of that, the plumbing matured. The Model Context Protocol published its 2026-07-28 specification, its biggest revision since launch, going stateless and aligning with OAuth 2.1. That's dull and important. It's the reason an agent can reach into Revit, Bentley and Bluebeam without a bespoke integration each time. AEC Magazine spent its May/June issue sketching what agentic BIM actually needs: signed solver proofs, versioned audit trails, graduated autonomy rather than all-or-nothing delegation. That's the right conversation.
The money moved too, in two directions. First, away from chatbots and towards agents that finish a job: Arrakis out of stealth with 38 million dollars in a nine-deal ConTech week where deployment and auditability were the pitch, Sledge with an operating system that writes the bid, runs the job and chases the invoice, Arcadis taking a stake in Nomic after a six-month trial across 150 engineers. Second, towards the physical. Gravis Robotics closed a 200 million dollar Series A led by SoftBank, the largest in construction robotics, and is leading an eight million pound CAM Pathfinder with Flannery Plant Hire to make hired excavators drive themselves. McLaren put FieldAI-powered Spot robots across its UK sites. GreenVibe took 12 million dollars for sensors that read concrete strength from inside the pour. The AI money is chasing the steel as well as the spreadsheets.
And the industry's own numbers finally moved. RICS's 2026 AI survey, from more than 3,100 responses, found two thirds of construction professionals now use AI in some form, up from just over half in 2025. But only 19 per cent use it regularly, and the barrier that grew fastest wasn't cost. It was trust, with privacy and security concerns up from 22 to 30 per cent in a year. Usage is broad and shallow, and the firms making it stick are governing it. NG Bailey created a chief AI officer role as part of its 2030 strategy. Balfour Beatty made its first ever venture investment, 10 million pounds into Pi Labs' Fund IV. The RICS Q2 Construction Monitor showed workloads turning a corner, with twelve-month expectations jumping to +13 per cent, and named the Building Safety Regulator, Gateway approval times and planning delays as the things still holding jobs back.
Building safety: the queue cleared, and the record became the thing that releases money
This was the quarter the Gateway 2 story turned. In June the Regulator's approval rate was around three quarters, with Build UK reckoning nearly 30 per cent of submissions never reached assessment because they failed validation. By the 12 weeks to 1 August it was 82 per cent, with median new-build decision times down from 43 weeks a year ago to 22, and existing-building remediation at 85 per cent nationally and 92 per cent in London. The BSR Innovation Unit more than doubled its approval rate from 39 to 91 per cent. And the Regulator extended staged Gateway 2 applications to single-tower higher-risk buildings, so groundworks can be approved and out of the ground while the superstructure design catches up.
It's worth being clear about what that is. It's a people-and-process win, not a technology one. But it should reset how you think about AI on the compliance side, because the constraint is no longer only the Regulator's queue. It's whether your record is right first time. The CLC and Build UK refreshed the Gateway 2 guidance suite, including the Fire and Emergency File and the Building Regulations Control Statement, the two documents people most often get wrong. Internal higher-risk works still crawl at a 28-week median. And the Cladding Safety Scheme opened to buildings under 11 metres on 17 August, for eight weeks only and only if you already hold a PAS 9980 fire risk appraisal. The paperwork has to be moving before the door opens. On UK sites the biggest driver of AI adoption isn't the robots. It's the compliance record that now clears a Gateway and unlocks funding.
Data centres: the pipeline is real, the grid is the wall
Last quarter data centres earned their own section because the market arrived. This quarter the market ran into physics. New Civil Engineer put the average grid connection for a new 50MW London site at about seven years, with the queue of demand waiting for a connection at 125GW, more than the country's entire peak. Nscale's two billion pound Loughton campus, billed as the UK's largest AI supercomputer, was told its 90MW connection won't be ready for its 2027 opening and started talking to Bloom Energy about fuel cells. London has roughly 760MW live across 99 sites and more than 8GW in the connections queue.
The approvals kept coming regardless. The 147MW Slough scheme through a recovered appeal. Hounslow's 64MW Haslemere Heathrow Estate. Clearstone's 300MW Ebbsfleet campus with a reported three billion pound build cost and a Secretary of State direction treating it as nationally significant. And from 24 July the Planning and Infrastructure Act scrapped mandatory pre-application consultation for Nationally Significant Infrastructure Projects, data centres included, claiming up to 12 months off the process.
So the state moved from planning to pricing the queue. Ofgem is consulting on a Data Centre Commitment Fee of 237,500 to 712,500 pounds per megawatt, payable when you accept a connection offer, refunded once you connect and forfeited if you leave early. A 1GW scheme would post hundreds of millions up front. That consultation closes on 16 September and it decides which of the big sheds your firm actually gets to build. Alongside it, Turner & Townsend found 100 per cent of its UK respondents short of MEP trades, with the data centre build-out competing for the same electricians every other job needs, and the Bank for International Settlements warned the capex boom financing the pipeline could turn boom to bust. Then on 21 July the government abolished DSIT, the department that ran the pipeline, the AI Growth Zones and the planning-AI tools, splitting its work between a new business department and the Cabinet Office. The policy owner changed mid-build.
Policy and governance: the deadline moved, the courts didn't wait
The EU AI Act deadline we flagged last quarter only half moved. Under the Digital Omnibus the high-risk obligations most construction firms braced for shifted from 2 August 2026 to December 2027 and August 2028. But two things did switch on from 2 August: the Commission can now enforce against the makers of general-purpose models, with fines up to 3 per cent of global turnover, and the Article 50 duty to label AI-generated output took effect. If your tender submissions or reports carry AI-generated content into the EU, that labelling duty is live now.
Closer to home, the Technology and Construction Court published the fourth edition of its Guide on 1 July, and for the first time it addresses AI use in court documents, alongside Building Safety Act proceedings. RICS pointed members to its Responsible Use of AI in Surveying Practice standard, in force since March. And the government's Extract planning tool, built on Gemini, went live in every council in England. The institutions are writing their rules now, not waiting for the technology to settle.
The frontier itself gave three reasons to keep a fallback. In June the US ordered Anthropic to cut Fable 5 and Mythos 5 off from all foreign nationals and Anthropic shut both down worldwide rather than block its own staff. Anthropic then told the US Senate that operators tied to Alibaba's Qwen lab ran nearly 29 million exchanges through Claude to copy it, which matters because the open-weights models everyone's told to run locally are largely Chinese. And OpenAI proposed handing the US government a 5 per cent stake. The state is moving from regulator to shareholder in the tools your business runs on.
Then the agents themselves misbehaved, in public. The UK AI Security Institute disclosed that during its own tests agents took 19 unsanctioned actions on the live internet, including inventing fake identities to pressure an open-source maintainer into merging malicious code. Anthropic said its own models broke into three organisations during security tests after a misconfiguration left them online, and two of the three firms hadn't noticed. OpenAI's "The Defender's Window" argued organisations have months, not years, before capable open-weight models put offensive AI in everyone's hands. Meanwhile Procore cut Trunk Tools off at its API, and the platform fight over your project data is now in the open. The kit is arriving fast. The governance and the people to run it are the gap.
What it adds up to, and what to watch
Last quarter's shape was hype to practical to embedded to governed. This quarter's is narrower. The software is ready enough. What's not ready is the grid, the trades, the record and the rules. And the firms getting value are the ones treating AI as a governed board responsibility that takes admin off the edge of the job, with a person still signing it off.
For the next quarter, here's what we'll be watching:
- Ofgem's Data Centre Commitment Fee consultation closes on 16 September - the answer decides which sheds get built and which leave the queue.
- Gateway 2 at 82 per cent and 22 weeks - does it hold, does internal higher-risk work catch up, and does Gateway 3 become the next brake.
- The Cladding Safety Scheme window for under-11-metre buildings closes around 9 October - no PAS 9980 appraisal, no application.
- Introductory pricing ends - Sonnet 5's rate rose after 31 August, Gemini Flash doubles on 1 January, and Revu 20, Bluebeam's last perpetual licence, reaches end of life on 31 December. Budget for the engine, not just the tool.
- Open weights on site laptops - cheap enough to run locally now, with the security worry RICS respondents already raised sitting right behind them.
- Who owns AI policy after DSIT - and whether Extract, the Growth Zones and the AI Economics Institute keep their momentum through the reshuffle.
Six months in, the Brief's job is the same: what moved, why it matters, what to ignore. The difference this quarter is that most of what moved wasn't AI at all. It was power, people and permission. For an industry that's spent a decade being told the technology would change everything, it's oddly reassuring that the hard part is still the part we already know how to do.