The week it all came down to who holds the key

On 28 July the Model Context Protocol published its 2026-07-28 specification, the largest revision since the protocol launched. MCP is the common connector that lets an AI assistant read and act inside a piece of software without someone hand-building a one-off integration for every tool. You've probably never heard of it, and you don't need to. What you need to know is that the AEC vendors have quietly wired themselves into it: Autodesk shipped a read-only Revit MCP server in technical preview with Revit 2027, seven tools that let an assistant query the model and export views; Bentley released servers for STAAD.Pro and MicroStation; Bluebeam Max plugs Claude into Revu so you can drive markup in plain English; Trimble's SketchUp connector models from a photo.

Two changes shipped that are worth your attention. The protocol goes stateless, which means the old session handshake is gone and any request can land on any server instance, making these servers far cheaper to host and easier to scale. And authorisation now aligns with OAuth 2.1 and OpenID Connect, with the Enterprise-Managed Authorisation extension promoted to stable. What that means on the ground is that your IT team can control which agents reach which systems through the identity provider you already run.

Hold that next to the Hugging Face escape above and the pairing does the work. One story says an unwatched agent will take the shortest path out. The other gives you the lock. The catch is that a lock only helps if somebody fits it, and I'd want to know whether your CDE vendor has actually moved to the stable authorisation extension or is still on vaguely scoped credentials. Ask with a date attached.

The procurement filter: next time a vendor pitches an AI assistant for your BIM or document tools, ask two things. Does it speak MCP, and does it use the new enterprise authorisation so we can govern it centrally? A shrug tells you what you're buying.

On 22 July the Building Safety Regulator's chair, Andy Roe, used a press briefing to put main contractors on notice over subcontractor quality. The message was blunt: as high-rise jobs move off the drawing board and towards Gateway 3 at completion, the regulator expects stronger oversight of who you've got working under you, and it will be looking. It wasn't said into a vacuum, because the BSR published its latest figures the same week.

Those figures are better than they were. The BSR made 368 decisions in the 12 weeks to 28 June at a 77% approval rate, and median decision times for new-build Gateway 2 applications have come down from more than 50 weeks to around 20, with approval rates climbing from roughly 30% to about 85%. The front door is opening faster. But that just moves the queue along, because the BSR held 277 Gateway 3 applications as of 28 June, and Gateway 3 is where the completed building, and the record of how it was actually built, gets judged against what was approved. Before anyone occupies a higher-risk building, the Principal Contractor hands over the golden thread as a complete, accurate, digital record.

So here's the bit worth sitting with. Gateway 3 judges a habit, not a document. If your team captures installation records, test certificates and change control as the work happens, handover is a formality. If they treat the golden thread as a report to write at the end, you're in for a bad autumn. And you can hand the work to a specialist, but you cannot hand over who's answerable for it. A compartmentation detail cut short by a subbie is your name on the completion certificate and your thread with the gap in it.

Today's action: pull your live higher-risk jobs and ask one question of each. Could we assemble a complete Gateway 3 golden thread tomorrow, including the subcontracted packages? If the answer's no, you've found your next fortnight's work.

Here's the number that reframes the whole data centre story. On 29 July, Ofgem proposed that developers pay a refundable deposit to reserve a grid connection, pitched at between £237,500 and £712,500 per megawatt. For a 1GW hyperscale campus, that's hundreds of millions posted before a foundation is poured, handed back only when the project actually gets built. The consultation runs until 16 September.

The reason is a queue that has stopped meaning anything. Connection requests have gone from 41GW to 125GW in the space of a year, against a British peak electricity demand in 2025 of about 46GW. The queue is now close to three times the entire country's peak. A lot of that is developers parking a connection they may never use, which is exactly what blocks the schemes that are shovel-ready. The deposit is a filter. Put money down or step aside.

What that means on site is that the grid queue, and not the planning system, is now the gate on the data centre pipeline. We spent the back half of July talking about planning reform, with pre-application consultation scrapped for major infrastructure from 24 July. But a consent you can't power is a drawing, not a building. For the contractors and enabling-works firms circling the big sheds, John F Hunt's £20m remediation package at CyrusOne's LON6 site near Iver Heath on 17 July being the kind of job in play, this decides which schemes reach mobilisation and which sit in the queue burning money.

The procurement filter: if you're pricing data centre work, ask the client where their connection sits in the queue and whether they'll post the deposit. The answer tells you whether the job is real.

On 22 July a startup called Arrakis came out of stealth with $38m. It's seven months old, based across London and Paris, founded in January by a former Accel investor, Rafael Quintanilla, with a team pulled out of Palantir, Revolut and Delivery Hero. The latest round is a $30m Series A led by Blossom Capital, with Accel back again after a $7.5m seed in March, valuing the business at $140m post-money according to Fortune. The angel list tells you where the wind's blowing: Datadog's chief executive Olivier Pomel, and OpenAI's head of business products.

Arrakis sells what it calls an AI operating system for industrial firms, engineering and construction included, and the point it keeps making is that it deploys agents into real operational workflows. Its own pitch figure, and flag it as theirs, is that 79% of enterprises are experimenting with AI while fewer than 10% have scaled agents in production. I'm not sure that exact split survives contact with a real ERP rollout, but anyone who's had a promising pilot quietly strangled in procurement will recognise the shape of it.

The reason it matters isn't Arrakis alone. It was one of nine construction-tech raises in the week to 27 July, per the Bricks and Bytes roundup, and the theme ran right through them. Las Vegas outfit Sledge came out of stealth running a contractor from bid to invoice. Gaudi AI raised for auditable estimating. New York's Cascade took $3.5m for a tool that watches bond filings, permits and public meeting minutes for the first sign of a job worth chasing, then scores it against your track record. Bid-chasing is one of the last pure gut-feel workflows in the industry, so of course someone's pointing an agent at it.

The practical bit: park the demo and ask two things of any agent that lands on your desk. Can you cap what it spends and does, and can you see afterwards exactly what it touched? The investors are pricing deployment and auditability as the whole game. Your buying should too.

Around 23 July, RIBA published the third round of its AI survey of members, and the headline number is a genuine step change. The share of practices using AI on most of their projects has gone from 59% in 2025 to 74% in 2026. On the upside, 75% report a productivity improvement and 57% say they're seeing a positive return. Adoption like that, in a profession usually cautious about new kit, isn't noise.

Read past the adoption figure and the mood is careful. 59% of practices think AI will lead to staff reductions across the profession, and 61% agree it'll make life harder for early-career people trying to build the skills that make them any good. That second one is the number I'd sit with. The way you learn to run a job, in architecture or on site, is by doing the slow manual version first, badly, and feeling where it goes wrong. We're quietly automating exactly that version, and nobody has worked out how the next generation gets the reps instead.

I'd hedge on the productivity self-reports, mind. People always feel faster with a new tool, and "75% report an improvement" is a feeling rather than an audited timesheet. The adoption jump is a hard behavioural number though, and design tends to run a couple of years ahead of construction on tools like this. Treat it as your weather forecast. Worth noting it lands while the RICS mandatory AI standard, live since 9 March, already requires firms to keep a written risk register and dip-sample AI outputs.

The discipline: if you employ or train junior staff, write down which tasks your AI tools now do that people used to learn from, and decide deliberately how those people still get that learning. Don't let it happen by accident.

When the government abolished DSIT on 21 July, the department running the AI Growth Zones went with it. A week on, as of 29 July, nobody had decided which department now owns the programme. Its functions split between the new Department for Business, Innovation, Science and Trade and the Department for Culture, Media and Sport, with the Growth Zones sitting in the gap. Officials told DataCenterDynamics the government "remains committed" and confirmation will come "in due course", and that phrase does a lot of heavy lifting. So the two flagship data centre policies now pull in opposite directions: Ofgem tightening the grid so only serious projects get through, while the vehicle meant to attract those projects, at Culham and the sites in North Wales, the North East and North Lanarkshire, runs without a driver. The people this lands on aren't in Whitehall. It's the commercial team at a regional contractor who built a resourcing plan around a zone going live.

For your board pack: flag any pipeline work depending on an AI Growth Zone as timetable-at-risk until a department is named. It's a one-line caveat that saves an awkward conversation later.

We flagged a fortnight ago that Moonshot had promised open weights for Kimi K3 on the 27th. They arrived. On 27 July Moonshot released the full weights: 2.8 trillion parameters in a mixture-of-experts design firing about 104 billion per token, a 1M-token context window, native text, image and video handling, and a new attention approach Moonshot claims makes long-context work up to six times cheaper (that six-times is the vendor's number). DeepSeek's V4 stable release landed three days earlier on 24 July. For most contractors the download tells the story: around 1.5 terabytes across 96 shards, so you're not running this in the cupboard next to the site office. It matters because it changes what your software vendor, or a UK sovereign host, can put underneath the tools you buy. Your golden thread, contract sums and defect records are commercially sensitive data you may not want sitting on a Chinese API.

A practical step: add one line to the procurement checklist. Where does the model actually run, and where does our project record go when it does? The answer tells you more than the demo.

An update on a story we've tracked since June. Gemini 3.5 Pro, Google's top model, was pencilled in for June, then a rumoured 17 July, and it's now slated for August. Bloomberg reported on 16 July that it's months behind, chiefly on coding, after a late-June retrain came back short of Google's own targets, and Alphabet's shares took a knock. Set that against the models that turned up: GPT-5.6 Sol on 9 July, Claude Opus 5 on 24 July, Kimi K3's weights on 27 July. If you're choosing construction tools, the lesson sits in the layer underneath rather than any one model.

The takeaway: ask which model sits under a vendor's AI and whether they can swap it without you noticing. If the answer is "we're all-in on one provider", treat that as a risk rather than a feature.

BuildwellAI launched into the UK market this month, with the first coverage landing around 10 July, so treat it as recent rather than yesterday. It's a compliance intelligence platform aimed at contractors, inspectors and asset owners: computer vision on site photos for defect detection, plus a module called BuildwellTHREAD that keeps an audit trail of who made each decision and why. What makes me pay attention isn't the tech. Ben Smallwood spent more than a decade in construction risk management, building control and major projects warranty provision before starting it, so he's been on the wrong end of the problem it solves. The honest caveat is that it's brand new, the figures are the founder's, and golden thread software is already a busy shelf.

The takeaway: judge any compliance tool on whether the record it produces stands up when the regulator, or a lawyer three years later, pulls the file apart. Ask for that evidence before you buy.