The week control became the question

On 12 June 2026 the US government, citing national-security authorities, issued an export-control directive suspending all access to Claude Fable 5 and Claude Mythos 5 by any foreign national, inside or outside the United States, and including Anthropic's own foreign-national employees. Anthropic disabled both models worldwide rather than block a vast slice of its users and staff. Every other Claude model (Opus, Sonnet, Haiku and the rest) keeps running. Anthropic's own statement is that the order traces to a narrow jailbreak of Fable 5 and that the capability shown is already available in other publicly deployed models, including GPT-5.5. We covered Fable 5 on 11 June as the document-reasoning model best suited to golden-thread and compliance work. A UK practitioner is a foreign national to a US directive. If you'd started leaning on Fable 5, you lost it overnight with no notice and no appeal.

The most telling detail, in my reading, is that Anthropic took the reputational hit of a global shutdown rather than carve out its foreign staff. That tells you how blunt these orders are and how little room a vendor has to protect you when one lands.

For your board pack: Add one line to your AI risk register, "single-provider exposure", and name the workflows that would stop if your primary model were withdrawn. If the list isn't short, that's the project.

On 12 June 2026 Engineering News-Record ran a piece with a blunt headline: construction platforms are already fighting over data to train AI agents. The reporting is specific. Procore pulled Trunk Tools (an agentic-AI provider used by US general contractors including Gilbane and Suffolk) off its API last September, refunded its booth at Procore's Groundbreak conference, and then on 20 January bought rival agent platform Datagrid. Procore's public line is that it's protecting "the integrity and security of all our customers' data". The prize, ENR notes, is geometry, getting AI to actually understand the spatial relationships in a model, and whoever holds the most project data has the best shot at training one.

This isn't a US-only curiosity. SAP made the same move in its API Policy v4/2026 restricting autonomous AI from chaining SAP API calls. If you run a cloud platform as your ISO 19650 common data environment, the question of whether your own (or a third party's) AI agents can read that data through the API is a commercial term, not a technical detail. McCarthy's multi-year Palantir deal on 4 June for an in-house "Pulse" platform is the expensive answer. Reading your contract is the cheap one.

The procurement filter: Before any CDE or platform renewal, find the two clauses that matter, who owns the data and whether your own (or a third party's) AI agents can read it through the API. Silence is the vendor's option, not yours.

At the Google Cloud Summit in London on 17 June 2026, MHCLG and DSIT confirmed that Extract, an AI tool the government's own i.AI incubator built on Google's Gemini foundation models, is now available to every local planning authority in England after trials across 20-plus councils. It turns scanned historic maps and planning documents into standardised data. The government says a two-hour task drops to about two minutes, saving an average council around 255 hours a year (these figures are government-reported, not independently audited). Fifty authorities are already using it, with more than 1,000 documents processed, mostly Tree Preservation Orders.

The more interesting tool is the one still in the lab. Augmented Planning Decisions (APD), built with Google Cloud, Google DeepMind and the AI firm Faculty, is in alpha with Barnet, Dorset and Camden. The stated target: halve householder application times from eight weeks to four, with national rollout pencilled in for 2027. That timeline will slip (these always do), and "helping officers reason about policy" is a far harder thing than digitising a TPO. But the direction is unmistakable. The body that determines your applications is now reaching for the same tools you are.

Today's action: Audit how your planning submissions actually arrive at the council, file formats, naming, whether your drawings and documents are machine-readable. The authority's tooling is changing; your inputs should too.

On 10 June 2026 housing secretary Steve Reed approved the 147MW Manor Farm data centre at Poyle Road in Slough, a 74-acre, part-greenfield, part-brownfield site brought forward by developer Manor Farm Propco. This came through a recovered appeal, the route where central government pulls the decision up to the Secretary of State, and Reed concluded the benefits of the scheme were "collectively sufficient" to outweigh the harms of building on the land. Property Week reported the approval at 448,000 sq ft, with Tritax in the frame on the investment side. Granted permissions hit roughly 2.8GW by April 2026, with the pipeline rising to about 6.2GW once unconsented schemes are counted (House of Commons Library figures).

One recovered appeal isn't a policy, and a part-greenfield approval will draw the usual (and legitimate) objections about precedent. But if you're advising clients or contracting on data-centre schemes, the reasoning in that decision letter is the ground you'll be fighting on. Read it before the planning balance gets quoted back at you in a committee room.

Worth doing: Pull the Manor Farm decision letter and lift the planning-balance reasoning into your data-centre pitch template now. The "benefits outweigh harms" framing is the one winning consents.

On 2 August 2026 the EU AI Act's obligations for high-risk AI systems become generally applicable. The Act applies extraterritorially: if an AI output your firm produces ends up being used in the EU, or you supply an AI-enabled service into an EU market, you can be in scope regardless of where you sit.

Browne Jacobson's 2026 construction horizon-scan, published as part of its sector series, makes the contractual point that should worry every commercial director. Standard-form construction contracts have not kept pace with AI adoption. There's real uncertainty about how existing liability provisions apply to AI-generated outputs, and about the rights in the underlying data, who owns it, what licence terms attach, and who carries the can when an AI-generated design or document turns out to be wrong. JCT and NEC weren't written for a world where an agent drafts your RFI response. Most amendments haven't caught up either. A lot of AI use on live UK projects right now sits in a contractual grey zone.

For your board pack: Map where AI touches anything that could be called high-risk, safety-critical design, automated decisions about people, anything exported into the EU. Get the data and IP ownership clauses in your live contracts reviewed before August, not after the first dispute.

On 11 June 2026, at London Tech Week's AI Summit, AI minister Kanishka Narayan MP announced the RIBA x DSIT Data Centre Design Challenge, the UK's first government-backed design competition of its kind, run jointly by the Department for Science, Innovation and Technology and the Royal Institute of British Architects. It's open to architects, designers and engineers, with communities pulled into the framing. The brief asks entrants to reimagine data centres as "places of genuine civic value", buildings people can be proud of, in the government's words, rather than ones they merely tolerate.

Read it in context. After a 147MW Slough shed got approved on a part-greenfield appeal the day before, and Ravenscraig steelworks went in the week before, a national design competition is, in part, the cultural answer to the political problem. So be clear-eyed: it's a design-ideas competition, not a planning requirement, with full eligibility rules, timeline and judging criteria still to be confirmed. Treat it as influence rather than regulation. But influence at the design-brief stage is exactly where a typology gets set, and most data centres are currently being drawn to a logistics-shed standard because nobody has asked for better.

The practical bit: If your practice touches the data-centre pipeline, watch for the full competition brief from RIBA and put an entry in. The design standard set here will become the language planners reach for to push back on the next bad shed.

In the first week of June 2026, Stockholm-based Endra closed a $50m Series A led by Andreessen Horowitz, with continued backing from Notion Capital and Norrsken VC (Axios broke the round on 1 June; figures are vendor-and-investor-reported). Endra is building an AI platform for mechanical, electrical and plumbing design, the coordination work that decides whether a building's services actually fit together before anyone gets on site. Last week's brief flagged M&E labour as the genuine bottleneck on the UK data-centre and defence pipeline. A tier-one venture firm has now put $50m behind software that attacks exactly that coordination problem. When the money stops chasing site cameras and starts chasing services design, that tells you where experienced people think the margin pain actually is.

A practical step: If MEP coordination is where your projects bleed, ask the new entrants one question, show me a clash set resolved on a live model, not a demo. Then measure the hours saved before you commit.

On 10 June 2026, US construction QA startup Structured AI raised a $4.2m seed round led by FCVC, taking total funding to $5m. Y Combinator, 20VC, Cherry Ventures, Zero Prime Ventures, Transpose Platform and Sequoia Scout are all in (reported figures). The product uses optical-recognition and computer-vision models to analyse completed fieldwork against the construction documents, and runs QA checks across entire drawing sets before an engineer reviews them. The founders describe it as spell-check-style: it recommends changes, a human approves them, and it plugs into Revit. Syska Hennessy Group has been co-developing it, with MEP coordination as the focus. Put it next to the Endra round and last week's Bluebeam-mbue deal and three separate bets land in the same fortnight on the same workflow.

The procurement filter: Before you buy any drawing-review AI, hand the vendor one of your genuinely difficult, mid-revision drawing packs and ask it to find the clashes you already know about. If it misses the ones your team caught, you've learned something cheap.

Renewable-energy developer Apatura submitted plans this fortnight to turn the former Ravenscraig steelworks (160 acres of brownfield land about 20km east of Glasgow) into one of the UK's largest green AI data centres, paired with battery energy storage and around 550MW of grid connections expected by 2030 (Apatura's figures; the project is reported at roughly £3.9bn). One route fights greenfield land through the appeals system on the argument that national compute needs outweigh local harm. The other reuses derelict heavy-industry land next to spare renewable capacity, sidestepping a chunk of the green-belt and grid-constraint argument before it starts. The catch in Scotland, as ever, is grid connection timing. "Expected by 2030" is doing real work in that sentence.

Worth doing: If you chase data-centre work, keep two pipelines in view, contested greenfield schemes through appeals, and brownfield-plus-renewables sites like Ravenscraig. The second is where the planning path is quietly easier.

On 16 June 2026 Anthropic shipped an update to Claude Code that's more interesting for what it signals than what it adds to the changelog. Two changes stand out. Sub-agents can now spawn their own sub-agents, hierarchical, self-directing chains of AI doing work without a human at every step. And there's a new "safe mode" that isolates a broken or misbehaving configuration so you can troubleshoot it, sitting alongside stronger permission rules and tighter admin controls. More autonomy and more containment, announced together. This is the layer underneath the agents your software vendors will wire into your project data through MCP. The fact that Anthropic shipped the brakes in the same release as the engine tells you the people building this know the supervision problem is real.

The procurement filter: Ask any agent vendor for three things, a full audit trail, granular permissions and a hard stop. Anthropic just built all three into its own tool; accept nothing less in yours.

Google announced Gemini 3.5 Pro at I/O on 19 May 2026, then told the audience to wait a month for access. As of mid-June the model is in limited Vertex preview with general availability expected this month. Reporting on 6 June put the headline specs at a 2-million-token context window and a "Deep Think" reasoning mode. Two million tokens is roughly enough to hold an entire O&M manual, a full set of specifications, or a stack of fire-safety documentation in a single context, and reason across all of it at once rather than chunking it and hoping retrieval picks the right page. That's the shape of a golden-thread query. Whether Gemini 3.5 Pro is actually better at that than the models already shipping is unproven until general release. A big context window isn't the same as good judgement about what's in it.

A practical step: When Pro lands, benchmark it on a completed project's O&M pack with known gaps. If it finds the gaps you planted, it's ready for a real one.

On 17 June 2026, alongside the planning AI news at the Google Cloud Summit, Deloitte and Google Cloud announced a London AI Studio sited at Deloitte's London campus. The explicit job is to move British organisations past AI experimentation into deploying "autonomous, action-oriented" agentic systems at scale. Google framed the whole Summit around the shift "from AI potential to agentic reality" and named a clutch of new UK partnerships alongside. The big advisory and delivery firms are where a lot of UK infrastructure and public-sector programmes get shaped, costed and run. When Deloitte stands up a dedicated agentic-AI studio with a hyperscaler, the agents that come out will end up inside the programme controls, cost models and assurance processes that sit over schemes you're delivering.

The takeaway: Before you sign up to any "agentic" delivery or platform pitch this year, ask two things, what the agent is allowed to do unsupervised, and how it's protected from prompt injection on the documents it reads. If the vendor can't answer both crisply, you've found your answer.

The 52nd G7 summit ran in Évian from 15 to 17 June 2026, hosted by France. AI was on the agenda, but the trajectory reported across the week was that the governance language would dilute compared with prior years, the focus pivoting to economic upside, with the US resisting multilateral commitment. All three frontier-lab chiefs (Sam Altman, Dario Amodei, Demis Hassabis) were reported in attendance, said to be the first time the rivals have appeared together before G7 leaders. The final communiqué language was still emerging as of 17 June, so the detail is provisional. Either way, the missing global rulebook is the story. If the G7 won't harmonise, the rules that bind a UK construction firm stay exactly where they sit today, the UK's own emerging regime and the EU AI Act on 2 August. A softer G7 line removes the excuse to wait for someone to standardise it for you.

For your board pack: Fragmentation means more compliance surfaces to manage, not fewer. Don't let "the G7 went soft on AI" turn into a reason to slow your own governance work.