Weekly Roundup
The week AI met regulation in UK construction
Gateway 2 compliance checking, nationwide planning digitisation and the EU AI Act clock — this week's strongest construction AI stories were the unglamorous, regulatory ones.
The week construction AI grew up — and the regulator turned up too
This week, the most important UK construction AI stories were not flashy product launches. They were regulatory.
Truelens — tested and adopted by CAST Consultancy — cross-references a Gateway 2 submission against every relevant Approved Document and BSR requirement, compressing a 10-day manual check into roughly an hour. With Gateway 2 approval times sitting at 13–14 weeks and barely 10 per cent of new-build submissions reportedly approved, getting submissions right first time is the real bottleneck. The vendor is explicit: Truelens does not make compliance decisions. The judgement, and the liability, stays human. That framing is the right one.
The government's Extract tool, built by i.AI on Google Gemini, is rolling out to every English council by Spring 2026 — turning old planning records (blurry maps, handwritten notes included) into clean digital data in about three minutes versus one to two hours manually. The digitised planning baseline every site appraisal draws on is about to improve nationally.
Both stories share a pattern. The strongest construction AI use cases right now are the unglamorous ones: word-by-word pre-submission QA, planning record digitisation, computer-vision safety. The kind of work that compounds quietly across a portfolio. The generative razzle-dazzle is still mostly demo.
But, the regulator turned up at the same time. The EU AI Act's high-risk obligations go live on 2 August 2026. Worker monitoring via computer vision, automated compliance flagging and safety-critical AI all potentially fall in scope — including the same systems delivering the reported 47 per cent fall in PPE violations on UK sites. The UK's own AI Bill has slipped to H2 2026 at the earliest, so for any firm operating cross-border or selling into the EU, the EU Act is the binding constraint. Adopting the tool and managing its conformity are now the same project.
Underneath the regulatory story, the wider AI stack kept moving. UK Construction Week London opened on Tuesday with the ConTech & AI Hub as the headline programme — the densest UK audience of construction AI buyers in the calendar. ProcurePro raised US$11m led by QIC Ventures with Bouygues moving from customer to investor, explicitly to scale UK and Middle East procurement-AI adoption. Anthropic shipped Claude Code Agent View, the first mainstream supervision dashboard for parallel coding agents. OpenAI launched Daybreak as a direct counter to Anthropic's Mythos/Project Glasswing — frontier AI is now an enterprise security category, and the alliances forming around it (AWS, Apple, Cisco, Microsoft, NVIDIA, JPMorganChase) will set the defender ecosystem your clients and insurers expect you to plug into.
And the wider research kept the ceiling in view. GPT-5.4, Claude Opus 4.6 and Gemini 3.1 all scored exactly 0 per cent on the new ARC-AGI 3 benchmark, where untrained humans score 100 per cent. Use that number deliberately when defending which decisions on a project should stay human. Palisade Research showed frontier models can self-replicate across vulnerable networks at up to 81 per cent success — agent permissions, sandboxing and audit trails are now first-order procurement questions, not IT hygiene. David Jones (ICM) put the CDM and PI accountability question on AI-generated designs squarely on the table.
Pull all of that together and the test is shifting again. A month ago it was: can your firm govern, fund and lead AI? This week it is: can your AI defend a regulatory bottleneck, survive a conformity assessment and audit a self-replicating agent? Can your procurement language ask for least-privilege permissions, exclusion transparency, tracked-change audit trails and EU AI Act documentation?
That is a less exciting list than "agentic, multi-modal, self-improving". It is also the one that will decide whether AI earns its place on a UK project this year.
Top Stories This Week
Truelens + CAST: Gateway 2 compliance check that turns 10 days into an hour
Truelens — tested and adopted by CAST Consultancy — cross-references every word of a Gateway 2 submission against the relevant Approved Documents and Building Safety Regulator requirements, producing a structured report of where information is missing, inconsistent or potentially non-compliant. CAST reports a manual check that takes around 10 days compressed to roughly an hour, with a 65–73 per cent reduction in checking time. Truelens stores data on UK-based infrastructure and is explicit that the tool does not make compliance decisions — the judgement and the liability stay human.
This is construction AI in its most defensible form. With Gateway 2 approval times at 13–14 weeks and barely 10 per cent of new-build submissions reportedly approved, a word-by-word pre-submission QA pass is exactly the right place to put AI.
Why it matters
If your firm touches higher-risk buildings, a pre-submission AI assurance pass is now a credible, liability-aware workflow. Lead with the regulatory bottleneck, not the time-saving.
Extract goes nationwide — planning digitisation reaches every English council
Extract — built by i.AI on Google Gemini — turns old planning documents, including blurry maps and handwritten annotations, into clean digital data. Trials at Hillingdon, Westminster, Nuneaton & Bedworth and Exeter digitised planning records in around three minutes each, against one to two hours manually — roughly 100 records a day per user. The tool will be available to all English councils by Spring 2026, with the aim of digitising every planning document type by year-end.
For land, feasibility and development work this matters directly. The digitised planning baseline every site appraisal draws on is about to improve nationally. It is also a credible signal of where public-sector AI adoption is heading.
Why it matters
Factor the improving planning-data baseline into how you scope feasibility and due-diligence work through 2026. The manual-records tax on early-stage appraisals is starting to fall.
EU AI Act high-risk obligations go live on 2 August — UK Bill has slipped
The EU AI Act's high-risk system obligations become enforceable on 2 August 2026. Several common UK construction workflows could fall in scope: worker monitoring via computer vision, automated decision-making such as compliance flagging, and safety-critical AI applications. The UK's own AI Bill has slipped to H2 2026 at the earliest — so for any firm operating cross-border or selling into the EU, the EU Act is the binding constraint, not the UK's lighter-touch position.
The same computer-vision safety systems delivering the reported 47 per cent PPE-violation reductions on UK sites are exactly the systems most likely to trigger high-risk classification. Adopting the tool and managing its conformity are now the same project.
Why it matters
Start the AI Act conformity assessment now — not in July. Map which deployed or planned AI systems could be high-risk, and who owns the documentation.
UKCW London 2026: ConTech & AI Hub draws the densest UK buyer audience of the year
UK Construction Week London ran 12–14 May at ExCeL, co-located with Futurebuild and the Stone & Surfaces Show. Over 25,000 built-environment professionals, 600+ exhibitors, 700+ speakers and 14 stages combined into a "super event" format. The standout addition for 2026 was the ConTech & AI Hub — sessions covered automation, robotics, digital twins, AI-driven decision-making, generative design, on-site risk anticipation, automated estimating and AI in construction sales and marketing. Two new-for-2026 features were worth the trip: the Live Demonstration Zone and the Future Tech Bunker.
The value now is in the debrief. Pick the three vendors to scope in the next two weeks, the regulatory or compliance issue you heard at the most stages, and the one operational pattern to put a small pilot behind this quarter.
Why it matters
This was the densest UK audience of construction AI buyers in the calendar. Convert show-floor energy into a 90-day plan this week, not next month.
ProcurePro raises US$11m with Bouygues moving from customer to backer
ProcurePro — the Brisbane-founded, globally pitched procurement-AI platform — closed an US$11m (A$15m) round led by Queensland government-backed QIC Ventures, with existing investors AirTree and Glitch Capital topping up and French main contractor Bouygues joining the cap table via Paris-based ISAI. The round values the six-year-old company above US$100m. ProcurePro has run more than 6,000 construction projects globally and will hire 100 staff across product, engineering and go-to-market, with UK, Middle East and US expansion explicit.
The platform's pitch — consolidating 20+ fragmented procurement workflows so main contractors can compare, select, approve and contract on one AI-powered system — is procurement's version of what BIM tried to do for design. The Bouygues move from customer to investor is the credibility signal.
Why it matters
Procurement AI is now a venture-priority subsector with UK money behind it. Expect more visible RFPs from main contractors over the next quarter and tighter integration demands on every system that touches their supply chain.
OpenAI Daybreak vs Anthropic Mythos/Glasswing — frontier AI becomes a security category
OpenAI launched Daybreak on 11 May as a cybersecurity platform built on GPT-5.5 and the Codex Security agent — connecting to codebases and infrastructure, simulating attack routes, surfacing vulnerabilities, generating and testing patches inside repositories and producing audit-ready validation. It is a direct counter to Anthropic's Claude Mythos and Project Glasswing, whose partner ecosystem already includes AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks.
The construction read-through is sharp. Last week, Beazley ranked construction the least prepared industry for cyber threats. The same week, Palisade Research showed frontier models can self-replicate across vulnerable networks at up to 81 per cent success. Now both major frontier labs are racing to build defender programmes — but those programmes are gated to enterprises in their partner ecosystems.
Why it matters
Pick an ecosystem and get on the right defender table early. The "AI defender" capability your insurers and clients will start asking about in 2026/27 is being decided in these alliances now.
50 free Intelligence Units. See what AI can do for your projects.
Also Worth Noting
CDM accountability for AI-generated designs is no longer theoretical
David Jones (Director of Education & Training at the Institute of Construction Management and 2025 ICM President) put a sharp question to UK practitioners on LinkedIn: when AI generates the design, who is the "designer" under CDM? Who carries Principal Designer duties under the Building Safety Act? Who holds the PI insurance when something goes wrong? Generative-design tools that go from prompt to compliant floor plan are now production-grade, and existing accountability frameworks are under stress.
Why it matters
Draft a one-page firm position on AI-assisted design — competence, sign-off, PI scope, audit trail — before the question lands in a tender or HSE conversation.
Source: Construction News — PI insurers question use of AI in construction →
Palisade Research: frontier agents can now self-replicate across networks
Palisade Research's "Language Models Can Autonomously Hack and Self-Replicate" tested GPT-5, 5.1, 5.4; Claude Opus 4, 4.5, 4.6; and Alibaba Qwen against systems with deliberately planted security flaws. Claude Opus 4.6 succeeded in 81 per cent of self-replication tests; GPT-5.4 reached 33 per cent. The trajectory has moved from around 6 per cent to 81 per cent in twelve months.
Why it matters
Agent permissions, sandboxing and audit trails are now first-order procurement questions, not IT hygiene. Treat them like a JCT clause.
Source: Euronews — AI models can hack computers and self-replicate →
ARC-AGI 3: every frontier model scored 0 per cent on novel reasoning
GPT-5.4, Claude Opus 4.6 and Gemini 3.1 all scored 0 per cent on the new ARC-AGI 3 benchmark, where untrained humans score 100 per cent. The same models simultaneously score higher on code, professional exams and long multi-step tasks — so this is a structural limitation, not a regression.
Why it matters
This is the single most credible counterweight to "AI can do anything" claims at boardroom level. Use it deliberately to defend human-in-the-loop on judgement-heavy decisions.
Source: MindStudio — Why GPT-5.4, Claude 4.6 and Gemini 3.1 all scored 0% on ARC-AGI 3 →
Airbnb says AI now writes ~60% of new code
On its Q1 2026 earnings call, Airbnb CEO Brian Chesky said AI wrote around 60 per cent of code produced by Airbnb engineers during the quarter, and that AI now resolves 40 per cent of customer support issues without human escalation. The framing is human-supervised AI agents allowing small teams to do work that previously required large ones — not full replacement.
Why it matters
Anchor AI-investment conversations to this number, with caveats. It will be quoted at every digital transformation board through 2026 — decide whether you use it to justify investment or set realistic expectations.
Claude Code Agent View — agent supervision becomes the product layer
Anthropic shipped Claude Code Agent View as a Research Preview on 11 May. The dashboard puts every Claude Code session into a single CLI table — running, blocked waiting on you, or done — with the ability to launch new sessions and navigate by keyboard. The "open six terminals and a tmux" phase of agentic AI is closing.
Why it matters
Treat agent supervision as the first-class procurement question, not an afterthought. Audit trail, permissioning model and team productivity all sit at this layer.
Helix 02 and production Atlas turn humanoid demos into deployment plans
Figure AI's Helix 02 — a vision-language-action system controlling full-body humanoid movement directly from pixels — has demoed 4-minute continuous dishwasher loads and unprompted bedroom resets. Boston Dynamics' production-ready Atlas, with 56 degrees of freedom and 50kg lift, has its 2026 deployment slots committed to Hyundai's RMAC and Google DeepMind. For UK contractors with Unitree-class humanoid pilots already in flight (Tilbury Douglas's "Douglas" being the visible example), the gap to "deployable site asset" narrowed measurably this quarter.
Why it matters
Get your next-step business case ready. The contractors with the cleanest "physical AI" roadmap for 2027 will be the ones that treated 2026 as a build-the-evidence year.
Open-source agent stack got serious: Hermes 0.13 + Desktop, Mistral Vibe 2.0
Nous Research shipped Hermes Agent 0.13 "Tenacity" (durable multi-agent kanban, /goal locking, pluggable models including DeepSeek v4 Pro) and a Hermes Desktop app for always-on local agents. Mistral Vibe 2.0 added custom subagents and remote cloud agents. For UK construction firms with confidentiality clauses that block cloud-only AI, the self-hosted option just got materially stronger.
Why it matters
If data residency currently blocks cloud-only AI tools, give Hermes Desktop a focused two-week evaluation.
Computer vision on UK sites: the safety case stands on its own
Reported figures from UK sites running embedded AI vision (fixed cameras and drone scans compared against BIM in near real time) point to a 47 per cent fall in PPE violations within six months, a 36 per cent fall in site safety incidents, and 52 per cent better hazard identification. Video-clip analytics tools like Fyld report up to a 48 per cent reduction in serious incidents.
Why it matters
If your AI-on-site business case has been stuck waiting for a productivity number, reframe it as a safety-and-insurance case. The evidence base there is now stronger.
Source: Construction Magazine — AI live on UK construction sites →
MyQS.ai and Claude for Word — the audit trail is the procurement criterion
Two practitioner-grade tools worth flagging. MyQS.ai (built by UK QS Sofoclis Patsalos) generates NRM1-tagged Bills of Quantities from drawings, voice notes or text — and explicitly flags what it has excluded (perimeter fencing, kitchen units etc.) rather than hiding gaps behind a polished number. Claude for Word proposes every edit as a tracked change in Word's native review pane, with each change individually acceptable. Both point to the same procurement criterion: exclusion transparency and tracked-change auditability beat benchmark scores.
Why it matters
When evaluating AI tools for tender pricing, contracts or BSA submissions, weigh auditability as a first-class criterion. It is a better signal of production-readiness than accuracy claims.
Adoption: AI in UK construction projects 15% → 75% in two years
PBC Today's most recent reading reports AI usage in UK construction projects has jumped from 15 per cent two years ago to around 75 per cent today, with significant variation by firm size, region and project type. The line worth using internally: you are no longer leading by trialling AI; you are catching up by not deploying it.
Why it matters
If your firm is still in pilot-only mode, the conversation has shifted around you. Rewrite the AI narrative for the catching-up frame, not the leading-edge one.
Foxglove keeps the data-centre carbon and AI Growth Zone scrutiny rising
Foxglove's legal and journalistic campaign over UK data-centre carbon and AI Growth Zone job-creation claims continued through the week — January's government admission of a "serious error", March's council approval of a data centre with ~1Mt annual climate pollution, and FT-backed analysis of Essex Google facility emissions. This is a hardening accountability environment for AI Growth Zone bids.
Why it matters
If you are tendering on AI Growth Zone or hyperscale data-centre work, your carbon, water, grid and community case has to be defensible at the level of the Foxglove and FT critiques — not the marketing brochure.
Bentley MicroStation 2026: AI for scripting and onboarding, not generative theatre
Bentley's flagship MicroStation 2026 release leans pragmatic. The Python Assistant generates, explains, edits and reuses Python automation scripts conversationally. Bentley Copilot — a tech preview — is a chat-based help layer for learning and onboarding. The two are kept deliberately separate.
Why it matters
The highest-ROI AI in CAD-heavy teams right now is automation scripting and onboarding, not generative design. Scope your tooling decisions accordingly.
What matters most
- →Lead AI business cases with the regulatory bottleneck (Gateway 2, planning, safety), not the time-saving.
- →Standardise on one agent supervision and permissioning control plane before you scale anything.
- →Start the EU AI Act conformity assessment now — not in July.