Weekly Roundup
The week construction AI got bankable
EY and Cambridge put a board-grade case behind agentic AI, Skanska showed where the moat really lives, Google I/O reset the cost curve, and a robot bricklayer quietly went back to work in Durham.
Board-grade case, bankable wins — and the moat moves to your own data
Three things matured at once in construction AI this week.
First, the productivity case got formal. EY, the University of Cambridge (Laing O'Rourke Centre and the Centre for Smart Infrastructure and Construction), NVIDIA and Futurity Systems published "The intelligence layer", reviewing more than 100 construction productivity studies and concluding that agentic AI is the most credible route to closing the sector's productivity gap. Cambridge is the part that matters. This is now the report you can put in front of an infrastructure client board, a procurement panel or an internal finance director without having to defend the source. Houzz's first UK State of AI in Construction & Design report added a directional figure — early adopters reporting roughly 3 hours/week and around £23,000/year per practitioner saved.
Second, the moat moved. Skanska's "Expert Sidekicks" — purpose-built agents trained on thousands of documents from its safety leaders, and crucially on how those experts reasoned through decisions in the field — are the construction AI signal that matters more than any model launch. Large contractors are now building their own domain agents on proprietary corpora rather than waiting for vendors. The defensibility is in the data and the captured expertise, not the underlying model. For UK firms, the strategic question follows: what is your equivalent corpus? Decades of project records, RAMS, lessons-learned and senior-engineer judgement are the asset. The model is increasingly a commodity.
Third, the bankable wins surfaced. The WLTR "Walter" robot bricklayer is building 27 homes in Durham — up to 200 m² of masonry a day, walls to 3.5m without scaffolding, framed openly around the 35,000+ unfilled construction roles and the 1.5-million-homes target. Sir Robert McAlpine and Vinci's JV used Buildots on Royal Bournemouth Hospital; Vinci Construction UK reports saving more than 6,000 documentation hours annually with OpenSpace. None of those wins are autonomous. They are admin-cutting, dispute-reducing, ground-truth-giving. They are also the easiest pilots to justify to finance.
Underneath those three, the plumbing kept settling. Google I/O delivered a substantive reset on Tuesday night: Gemini Omni folds text, image and video generation into a single architecture, Gemini 3.5 Flash now beats last year's Gemini 3.1 Pro on several agent benchmarks at roughly 40 per cent lower cost, and Spark and Android XR repositioned the wider Google offering toward proactive agents. The coding-agent market became a three-horse race when xAI launched Grok Build with native Anthropic-skill compatibility and $99 intro pricing. Anthropic published "Explore → Plan → Code → Commit" as the canonical Claude Code methodology and a separate engineering blueprint for agents that run for hours — planner/generator/evaluator, adversarial evaluators over self-evaluation, structured handoffs over naive context compaction. Claude Opus 4.7 task budgets let you cap an agent's token spend across a whole run. The MCP versus ADK debate settled into a clean answer: MCP for agent-to-tool, A2A/ADK for agent-to-agent, design for both. Procore's Q1 results positioned agentic AI as embedded platform capability, not a roadmap item — UK clients on Procore should expect AI features in their existing licences. SEGRO and Pure DC secured planning approval for a £1bn, 72MW hyperscale data centre at Premier Park, west London, with construction starting in 2026.
But, the most useful single sentence of the week came from the UKCW "AI or Die" debrief. The sharpest part of the debate was not about capability. It was about trust, data security and the cultural resistance teams hit when introducing new tooling. The market has moved past "should we use AI" to "how do we govern it and get people to trust it". The capability case is largely won. The trust case is where deals and rollouts now stall.
So pull all of that together and the test is shifting again. A week ago it was: can your AI defend a regulatory bottleneck? This week it is: can it earn trust, and does the value come from your data rather than your vendor's model? Can you point to a bankable win — a robot, a reality-capture saving, a routed cheaper-model workflow — instead of a roadmap?
That is a more grown-up question. It is also a much easier one to answer in the room.
Top Stories This Week
EY × University of Cambridge × NVIDIA: agentic AI as the "intelligence layer" for infrastructure
EY released "The intelligence layer: how agentic AI can connect the infrastructure industry" — produced with the University of Cambridge (Laing O'Rourke Centre and Centre for Smart Infrastructure and Construction), NVIDIA and Futurity Systems. It reviews more than 100 construction productivity studies and concludes that agentic AI is the most credible route to closing the sector's productivity gap. Agents act as an intelligence layer connecting siloed BIM, ERP, scheduling and sensor systems, surfacing decisions in real time. The report puts global infrastructure investment need at US$140tn by 2050 against a current US$64tn funding gap, and reiterates the 15 per cent rework-cost baseline.
The Cambridge endorsement is the part that matters. This is the report you can put in front of an infrastructure client board, a procurement panel or an internal finance director without having to defend the source.
Why it matters
This is now the most quotable, board-grade reference for agentic AI in infrastructure. Cite it in proposals, internal investment cases and client briefings.
Skanska's "Expert Sidekicks" — the proprietary-corpus agent is the real story
Skanska is building a suite of generative-AI tools it calls Expert Sidekicks — purpose-built agents infused with Skanska-specific knowledge. The Safety Sidekick is the visible example: trained on thousands of documents from experienced safety leaders (policies, procedures, EHS manual, OSHA standards) and, crucially, on how those experts reasoned through decisions in the field — not just the rules they applied. Skanska frames the goal explicitly as closing the industry's knowledge gap as experienced people retire.
This is the trend that matters more than any model launch. Defensibility is shifting to the proprietary corpus and the captured expertise. For UK firms, decades of project records, RAMS, lessons-learned and senior-engineer judgement are the asset. The model is increasingly a commodity.
Why it matters
Your competitive moat in AI is your proprietary data and your experts' captured judgement. Start treating your document archive and your senior people's reasoning as a training asset.
Google I/O 2026: Gemini Omni, 3.5 Flash, Spark — and a Flash that beats last year's Pro
Google I/O 2026 delivered a substantive reset on Tuesday night. Gemini Omni folds text, image and video generation into a single architecture. Gemini 3.5 Flash — now the default behind the Gemini app and AI Mode in Google Search — out-scores the prior Gemini 3.1 Pro flagship on several agent benchmarks (76.2 per cent on Terminal-Bench 2.1, strong on MCP Atlas, Toolathlon and Finance Agent v2), runs at roughly 289 tokens/sec, and costs around 40 per cent less than 3.1 Pro. Spark is a redesigned creative surface; Stream to Cursor deepens IDE integration; Android XR previewed for smart glasses ahead of Samsung's Galaxy Glasses at Unpacked in July.
For construction, the relevance is in high-volume, repetitive workflows — document triage, classification, RFI summarisation, compliance-record search at scale. The cost per task on those workloads just dropped meaningfully.
Why it matters
Build a routing layer that swaps to the cheapest model clearing your quality bar, rather than hard-wiring one frontier model into every workflow. Re-cost any high-volume document workflow you run.
The WLTR "Walter" robot bricklayer is back on UK sites — framed around the labour gap
The WLTR ("Walter") wall-laying robot is building 27 homes in Durham for JT Lifestyle Homes — up to 200 m² of masonry a day, walls to 3.5m without scaffolding, operating in rain, wind and temperature extremes, using DryFix adhesive foam rather than traditional mortar. Future deployments are lined up at a former Post Office redevelopment in Hull, new homes at Haywood Park in Southampton and an apartment refurbishment in Nottinghamshire.
The framing is the part to study. With the trade's average worker age at 46, too few young entrants, 35,000+ unfilled roles and a 1.5-million-homes target, the pitch is "automation complements human labour" and "fills the gap" — not headcount reduction. That is a far more durable narrative with clients, unions and planners.
Why it matters
When you make the case for any AI or automation on site, lead with workforce resilience and the skills shortage. It lands better with operations leads and procurement than a pure productivity pitch — and it's the truthful framing right now.
Reality capture has the hard UK numbers now — Buildots and OpenSpace
Sir Robert McAlpine and Vinci's joint venture (IHP) ran Buildots — 360° site video automatically mapped to BIM and the programme — on the Royal Bournemouth Hospital project, using it for QA, billing and delay-risk mitigation; Sir Robert McAlpine has adopted Buildots as a preferred partner across large areas of live work. Separately, Vinci Construction UK reports saving more than 6,000 hours annually on documentation alone using OpenSpace for automated progress capture and remote verification.
These are the unglamorous wins that pay back. They cut admin load and reduce disputes by giving every party the same visually-verified ground truth — rather than promising autonomy or replacing judgement.
Why it matters
If you want a defensible first AI deployment, reality-capture progress tracking is the lowest-risk, best-evidenced starting point.
UKCW debrief: "AI or Die" panel lands on trust, not capability
The Construct UK panel at UK Construction Week — "AI or Die: Rewriting the Rules of Construction Sales and Marketing" — played to a full-capacity room. The sharpest part of the debate was not capability. It was trust, data security and the cultural resistance teams hit when introducing new tooling. The market has moved past "should we use AI" to "how do we govern it and get people to trust it".
For anyone selling or deploying construction AI, the objections you will hit are now overwhelmingly about trust and security, not whether the tool works.
Why it matters
Lead your AI conversations — internal and client-facing — with trust, data residency and change management. The capability case is largely won; the trust case is where deals and rollouts now stall.
50 free Intelligence Units. See what AI can do for your projects.
Also Worth Noting
Houzz 2026 UK State of AI: £23k/year per practitioner
Houzz published its first UK State of AI in Construction & Design report. Early adopters report savings of around 3 hours/week and a £23,000 annual productivity uplift per practitioner. Top use cases: administrative work (79 per cent of pros overall, 89 per cent construction), sales and marketing (52 per cent) and planning/design visualisation (33 per cent). Three in five respondents believe AI will transform the industry within five years.
Why it matters
The "AI is for tier-one contractors" narrative is hard to defend. SME and mid-tier practices are seeing measurable benefit, and the tooling is cheap enough to deploy without enterprise budgets.
Procore Q1 2026: agentic AI is now embedded, not a roadmap item
Procore reported Q1 2026 revenue of $359m (up 16 per cent YoY) and used the call to position agentic AI as embedded platform capability — "autonomous AI-driven software agents embedded in Procore's platform". The company expects increased sales of its AI offerings by Q3 FY2026 and shipped a wave of platform updates including a Planning feature beta and a redesigned submittal experience.
Why it matters
UK contractors and consultants whose clients run Procore should expect AI features in their existing licences. The "do we need a separate AI tool?" question will land by end of summer — be ready with the answer.
Source: The Motley Fool — Procore Q1 2026 earnings transcript →
Anthropic's long-running-agent blueprint + Opus 4.7 task budgets
Anthropic engineers published a workshop and engineering write-up on building agents that run coherently for hours: a three-agent structure (planner, generator, evaluator), adversarial evaluators over self-evaluation (self-grading reliably becomes self-praise), structured handoffs over naive context compaction. Separately, Claude Opus 4.7 introduced task budgets — a parameter that caps an agent's token spend across an entire run, with the model self-regulating against the depleting budget.
Why it matters
If you commission agentic workflows, require an independent evaluator in the architecture. If pilot economics were your blocker, task budgets remove a major source of unpredictability — retest the cost case.
£1bn west London hyperscale data centre approved at Premier Park
SEGRO and Pure Data Centres secured planning approval from the Old Oak and Park Royal Development Corporation for a £1bn, 72MW hyperscale data centre at Premier Park — 22,365 sq m, three storeys, nine data halls, designed by Scott Brownrigg. Sustainability brief: BREEAM Excellent target, A-rated EPC for offices, PV panels, rainwater harvesting and potential heat recovery. Construction starts in 2026.
Why it matters
If you bid or design into hyperscale data-centre work, lock in your AI-driven scheduling, programme controls and compliance-evidence story early. London data-centre approvals are hitting construction starts faster than traditional commercial.
xAI Grok Build CLI — coding-agent market becomes a three-horse race
xAI launched Grok Build on 15 May — an agentic CLI for professional software engineering, in direct competition with Claude Code and Codex. Beta access via SuperGrok Heavy ($300/month) with $99/month intro pricing for the first six months. Eight concurrent specialised sub-agents, plan mode, native compatibility with the Anthropic skill format (drop ~/.claude/skills/ into Grok's working directory unmodified), and MCP server discovery.
Why it matters
If you're locked into one coding-agent vendor, this is the right month to benchmark alternatives — particularly on permissioning, audit and skill-portability dimensions.
32 production agent system prompts now public
A GitHub repository (Lucas Valbuena's x1xhlol/system-prompts-and-models-of-ai-tools, plus the EliFuzz/awesome-system-prompts mirror) collects the leaked system prompts and tool schemas of more than 28–32 production AI coding agents — Cursor, Windsurf, Claude Code, Devin, v0, Lovable, Replit, Manus, Augment Code and others. 137k stars and climbing. Crucially, the JSON tool schemas are included alongside the prompt text.
Why it matters
Before building or buying an agent for project-critical workflows, read several of these prompts back-to-back. The differences in tool permissions are where security risk actually lives.
MCP vs ADK settles — design for both, not either
IBM Technology's Day-1 video this week crystallised the buying question. Anthropic's Model Context Protocol (MCP) is the agent-to-tool layer — APIs, file systems, GitHub, Drive, Slack. Google's Agent Development Kit (ADK) and the Agent-to-Agent (A2A) protocol are the agent-to-agent layer — orchestration and handoffs. They are complementary, not competing.
Why it matters
Procurement decisions should require vendors to support MCP for tool connectivity and A2A or ADK for multi-agent orchestration. Anything less is a single-layer answer.
Open-weight quality catches up — Gemma 4 and Microsoft MAI
Google's Gemma 4 ships under Apache 2.0 in four variants (2B, 4B, 26B MoE, 31B dense) with native multimodal support; the 31B model ranks #3 on Arena AI at 89.2 per cent on AIME 2026. Microsoft's MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 are the first production models built independently of OpenAI — MAI-Transcribe-1 hits 3.8 per cent average WER across 25 languages at ~50 per cent lower GPU cost than leading alternatives.
Why it matters
If a vendor told you twelve months ago that open-weight quality was a year out, that answer is now wrong. Re-open the "local AI" feasibility conversation, particularly for HRB and BSA Golden Thread workflows.
All3 raises $25m to put legged robots above the foundation
UK construction-robotics startup All3 closed a $25m seed led by RTP Global, with SuperSeed, Begin Capital, s16vc and VNV Global participating. The pitch is end-to-end: AI generates building designs tailored to a site, robotic factories produce custom parts off-site, and mobile robots — including Mantis, a four-legged autonomous platform — handle installation and finishing. Initial commercial deployment is on active sites in Germany.
Why it matters
If you operate at the off-site/MMC end of the market, watch where All3's first UK pilots land. This is a 12-month partner conversation, not a 24-month one.
NotebookLM as the low-risk way to interrogate specs, RAMS and Gateway 2 evidence
Source-grounded assistants like NotebookLM answer only from the documents you provide — PDFs, drawings, specs, O&M manuals, contract documents — rather than the open web. That sharply reduces hallucination risk on contractual or compliance-critical material. For BSA evidence, Golden Thread documentation, NEC clause interrogation or O&M handover, this is a low-risk entry point.
Why it matters
If trust is the blocker (UKCW says it is), source-grounded assistants are the lowest-risk way to demonstrate value. Pilot one on a real document set this fortnight.
Source: Construction Magazine UK — AI on live UK construction sites 2026 →
Keep your scepticism calibrated — the Builder.ai cautionary tale
A useful counter-current: recurring references to the Builder.ai collapse — the "AI" app-building service later reported to have leaned heavily on hundreds of human engineers behind the scenes. The fastest way to lose credibility with a sceptical construction audience is to oversell autonomy.
Why it matters
Calibrated scepticism is an asset. Be the person in the room who asks "what's actually automated here?" — it builds trust faster than enthusiasm does.
Source: Browne Jacobson — AI and emerging legal challenges →
What matters most
- →Lead AI conversations with trust, data residency and change management — capability is no longer the blocker.
- →Treat your document archive and your senior people's reasoning as a training asset, not just a record.
- →Build a routing layer that picks the cheapest model clearing your quality bar, instead of hard-wiring one frontier model into every workflow.