Construction AI got plumbing - and the perimeter started to harden

A useful week to step back. Two themes ran in parallel, and they explain a lot about where the next quarter is headed.

The first theme is plumbing. Bluebeam Max went live globally on 19 May with Anthropic's Claude embedded directly inside Revu via MCP. Smart Overlay catches design changes across phases and scales. Smart Review flags scope gaps before they become RFIs. Magic Markups cuts the repetitive click-by-click work out of take-offs. The point isn't the feature list. It's that frontier AI is now built into the software UK contractors, consultants and QSs already open every day. No new platform to procure, no separate login, no migration project. The "should we adopt AI?" decision is being made for you as it lands inside the tools you already pay for.

Underneath Bluebeam, the model and tooling layer kept commoditising. Google's Antigravity 2.0 brought multi-agent coding orchestration to a free starting tier. Cursor's Composer 2.5 moved to standard pricing and still posts near-frontier coding benchmarks at roughly a tenth of the API cost of Claude Opus 4.7 or GPT-5.5. Alibaba's Qwen 3.7 Max arrived with a 1-million-token context window and the lowest reported hallucination rate among frontier models (worth a fact-check on the "open-weight giant-killer" framing some clips ran with - it's closed-weights, API-only, and sits around fifth on the leaderboards). The takeaway across all of these is the same one we keep returning to: capable agentic work is now a commodity, and the discipline that pays is routing each workload to the cheapest model that clears your quality bar.

The second theme is the perimeter. The governance question that's been quietly building for months started showing up in product launches. Anthropic added self-hosted sandboxes (public beta) so the actual execution of an agent's tools runs on infrastructure you control, with Anthropic still handling orchestration and recovery - and MCP tunnels (research preview) that let agents reach private MCP servers without exposing them to the public internet. Microsoft Copilot Studio's computer-use agents reached general availability with Azure Key Vault credential storage, Microsoft Purview audit logging, and configurable human-in-the-loop review. These are the features that decide whether an AI pilot ever reaches a live project on regulated work. The headlines this week were governance features dressed as product launches.

But the perimeter isn't only technical. The Royal Town Planning Institute warned that the planning system can't keep pace - the Economic Needs Assessment guidance authorities rely on for land allocation dates to 2019, predates the current AI wave entirely, and local authorities can no longer reliably forecast future employment levels. For anyone doing development, infrastructure or data-centre work, this is the policy story that touches the pipeline directly. AEC Magazine made the parallel argument from the standards side: the draft DIS/ISO 19650-1:2026 (out for consultation in March) says nothing about agents or autonomous workflows, even as vendors ship MCP-hooked platforms. MCP gets intent into an app quickly - it was never designed as a delegated-authority framework.

The most consequential perimeter story landed in PI insurance. Some professional indemnity insurers are writing AI out of their cover. Berkley has gone furthest with an absolute exclusion across D&O, E&O and fiduciary policies, naming ChatGPT, Bard, Midjourney and DALL-E specifically. Others are taking softer routes - sub-limits, disclosure requirements, narrower exclusions tied to specific use cases - but the direction is the same. If anyone in your business is using generative AI on chargeable work, this is the part of the AI story that lands in your wallet first. Audit who uses what on which deliverables, get the policy wording reviewed against that audit, and move ahead of the conversation rather than waiting for a claim.

Two practical bits for the diary. Digital Construction Week is at ExCeL on 3-4 June. The speaker bill is the strongest it's been: Arup, Balfour Beatty, Skanska, Mace, Laing O'Rourke, Heatherwick Studio, TfL, NHS, Tata Steel. The AI sessions to actually go for are "Will an AI-enabled construction industry still need main contractors?" on the Inspire Stage, and "Beyond BIM: AI, Digital Twins & Real-Time Delivery" on the Information Management Stage Wednesday at 16:00. Pick three sessions before you arrive, list two questions you want answered, and book one demo by appointment rather than queuing.

And one note on distribution. LinkedIn confirmed on 18 May that it's throttling reach for AI-cadence posts. The "it's not X, it's Y" structure, recycled thought leadership, perfectly parallel tricolons - all flagged. Posts still show to direct connections but get dropped from the recommendation feed. If your reach has gone quiet recently, the platform is part of the explanation. Use AI for research, outlines and proofing, but the final voice needs to be a person. I've drafted this issue with that in mind.

Pull all of it together and the test for the next quarter is clearer than it's been. Capability is no longer the constraint. Assurance is. The teams that move fastest will be the ones who pair every AI deployment with a defensible answer to three questions: where does the execution happen, who can see the data, and where's the audit trail? Get the data-classification, approval-gate and audit-logging house in order now, and you can move quickly when a pilot proves out. The frontier moved into your existing tools this week. The work now is making sure your perimeter moved with it.