Construction AI: the Building Safety use case lands, and the EU clock starts ticking

Truelens - an AI compliance-checking tool that CAST Consultancy helped test, shape and now uses in its Gateway 2 service - cross-references every word of a submission against the relevant Approved Documents and Building Safety Regulator requirements, producing a structured report of exactly where information is missing, inconsistent or potentially non-compliant. CAST reports a manual check that takes an experienced professional around 10 days being compressed to roughly an hour, with a reported ~65-73 per cent reduction in checking time. Truelens stores all client and project data on UK-based infrastructure, encrypted in transit and at rest. Crucially, the vendor is explicit: Truelens does not make compliance decisions and does not replace the architect, principal designer or BSA consultant - the judgement, and the liability, stays human.

This is the construction AI use case in its most defensible form. With Gateway 2 approval times sitting at 13-14 weeks - and barely 10 per cent of new-build Gateway 2 submissions reportedly approved - the bottleneck has moved upstream to getting submissions right first time. A word-by-word pre-submission QA pass is exactly the right place to apply AI.

Extract - built by the UK Government's AI Incubator team (i.AI) on Google Gemini - turns old planning documents, including blurry maps and handwritten annotations, into clean digital data. Trials at Hillingdon, Westminster, Nuneaton and Bedworth, and Exeter reported digitising planning records in around three minutes each, against one to two hours manually - roughly 100 records a day per user. The tool is being made available to all English councils by Spring 2026, with the aim of digitising all planning document types by the end of the year.

For anyone doing land, feasibility or development work, this matters directly. The digitised planning baseline that every site appraisal draws on is about to improve nationally - historic records become searchable, faster. It is also a useful reference point: a government department shipping a working, multimodal AI tool into local authority workflows is a credible signal of where public-sector adoption is heading.

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 before events occur. Separately, video-clip analytics tools like Fyld - where workers upload short clips rather than relying on fixed rigs - report up to a 48 per cent reduction in serious incidents among users.

The significance is that the safety benefit is now large enough to justify the spend on its own, without leaning on the secondary programme or quality arguments. That changes the internal business case: this is no longer "AI that also helps safety" - it is a safety intervention that happens to use AI.

Bentley's flagship MicroStation 2026 release leans into AI in a deliberately pragmatic way. The Python Assistant generates, explains, edits and reuses Python automation scripts conversationally, with the latest release focused on improved code reliability and reduced hallucinations. Bentley Copilot - shipped as a technology preview - is a chat-based help layer for learning, onboarding and workflow explanation. Notably, Bentley keeps the two separate: Copilot handles guidance, Python Assistant handles code.

The framing is the interesting part. This is not generative design or autonomous modelling - it is removing the scripting-capacity barrier for civils and infrastructure teams who have CAD-heavy workflows but no spare automation resource. For UK infrastructure consultancies, that is a more immediately bankable AI investment than most generative tooling.

A clear theme across this week's developer coverage: prompt injection and tool-misuse attacks against everyday agent setups are now being demonstrated with reproducible examples - including working data exfiltration via a poisoned MCP server. This lands directly on top of last week's Palisade Research self-replication paper and the Beazley poll that ranked construction the least prepared industry for cyber threats.

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. Meanwhile, 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.

This is the policy counterweight to a week dominated by genuinely useful regulatory AI tools. The same computer-vision safety systems delivering the 47 per cent PPE-violation reductions above are exactly the systems most likely to trigger high-risk classification under the EU framework. Adopting the tool and managing its conformity are now the same project.