Construction AI: a third of Gateway 2 applications still fail before assessment, and Anthropic wires up the controls for AI on your data

Here is the number worth pinning to the wall. In the 12 weeks to 30 May 2026, the Building Safety Regulator made 358 Gateway 2 decisions on higher-risk buildings, and around three-quarters were approved. That sounds like progress, and it is, set against the logjam that defined the back end of 2025. But it hides the real problem, which sits one stage earlier. Build UK reckons close to 30% of submissions are still found invalid at validation, which means they bounce before the regulator assesses the substance at all. You don't get marked down on your fire strategy. You get told the application wasn't put together in the required way, and you start again.

That gap between a 75% approval rate and a 30% rejection-at-the-door rate is the whole story. The schemes failing validation aren't necessarily unsafe. They are incomplete, or assembled wrong, or missing a document the regulator needs before it can even open the file. On 4 June the Construction Leadership Council and Build UK refreshed the Gateway 2 guidance suite, and tellingly the updates centre on the Fire and Emergency File and the Building Regulations Control Statement, two of the documents people most reliably get wrong. Build UK's validation checklist runs to 19 items. Most of the rejections come down to not working through it.

This is where AI earns its place, and where it doesn't. A model is genuinely useful for assembling a complete, consistently formatted application pack, cross-checking that every required document is present, and flagging the gaps before submission. It is not useful, and is frankly dangerous, as the author of the safety judgement itself. The honest framing is narrow: use the tools to stop yourself failing on a technicality, then keep a competent human accountable for the content. My own bias, for what it's worth, is that the firms who treat validation as a tedious clerical box-tick are the ones who keep losing weeks to it.

A practical step: Before your next Gateway 2 submission, run Build UK's items 1 to 19 as a final walkthrough, and have one named person confirm the Fire and Emergency File and the Building Regulations Control Statement are complete. A validation rejection is the most avoidable delay in the regime.

Less eye-catching than a new model, more relevant to anyone running a real project. Through June, Anthropic rolled out a set of enterprise governance features for Claude that matter the moment you stop experimenting and start pointing AI at live records. Two stand out. The first is enterprise-managed MCP connector access, starting with Okta, which puts centralised authorisation across Claude chat, Claude Code and Cowork for Team and Enterprise plans, so an administrator decides which connectors people can reach rather than leaving it to individuals. The second is managed agents that run inside a sandbox you control and connect only to your private MCP servers, keeping both the execution environment and the data it touches inside your own boundary. These are vendor-described capabilities, so test them against your own security requirements before you trust them, but the direction is clear.

Why this lands for construction specifically. The interesting AI use cases on a project all involve sensitive data: the golden thread, O and M information, commercial records, the very Gateway documents in the story above. The blocker has never really been whether a model can read a PDF. It is whether you can let it read your PDF without losing track of who authorised that, what it retained, and where the data went. Centralised connector control and a sandboxed agent are unglamorous answers to exactly that question. MCP became the connective standard across the AEC stack earlier this year, and this is the governance scaffolding catching up to it.

The competitive backdrop is worth a line, because it tells you how fast this is moving. On 15 June Anthropic paused a planned change that would have shifted Agent SDK usage onto separate monthly credits, and OpenAI used the same day to offer new business customers two months of free Codex. Both filed confidentially for IPOs this month. The vendors are fighting hard for enterprise seats, which is good for buyers as long as you read the governance terms rather than the launch copy.

The procurement filter: Before any AI touches project data, ask the vendor three questions: who can grant the connection, where does the agent run, and what is retained. If the answers are vague, that is your answer.

Strip both stories back and they rhyme. The Gateway 2 numbers say the constraint has moved off the model and onto the quality and completeness of what you put in front of the regulator. The Anthropic features say that once AI starts handling that material, the constraint moves again, onto who and what is allowed to see it. Neither is a story about clever models. Both are stories about discipline around data.

That sits comfortably with the standard already in force. The RICS professional standard on responsible use of AI has been mandatory since 9 March 2026, and its core principle is blunt: AI assists, it does not replace, and the surveyor stays accountable for every piece of advice regardless of the tools used. Apply that to this week and the action is concrete. Let AI assemble and check your Gateway pack, but sign it as a human. Let an agent read your project data, but only inside controls you set and can audit. The firms getting value from this aren't the ones with the newest model. They are the ones who fixed their inputs and their permissions first.

For your board pack: Adopt one rule this quarter. Any AI output that goes to a regulator, a client or into the golden thread carries a named human sign-off. It costs nothing and it is the difference between a tool that helps and a liability that hides.