Construction AI: the building safety queue clears faster, and the agent rails settle under one roof

The Building Safety Regulator published its latest Gateway 2 update and transparency data on 6 August 2026, covering the 12-week rolling period to 1 August. It's worth reading properly, because the direction has changed. Across all higher-risk building categories the BSR made 349 building control decisions and approved 82% of them. For new higher-risk buildings and conversions the approval rate is now 91%, with a median approval time of 22 weeks; in London it's 93% and 23 weeks. For comparison, a year ago that median sat at 43 weeks and the approval rate was 39%. So a job that used to disappear for the best part of a year now comes back in around five months.

The remediation side, which is where the human stakes are highest, has moved too. Existing-building remediation approvals reached 85% nationally and 92% in London, comfortably past the BSR's own 65% target for the end of 2026. Over the period the regulator determined 110 remediation decisions, approving 85% of them, covering 7,433 residential units. Legacy applications from 2024 have dropped to 13, down from 42 at the start of the year. What turned it around was not a piece of software. It was the BSR's Innovation Unit, set up in August 2025, bringing registered building inspectors, fire engineers and structural engineers into integrated teams; acting chief executive Charlie Pugsley says that cut assessment setup from months to days, and the unit's own approval rate has gone from 39% to 91% in its first year. These are BSR and government figures, not a vendor's, which is why I'd trust them more than most numbers we quote here.

Here's the bit that connects to everything else in this brief. The gateway is a document exercise before it is anything else, and the approvals ride on the quality of the application pack, the golden thread that follows a building from design to handover. What AI actually does well on this is the unglamorous part: keeping that pack current, flagging the drawing that changed after the last submission, chasing the missing fire-strategy sign-off before it becomes an invalidation. The comparison only goes so far, but it's a bit like a good document controller who never sleeps. Get that right and you're on the fast side of a 22-week median. Get it wrong and you're resubmitting. That's what it's about, and it lands on the person doing the paperwork, not head office.

On 20 August 2026 Google's Agent2Agent protocol, known as A2A, formally joined the Agentic AI Foundation, the Linux Foundation-run body that already stewards Anthropic's Model Context Protocol. If you've been reading this brief you'll know MCP by now: it's the layer that lets an agent reach into your project database, your drawings, your RFI log, and do something useful with what it finds. A2A does the other half, letting one agent hand work to another. Two protocols, two jobs, and from last week one neutral home. Axios flagged the move on 17 August; it was confirmed on the 20th.

Why does a governance reshuffle matter to a contractor? Because the construction agents we've covered plug straight into these rails. Bridgit's workforce-planning agents talk to ChatGPT, Claude, Copilot and Gemini through an MCP server. Procore's Digital Coworkers sit on the same kind of connection. When the standard those tools depend on is owned by one company, its roadmap is your risk. Under the foundation, the members list now runs from 49 a year ago to more than 250, with AWS, Anthropic, Google, Microsoft and OpenAI all signed on as top-tier backers, and changes go through a technical committee rather than a single boardroom. I'm not going to pretend a standards body is exciting. But the day one of your integrations breaks because a vendor quietly changed its mind, you'll be glad it wasn't theirs alone to change. The person who feels that is the site systems lead trying to keep three tools talking to each other.

Look at both items together and they're the same point wearing different clothes. The building safety queue clears faster when the application pack is clean. The agent economy works when the connections between your tools are governed and stable. In each case the value isn't in the clever bit at the front, it's in the boring discipline underneath: a record you can trust, a connection you can rely on, a human signing off at the end. The BSR's own line is worth keeping on the wall, that faster decisions will never be allowed to compromise safety. Same rule applies to your agents. Speed is welcome, but a person still owns the approval.

For your board pack: pick one gateway or handover pack you're preparing this quarter and ask a plain question, who is keeping this current, and could an assistant do the chasing so the humans do the judging. If the answer is nobody and yes, that's your pilot.