Construction AI: the planning system becomes an AI adopter, and Whitehall bets on Gemini to run it

Here's a development that matters more to construction than most model launches, and it slipped out at a cloud-vendor conference rather than a building one. At the Google Cloud Summit in London on 17 June 2026, the Ministry of Housing, Communities and Local Government and DSIT confirmed that Extract - an AI tool the government's own i.AI incubator built on Google's Gemini foundation models - is now available to every local planning authority in England, following trials across more than 20 councils.

What Extract does is unglamorous and genuinely useful. It takes the historic maps, site plans and planning documents that councils sit on - much of it scanned, hand-annotated, locked in formats nobody can query - and turns it into standardised, usable data. The government says that cuts a task that took roughly two hours down to about two minutes, and saves an average council around 255 hours a year (these are government-reported figures, and I'd treat the headline savings as directional until someone independent audits a full year of use). Fifty authorities are already using it, and in the early weeks they processed over a thousand documents - mostly Tree Preservation Orders, which tells you something about where the unglamorous bottlenecks actually sit.

The reason this matters to you isn't the productivity stat. It's that the body that determines your applications is now an AI adopter in its own right. The planning system has spent a decade being the thing that slows construction down; it's now reaching for the same tools we are. If your submissions still arrive as flat scanned PDFs and inconsistent site plans, you're handing a machine the worst possible input at the exact moment the machine starts to matter.

Today's action: Audit how your planning submissions actually arrive at the council - file formats, naming, whether your drawings and documents are machine-readable. The authority's tooling is changing; your inputs should too.

Extract handles the paperwork. The tool worth watching is the one that touches the decision itself. Alongside the Extract news on 17 June, MHCLG and DSIT shared progress on Augmented Planning Decisions (APD) - a prototype now in alpha with the London Borough of Barnet, Dorset Council and the London Borough of Camden, built as a collaboration between the government, Google Cloud, Google DeepMind and the AI firm Faculty.

The ambition is bigger than data tidying. APD is meant to help planning officers navigate complex local policy, and the stated target is to halve the time to determine a householder application - from eight weeks to four. National availability is pencilled in for 2027, subject to the alphas going well. That timeline will slip, these always do, and "helping officers reason about policy" is a far harder and far more contested thing than converting a Tree Preservation Order into a database row. But the direction is unmistakable: the state isn't just speeding up the admin around planning, it's putting AI next to the judgement at the centre of it.

For developers and consultants, the practical signal is the pilot list. If you have schemes going through Barnet, Dorset or Camden, you may be among the first to feel a different determination rhythm - and the first useful, real-world data point on whether AI-assisted decisions are faster without being worse.

For your board pack: Note the three alpha authorities and the 8-to-4-week target as a planning-risk assumption to test, not bank. If you build in those areas, ask the case officer whether APD touched your application - the early anecdotes will be worth more than the 2027 press release.

The same Summit carried a second thread that's easy to wave past as consultancy noise, but shouldn't be. On 17 June, Deloitte and Google Cloud announced a London AI Studio, sited at Deloitte's London campus, with the explicit job of moving British organisations past AI experimentation and into deploying "autonomous, action-oriented" agentic systems at scale. Google framed the whole Summit around the shift "from AI potential to agentic reality" and announced a clutch of new UK partnerships alongside it.

Why a construction audience should care: the big advisory and delivery firms are where a lot of major UK infrastructure and public-sector programmes get shaped, costed and run. When Deloitte stands up a dedicated agentic-AI studio with a hyperscaler, the agents that emerge will end up inside the programme controls, cost models and assurance processes that sit over schemes you're delivering on. This is the plumbing layer arriving, not a demo.

And there's a governance detail from the government side of the Summit worth keeping. Whitehall's planning models run on Gemini inside a protected environment, chosen specifically to mitigate prompt injection - the attack where hostile text hidden in a document hijacks the model reading it. If the government won't let a planning AI loose without that guardrail, neither should you let an agent loose across an O&M pack, a tender library or a CDE without asking the same question. The honest worry with agentic everything isn't capability - it's what happens when an agent reads a poisoned document and acts on it.

The procurement filter: Before you sign up to any "agentic" delivery or platform pitch this year, ask two things - what the agent is allowed to do unsupervised, and how it's protected from prompt injection on the documents it reads. If the vendor can't answer both crisply, you've found your answer.

Tie the week together and a pattern shows. For two years the construction-AI conversation has been about what we could do with these tools - estimating, drawing review, document control. This week the frame flipped: the part of the system that approves our work is now adopting AI too, and at national scale. Extract is live everywhere; APD is being tested on real decisions; the delivery firms above us are building agentic capability with Google.

That changes where the effort goes. When a human planning officer read your application, messy inputs were forgiven - they'd squint at the scan and ring you up. When a model reads it first, structure and clarity stop being courtesy and start being throughput. The firms that win the next two years won't be the ones with the cleverest internal AI; they'll be the ones whose data is clean enough that other people's AI - councils', consultants', clients' - can read it without choking. Same discipline we keep coming back to: get the data right, keep a human signing off, and treat vendor and government savings figures as a hypothesis to test on your own jobs, not a promise.

A practical step: Pick your next live planning submission and run it past one test - could a machine extract every key fact from it without a human translating? If not, that's your cheapest improvement this quarter, and it pays off the moment the council's tooling catches up.

A short, honest note - and you can skip it if founder updates aren't your thing. This week we took PlanOps live on LinkedIn for the first time. No countdown, no launch theatre: just an introduction to what it does and an open invitation to try it on a real job. The reason it sits here rather than in an advert is that it's already running on live UK sites, on exactly the work this brief keeps circling back to - progress reporting, comparing what changed between drawing revisions, pulling a plan together from the documents you already hold. You tell the project what you need; it routes the request to the right specialist assistant and brings back a structured, review-ready draft for a competent person to sign off. That last part matters: it supports the human, it doesn't replace the judgement.