Construction AI: Whitehall scraps the department behind the data centre push, and Google ships a cheaper agent model while its flagship stays missing

On 21 July the government abolished the Department for Science, Innovation and Technology as part of a wider machinery-of-government reshuffle the Prime Minister confirmed that day. DSIT was only three years old, set up in 2023 to put science and technology at the centre of government. Its work now splits three ways. Science, innovation and technology move into a new Department for Business, Innovation, Science and Trade under Jonathan Reynolds. AI policy and public sector AI adoption move into the Cabinet Office. Online safety goes to the culture ministry, and, in the detail that tells you how far the knife went, the Public Sector Fraud Authority ends up at the pensions ministry.

The headline is that AI gets a seat at the top table. Kanishka Narayan, appointed AI and online safety minister back in September 2025, becomes the first UK AI minister with the right to attend Cabinet, working jointly across the Cabinet Office and the new business department. A new AI Taskforce is being stood up inside a new Office for the Prime Minister and the Cabinet. So AI has gone up in the world. The question for our sector is what happens to the plumbing while the nameplates change.

Here's why this lands on a construction desk and not just a policy blog. DSIT owned the machinery that sits directly under the UK data centre boom: the Compute Roadmap and its 6GW-by-2030 target, the Sovereign AI programme, and the AI Growth Zones that were created specifically to fast-track data centre planning and grid connections. It also owned the planning-AI work aimed at construction output, including the Extract tool rolled out to English councils in June to digitise old planning records, and the AI planning assistant the government has been trailing to cut decision times on the 1.5 million homes target. That whole portfolio is now being handed between a new department, the Cabinet Office and a taskforce that doesn't have its furniture yet. I've sat through enough reorganisations to know the pattern: nothing gets cancelled, but everything gets slower for a quarter or two while people work out who signs what. If you're a developer, a contractor or a consultant with a data centre consent, a growth-zone designation or a DCO timetable riding on this, that delay is your risk, not Whitehall's.

For your board pack: list every live approval, funding line or fast-track designation your projects depend on that traces back to DSIT, and next to each one write the name of the department or office that now owns it. If you can't fill in the second column by the end of the month, that's the gap to chase.

On 21 July Google released three models at once: Gemini 3.6 Flash, a smaller 3.5 Flash-Lite, and 3.5 Flash Cyber, a version fine-tuned for finding and fixing security vulnerabilities. What Google is selling here is not a smarter model, it's a cheaper and leaner one. It says 3.6 Flash gets through multi-step agent tasks in fewer reasoning steps and tool calls, uses about 17% fewer output tokens than 3.5 Flash, and lands at $1.50 per million input tokens and $7.50 per million output, down from $9 on the output side. Those are Google's own numbers, so treat them as a claim to test rather than a benchmark you can bank. The genuinely useful change is quieter: the knowledge cutoff moves from January 2025 to March 2026, and the model was in GitHub Copilot on day one.

Why a Flash release matters more than a flagship one, for us at least, is about where the money goes. The heavy AI cost in construction isn't the occasional clever answer, it's the boring high-volume work: reading a set of drawings, extracting quantities, drafting and checking RFIs, sorting a document dump into something filed. That work runs on Flash-class models and it bills by the token. So a real cut in output tokens per task changes the unit economics of a pilot in a way a leaderboard score never does. The comparison only goes so far, but it's a bit like the difference between a van's headline top speed and its miles per gallon. On a fleet doing the same run every day, it's the fuel figure that decides whether you can afford to keep it on the road.

The other half of the story is what didn't ship. Gemini 3.5 Pro, the flagship I've now tracked through three missed dates, still isn't out, and Google used this launch to tease Gemini 4 instead. Holding back a model you're not happy with is the right call, and I'd rather a lab did that than push something shaky into a compliance workflow. But it's another reminder that the tool you can build on is the one with live API docs today, not the one in the keynote.

The version check: if you're running agents on project data, write down which model and version each workflow actually calls, and revisit it when a cheaper one lands. A 17% token saving is real money at volume, but only if someone is allowed to switch the pipeline over and test it.

Here's the item that won't make a headline on your site but will show up in a contract you sign this year. The big construction platforms have started fencing off data, and they're doing it because the same project data that runs your job is the fuel that trains AI agents. On 30 September 2025 Procore rewrote its API terms and published a new Developer Policy that bans bulk downloads of construction data for training large language models. It then removed Trunk Tools, an AI agent firm used by large US general contractors including Gilbane and Suffolk, from its API, and refunded the booth Trunk Tools had booked at Procore's Groundbreak conference. A few weeks later Procore completed its acquisition of Datagrid, a rival agentic-AI provider, on 20 January 2026. Protecting customer data with one hand, buying an agent company with the other. Procore frames the API move as security. I'd read it more plainly as a platform defending a moat, and there's nothing wrong with saying so out loud.

Trunk Tools answered in June 2026 with Cortex, pitched as an intelligence layer trained over years on real general-contractor projects, built to connect over the systems where the data already lives, including Procore and Autodesk Forma. So the battle line is drawn: platforms deciding whether your data can train their agents and which outside agents are allowed to touch it, and specialist firms trying to sit on top of the data wherever it sits. Most of the named players here are American, and I'm not going to pretend a UK regional contractor is in the room for those decisions. But the terms of service are the same terms of service, and the lesson travels without a passport.

What that means on the ground is that the data-rights clause has stopped being boilerplate. When you buy a construction platform now, the questions that decide your future flexibility are who can train on your project data, whether an agent you choose can get API access to it, and whether you can export your own records in a usable form if you leave. Those aren't IT questions. They're the difference between owning your golden thread and renting it back from whoever holds the platform.

Worth doing: on your next software renewal, get someone to answer three questions in writing before you sign: can the vendor train on our project data, can an AI agent we choose reach it through the API, and can we export it in full on exit. If the answers are vague, that's your answer.