Construction AI: the markup tool on every desk grows an agent, and a pricier tier to go with it

Here's the thing about how AI is actually going to reach most UK construction teams. It won't walk in the door as a new platform someone has to buy, roll out and train everyone on. It'll switch on quietly inside a tool they already use every day. And the clearest example landed on 30 July, when Nemetschek, the German group that owns Bluebeam, reported its half-year.

The headline for our sector was Bluebeam Max, the agentic AI now built into Revu. If you don't know Revu by name, your design office and your site team do. It's the PDF markup and takeoff tool that's been on nearly every AEC desk in Britain for years. Max folds Anthropic's Claude into it for natural-language automation, plus tools from Bluebeam's Firmus AI acquisition that flag design issues and scope gaps, and it links PDF markups back to the BIM model. Nemetschek says take-up has run ahead of plan since the global rollout in late May, with early users reporting time saved on drawing review, document analysis and quantity takeoff. That's the company's own account, so treat the enthusiasm as a seller's, but the use cases are exactly the grunt work a checker or estimator does all day. The Build segment that houses this grew 21.1 per cent in the half to 277.7 million euros, recurring revenue hit 95 per cent, and the group closed its largest ever acquisition, HCSS, on 1 July. This is where Nemetschek is putting its weight.

Now the bit an editor might cut. What Max really does is turn the AI on your desk into the top tier of a subscription. It comes in at about 590 dollars per user per year, against Revu Complete at 440 and Basics at 260. And the last version you could actually buy outright, Revu 20, reaches end of life on 31 December 2026, so anyone still nursing a perpetual licence gets moved onto the subscription this year whether they wanted to or not. Bluebeam ended perpetual sales back in September 2023, and the G2 reviews have been grumbling about the switch and the loss of multi-seat discounts ever since. So the agentic AI is real and useful, and it also arrives as another reason the annual bill climbs.

That's the case for the challengers, the broad AI-first platforms taking on the big incumbent suites, told from the incumbent's side of the fence. A challenger sells you published pricing, a tool you buy in the open, and a record that walks out with you when you leave. An incumbent sells you a better version of the thing you're already locked into, at a higher tier, with your old licence sunsetting on a date it chose. The comparison only goes so far, because Max is a genuinely capable feature and plenty of teams will happily pay for it. But if you're the person signing off the Bluebeam renewal this autumn, 31 December is a date worth putting in the diary now.

The procurement filter: Before you upgrade the whole office to Max, price the alternative. Ask what else does your markup, takeoff and document-check job, what it costs, and whether your marked-up drawings come with you if you switch.

While the vendors were reporting, Whitehall was quietly making the compliance load a bit lighter, and it's worth knowing what's changing before it lands. On 9 July the government published its response to the consultation on proportionality in building control for higher-risk buildings. The headline is that it intends to pull most works carried out solely within individual residential units in an HRB out of Category A, which means less Gateway documentation for the kind of flat-by-flat work that was clogging the system. Small-scale works in communal areas need more thought before anything changes there.

There's more coming in the next few weeks. From 1 September, certain telecoms works to higher-risk buildings drop out of the usual building control approvals, so drilling a small hole for fibre or fitting some rooftop kit stops needing a full sign-off. And the Cladding Safety Scheme opens to eligible buildings under 11 metres in England on 17 August, with an eight-week window to apply. All of this sits against a Building Safety Regulator that's finally shifting its numbers: 368 Gateway 2 decisions in the 12 weeks to 28 June, a 77 per cent approval rate, and existing-building remediation running at 85 per cent, well above target, per acting chief executive Charlie Pugsley.

So it's a genuine easing, and take it where it's offered. But read what it actually covers. The government is trimming the burden on minor works inside flats, not on the fire and structural decisions that the whole regime exists to catch. The temptation, when the paperwork gets lighter, is to let the discipline slip everywhere. Don't. What the easement does is free up your time and the regulator's for the things that carry real risk. Keep your Golden Thread tight where it counts, and this is a help rather than a loophole.

Tie the two ends of this brief together and you get the shape of the year. On 31 July DeepSeek released V4 Flash, an open-weights model under an MIT licence, and the number that matters isn't the benchmark, it's the bill: about 14 cents per million input tokens, which works out near three cents for a typical reasoning task, among the cheapest capable models going. Those are the lab's own numbers, so treat the exact figures with a pinch of salt, but the direction is not in doubt.

That's the reason Bluebeam can bolt agentic AI onto Revu and every other tool you own is about to do the same. When a capable model costs three cents a task, adding an AI feature stops being a moonshot and becomes a product decision. What that means for you is more buttons, arriving faster, at more tiers. And the thing that decides whether any of them actually help is the same thing it's always been: whether your project record is clean enough for the AI to work on. RICS said it a fortnight ago and it's still true. The brake was never the model. It's the messy data and the skills underneath. That's what it's about.

Today's action: Pick one workflow where you already have clean, structured data, drawing markups, an RFI log, a takeoff, and try the AI feature there first. It'll show its worth on a tidy record before it ever will on a messy one.