Construction AI: someone finally points the agents at the labour problem, and a cheap model gets cheaper until January

Here's the thing most construction AI has quietly ignored. Nearly all of it aims at the work, the drawings, the RFIs, the progress on site, and almost none of it aims at the one thing a UK contractor genuinely can't get enough of, which is people. So it's worth noticing when a tool turns the other way. Bridgit, a workforce planning outfit used across North America, launched AI agents for exactly that job, and it was picked up across the construction trade press in the week to 12 August 2026.

What the agents actually do is the resourcing graft a planner does by hand every Monday. They assemble a team for a project, spot the clash when two jobs both want the same site manager, flag where you're short and recommend who's free to fill the gap, and pull the reports together. Bridgit had already put out "Ask Bridgit" earlier in the year, the plain-English version where you could ask who's available and when. What's changed is that the agents don't just answer, they act on the planner's behalf. And there's a Bridgit MCP alongside them, which pushes live workforce data into the other AI tools a contractor's already running, rather than trapping it in one more dashboard. That MCP detail is the one I'd actually watch, because it's the same plumbing we talked about a week ago, now showing up in a real product instead of a spec.

The caveat is the usual one and worth saying plainly. Bridgit says it's trusted by 40 per cent of top contractors, which is a vendor's own figure, so treat it as marketing until an independent number turns up. And workforce planning is a narrow slice, not a whole platform, so this sits under tools rather than anything grander. But the reason it lands is human. On a UK site the binding constraint isn't the software licence, it's whether you've got a qualified person free on the week the job needs them. You can't download a bricklayer. A tool that helps you deploy the ones you've got, and does the rota admin that eats a planner's Monday, is pointed at the right problem. That's what it's about.

Worth doing: take one real resourcing decision you made last month by hand, the awkward one where two jobs wanted the same person, and see whether a workforce agent reaches the same call. If it does, you've found where it earns its keep. If it doesn't, you've learned that cheaply.

On 13 August 2026 Google released Gemini 3.7 Flash, and it matters to a builder for a duller reason than the benchmarks. Flash is the cheap, fast tier, the model you'd put behind the boring high-volume jobs, reading daily logs, sorting RFIs, drafting the first pass of a submittal, triaging a pile of site photos. This cycle Google tuned it for coding, agents and multi-step work rather than raw cleverness, and it kept the very large context window, which in plain terms means it can hold a lot of your project documents in its head at once. The gains it quotes are on software tasks, so read the headline numbers as a developer's story, not a promise about your quantity take-off.

The bit worth your attention is the price and the date on it. Introductory pricing is 75 cents per million input tokens and $3.75 per million output, which is half what 3.6 Flash launched at. Good news for anyone running document-heavy AI at volume. But it's introductory. On 1 January 2027 the list price doubles, to $1.50 and $7.50. I'm not going to pretend that's a scandal, it's a launch discount and everybody runs them. What it means for you, though, is simple. If you're picking a tool this quarter partly because the model under it is cheap right now, you're pricing on a rate that has a public expiry date. Build the process on what the model can do for you, measure it on a job you know the answer to, and assume the running cost roughly doubles in the new year. Then it's still a good decision when the promo ends.

The procurement filter: before you commit to anything priced on "AI included", ask what the underlying model costs and whether today's rate is introductory. The person who signs the renewal in January is the one who finds out the hard way.

Put the two together and there's a thread. The workforce agent helps you get more out of the people you have. The cheap model does enough of the admin that those people spend less of the day on paperwork and more on the work only they can do. Both are aimed at the same wound, which is that a UK contractor's scarcest thing is a competent person free on the right week, and every quarter the surveys keep naming skills and labour as the drag. Neither of these tools digs a trench or lays a course. What they do is stop your good people spending their afternoons on a rota spreadsheet or a stack of daily logs.

So the discipline stays the same as it's been all summer. Pick the job that eats your team's evenings, the resourcing puzzle, the log-writing, the report nobody wants to do, and put the tool on that first. Measure it against a week you already understand. And read the small print on price and portability before you sign, because the running cost and the exit terms are the two things a demo never shows you.

A practical step: name the one recurring admin task that most often pulls a qualified person off the tools this month. That's your first pilot, and the before-and-after is your proof.