Construction AI: someone finally draws the plumbing agentic BIM actually needs, and who gets to plug an agent into your data

For a year the agentic-BIM debate has been stuck on the same loop: a vendor shows an agent reading a model, the room nods, and nobody asks the awkward question. So it's worth flagging that AEC Magazine's current May/June 2026 issue has stopped asking whether agents are coming to BIM and started asking what they need to actually work on a real job. That's the more useful question, and the answer is gloriously unglamorous.

What the piece argues is that a proper agentic platform, what it calls runtime-native, would sign every solver output with a proof that the relevant constraints were honoured, and version those proofs so you can audit them later. Not a log of clicks. A signed record of what the agent changed, against which rules, that holds up after the fact. It adds graduated autonomy, the idea that you hand an agent a little decision-making at a time rather than flipping it from off to fully-trusted overnight. And there's one point most vendors walk straight past: the platform should publish provenance for the foundation models it depends on, because a defect in one of those models is a defect in every project the platform has ever touched. The article reaches for a Google DeepMind paper on AI delegation to frame what delegated authority should even mean. The comparison only goes so far, but think of it like signing off a subcontractor's work, you don't just take their word that it's right, you keep the record that says who checked it and against what.

I said in yesterday's brief that the model keeps getting cleverer while the thing you'd plug it into stays a filing cabinet. This is the first half-decent sketch I've seen of what the upgrade actually looks like. I'm not convinced any shipping platform is close to this yet, and a current-issue think-piece is a long way from a product you can buy. But the direction is right, and it gives you something concrete to hold a vendor to.

The procurement filter: Next time someone pitches you an agent for your model, ask one question. Can it show me a signed, versioned record of what it changed and the constraints it checked against? If the honest answer is no, it's a clever demo, not something you put near a Gateway 2 submission where the paperwork is the job.

Here's a story that reads like an IT release note and is actually a governance decision. On 18 June 2026, a fortnight back now, Anthropic shipped enterprise-managed authorisation for its MCP connectors, with Okta as the first identity provider. In plain terms: an admin approves a connector once, and everyone in the organisation inherits access automatically through the identity groups and roles they already have. No per-person consent screens, no queue of OAuth tickets. Seven providers are supported at launch, Asana, Atlassian, Canva, Figma, Granola, Linear and Supabase, with Slack named as coming next (vendor list).

Why does a construction firm care about connector plumbing? Because connectors are the doors an agent walks through to reach your systems, and this changes who holds the keys. What it means on the ground is that whoever runs your identity provider now decides which agent can reach your CDE, your finance ledger, your document store. That's a chunk of your AI risk surface, and it's just moved onto one person's desk, probably someone who doesn't think of themselves as making safety decisions. The same logic the AEC Magazine piece applies to the model, namely prove what it did, applies one layer up to access: prove who let it in. MCP is the wiring most agentic construction tools will run on, so this isn't a Silicon Valley curiosity, it's the substrate your future copilots sit on.

I'd not overstate the urgency, most UK firms aren't wiring agents into live project systems this quarter. But the firms that will are the ones with the cleanest data, and they're exactly the ones who should decide the access model on purpose rather than discover it later.

For your board pack: Name the person who owns connector and identity decisions for AI, and make sure they know it's a risk decision, not a convenience one. If that person is your IT manager and nobody's told them they're now guarding the data layer, that's a gap worth closing this month.

Strip the week down and you get one idea twice. The agentic-BIM piece says you can't trust what an agent did unless the platform can prove it. The connector story says you can't trust an agent's reach unless you can prove who let it in. Both are versions of the same thing every good QS already knows: keep the record, or it didn't happen.

There's a third leg, and it's the one UK construction keeps tripping over. You can't prove an AI tool saved you anything if you never measured the baseline. The RICS Construction Productivity Report 2026, drawing on nearly 3,000 professionals, found 22% of UK firms never track productivity at all, the lowest measurement rate of the regions surveyed, with the UK's own productivity net balance sitting at +18%. So before you pilot an agent on your next job, write down what the task costs you today, in hours and in pounds. That's not glamorous and it won't make a conference slide. But it's the difference between an AI pilot that quietly disappears and one that turns into a number your finance director believes. That's what it's about.

A practical step: Pick one repeatable task, takeoffs, RFI responses, O&M compilation, and measure it cold for two weeks before any tool goes near it. The baseline is the asset.