McKinsey published a piece on AI in AEC this month — How AI is Reshaping the Future of the AEC Industry. If you've been following this thread, the main argument will feel familiar: competitive moat comes from data and workflow ownership, not software features; adoption without process redesign stays capped at individual productivity; vendor data rights are being underread in most contracts. Three independent sources saying the same structural things in the same month is worth noting - even if McKinsey says it in considerably more words.
But there's one paragraph in the piece that I haven't seen anywhere else in this conversation:
Many tasks that AI is automating now are what junior staff traditionally have used to build experience and judgment. AI could concentrate expertise within a smaller group of senior professionals, weakening the training pipeline as firms become more dependent on expert judgment.
The Judgement Stack conversation — Martyn Day's piece I previously wrote about — is largely about capturing the expertise of senior practitioners before it walks out the door. That's a real and urgent problem. But there's a downstream question it doesn't answer: if agents are doing the repetitive work that juniors used to do to build pattern recognition, where does the next generation of senior practitioners come from?
The learning mechanism in AEC has always been exposure through repetition. You absorb judgment by doing the work and having it reviewed — hundreds of times, across dozens of project types. If agents handle the bulk of that work, the exposure loop breaks. You end up with a stack built from current expertise and no reliable way to replenish it.
McKinsey's prescription is "structured reviews, simulations, explicit standards, and exposure to real project failure cases." True enough, but that describes infrastructure most firms don't have.
What makes the timing interesting is that Architecture Studio v1.4 shipped last week — the same open-source repo from ALPA that Martyn Day used as the anchor for his piece, and that I wrote about in If You Have Done it Three Times.
Federico Negro's update notes are worth reading directly. The repo has been cloned 3,000+ times since May. The conversations since have mostly been about memory and governance — which is exactly where v1.4 focuses.
The key changes:
A proper governance layer: Studio and Project entities that give memory and decisions explicit boundaries, hardened provenance that tracks where numbers and citations come from even inside custom skills, and data that continues to stay in your own Anthropic account.
A memory layer: automatic memories integrated with the new Studio/Project entities, improved architecture decision records, better Obsidian support.
And the one that's directly relevant to the McKinsey training problem: a /learn skill. The description is deliberately open: designed to improve onboarding for those new to Claude Code, but "can also be repurposed to teach just about anything inside your studio."
That's not a complete answer to the training pipeline problem. But it's a step in the right direction — a structured mechanism for turning the institutional knowledge you're capturing into something that can teach the next person who needs it. The stack doesn't just store expertise. With /learn, it has a mechanism for transmitting it.
The governance additions matter for a different reason. Harpham's piece in the latest AEC Magazine made the case that adoption without verification discipline is a liability problem. Hardened provenance — knowing where every number and citation came from, even in custom skills — is the kind of audit trail that makes AI-assisted professional work defensible. It's governance built into the tool, not bolted on afterwards.
McKinsey describes these problems at industry scale, in a report that will circulate around board tables and strategy retreats. Federico Negro ships code. The gap between those two things — between identifying that the training pipeline is breaking and actually building a /learn skill — is where most of the interesting work in this space is happening.