When Anthropic shipped packaged legal skills into Claude, $285 billion was wiped from software stocks in a single day. Thomson Reuters down 16%. LexisNexis down 14%. LegalZoom down 20%. Not because of earnings. Because investors understood something the affected industries hadn’t yet: a general AI just upskilled itself into a professional domain overnight.
If law, why not specification review? If contracts, why not BIM coordination?
The software world had already moved. Twelve months of packaged knowledge bundles — reusable, shareable, installable from GitHub with a single command. Code review skills. Testing workflows. Documentation protocols. Not the automation of a task. Encoded expertise.
We kept talking about automating the model.
The cool output. The deliverable that photographs well in a client meeting. The thing AI does that the industry can point to. While the actual intelligence — the calibrated judgment of engineers with fifteen years of project failure behind them — kept walking out the door with every retirement.
One architecture firm. One school a year. Twenty years: twenty schools of accumulated knowledge. Why the acoustic spec on project seven failed. What school boards never mention in the brief. Which thermal detail always generates a change order two years later? That knowledge lives in email threads. In memory. In someone who might not be there next spring.
A skill isn’t a report. It’s an encoded judgment.
And unlike most competitive advantages in AEC, this one compounds. Every project encoded makes the next engineer faster. Every lesson captured makes the next mistake less likely. The firms that start now aren’t just retaining knowledge — they’re building a moat their competitors can’t see yet.
And we now have the infrastructure to build internal libraries of it. To extract what engineers know from decades of project work. To encode it into something an agent can actually use. To chain those agents into workflows, a new hire can query on day one — not a lessons-learned PDF buried in a folder no one opens, but an agent that knows what your firm knows.
The research is sobering: human-curated skills improve AI agent performance by 16 percentage points. Self-generated ones barely register. You can’t automate the capture. The expertise has to come from inside — from the people who carry it, while they’re still in the building.
The industry has been losing knowledge for decades.
Now it has the tools to stop it.
The question is whether it starts before or after the people who could fill those libraries are gone.
