Ask the executive floor and the AI transformation is on track: strategy approved, licenses purchased, workforce "upskilled." Ask the teams and you hear something different — nobody's sure what they're allowed to use, the sanctioned tools are worse than the forbidden ones, and the one engineer who's genuinely good at this stuff is now a bottleneck for four departments. Both floors are telling the truth. The gap between them is real, and the standard prescription — more training — barely moves it.
Here's why. When only your best engineers can use a capability productively, that's rarely evidence the rest need courses. It's evidence the capability requires heroics — undocumented setup, unclear guardrails, personal API keys, tribal knowledge about what's permitted. Infrastructure teams have seen this movie: Kubernetes had a "skills gap" for a decade, and the organizations that closed it didn't send everyone to certification bootcamps. They built platforms that made the right way the easy way, and suddenly median engineers were shipping to production.
What the paved road looks like for AI
Concretely: sanctioned model access through a gateway, so nobody manages personal keys and finance can see spend. Guardrails encoded in the path — data classifications enforced, logging on by default, the compliance question answered once by the platform instead of nervously by each engineer. Templates for the common patterns — a RAG service, an agent with tool access, an evaluation harness — that work on day one. And crucially, the escape hatch: a documented way to go off-road with review, because a paved road with no exits just recreates shadow IT.
Do that, and "AI skills" stops meaning arcane setup knowledge and starts meaning what it should: judgment about where the tools help, what to verify, and when not to trust the output. Judgment develops fast when experimentation is cheap and safe. It never develops when every attempt starts with a permissions argument.
The workforce shape underneath
There's a second, quieter shift underneath: the work that remains scarce is cross-boundary judgment — people who can hold the platform, the data, the security posture, and the business problem in one head. Organizations increasingly rent that judgment at the moments they need it — a build, a review, a hard decision — rather than carrying it as permanent headcount, and the senior end of the market is reorganizing around exactly that. I'd know; contracting that judgment directly is the entire premise of my LLC. The teams that thrive pair rented depth with a platform that makes their own median engineers effective. The teams that struggle buy training for a gap the platform created.