The people who laid the transcontinental tracks knew how a railroad worked. Not from a manual. From laying the tracks.
But that is the part we are about to lose.
Everyone gets promoted at once
The pitch for agentic AI is that your people move up a level. Stop chopping onions, start writing the recipe. Or stop writing the code, start writing the spec. And stop doing the work, start directing it.
It is a good pitch, and every organization I talk to is running some version of it.
So here is what nobody costs in. The line cook did not become a head chef by being told to. She became one by chopping onions for six years and watching what happened when the recipe was wrong.
Competence was a by-product
Three versions of this problem keep surfacing, and they are the same problem wearing different clothes.
The manager approving the agent’s output has never run the process herself, so she cannot recognize a bad output. The workflow you are automating was never the real workflow, because the real one included an accountant who quietly fixed a backward date every month and never told anyone. And the person authoring the spec learned what makes a spec good by writing bad ones and spending the night fixing them.
Approval judgment. Informal repair. Specification instinct. Your operating model depended on humans having done the work, and it never recorded what that dependency was supplying.
It never had to. The work produced the judgment for free.
This is not disruption. It is revelation.
So, nothing new broke. The dependency was always there. Automation removed the thing that was quietly paying for it, which is the only reason we can see it at all.
And notice where it lands. The whole demand-side playbook, the one I keep making the case for, rests on a human who can author a good spec and govern the outcome. That is still the right answer. However, it is also the one competence we have just stopped manufacturing.
The wrap-up
The real AI competence problem is not whether the system can perform the work. It is whether the organization can still specify, judge, and repair it.
This is not an argument for slowing down. It is an argument for being deliberate about something that used to happen on its own.
If specification and oversight are the durable human roles, then producing people who can do them is a design problem now, not a career-progression accident. Deliberate practice, retained execution, something. But chosen, not assumed.
Otherwise you end up with a railroad that runs beautifully, and nobody left who knows how to lay track.