Anthropic reported that more than 80% of the code it merges (opens in a new tab) is now written by Claude. A company whose entire business is judgment about AI has largely automated the doing of its core craft, and the humans are busier, more essential, and more accountable than ever. That is not a paradox. It is the whole point.
Here is the mechanism. Speeding up one part of a system doesn’t make the system faster; it just relocates the bottleneck (opens in a new tab). When the doing gets cheap and fast, the constraint moves to the thing upstream of the doing: deciding what to do, whether to do it, and how you’ll know it was right. Call that the judgment layer.
The electricity tell
We’ve seen this movie. When factories electrified (opens in a new tab), the machines that mattered weren’t the motors. They were the reorganized lines (opens in a new tab) that the motors made possible. The bottleneck moved from “power” to “the design of work.” AI moves the bottleneck from “execution” to “judgment.” The organizations that win are the ones that invest in the new bottleneck, not the old one.
What the judgment layer actually contains
- Direction. What’s worth doing at all. Agents are indifferent to strategy; humans are not.
- Boundaries. Where the machine is trusted, where it’s checked, where it’s forbidden. (This is why graded governance (opens in a new tab) matters: the leash should match the task.)
- Accountability. Someone owns the outcome (opens in a new tab). You cannot delegate that to a model.
- Taste. Knowing what good looks like well enough to reject confident, plausible, wrong output.
The uncomfortable implication
If the doing is cheap, then being good at the doing is no longer where your value sits. That’s threatening for a lot of expert roles, and it should be. The move is not to out-type the machine. It’s to climb into the judgment layer, to become the person who decides, bounds, and owns, because that is the part that stays scarce.
What to take into the room: for the work your team does, name the judgment layer explicitly. Who decides? Who bounds? Who owns the outcome? If those answers are vague, automating the doing will just let you make mistakes faster.