The Taste Bottleneck

For two years the AI story has been about capability: how much of the job can the model do. This week the story turned. Across an AI essayist, a brand strategist, and a credit analyst, the interesting claim is no longer what AI can produce but what humans still have to decide. The bottleneck moved upstream, from execution to judgment, and the people closest to the work are the ones naming it.

Dan Shipper, speaking publicly, put the shift most cleanly: “AI can tell us what is probable, but it cannot tell us what is worth wanting.” He frames the coming decade as “wisdom work” replacing knowledge work, arguing that “as intelligence becomes abundant, value will move again. This time towards the heart.” Strip the poetry and it is a supply-and-demand claim. When the marginal cost of competent output collapses, the premium attaches to whatever the model cannot supply, which is a point of view about what to build.

You can hear the same claim, in commercial dress, from a brand strategist. “AI makes mediocre ideas irresistible,” she said, and “now that tasteful but forgettable brands are cheap and easy to generate, that tyranny is more oppressive than ever.” Her line “agents are not loyal, they are rational actors” completes the frame: agents will optimize for whatever objective they are given, so the scarce input is the objective itself. Cheap generation does not raise the ceiling on taste. It raises the floor on everything else, which is the same thing as making taste the binding constraint.

Noah Brier gives the engineering version. “The most insidious failure mode in agentic engineering isn’t buggy code. It’s agents building fundamentally misaligned features, products, and systems.” His prediction that “the company with the best clock may soon beat the company with the best model” is another way of saying the model is no longer the moat. Coordination and intent are. A second builder in the same conversation notes that LLMs remain “quite limited in their ability to perform tasks that require ambiguity, open-endedness, and creativity,” which is roughly the definition of the work that is left.

The credit desk is watching the same thing from the opposite end. A high-yield credit analyst said that what he did as an analyst seven years ago has been “reshaped by AI where AI can do 90 to 95% of it.” He does not read that as junior analysts being doomed. He reads it as a threat to mid-career people “10 years into the industry” who assume they know enough and will not learn the tools. That is the wisdom-work claim restated as a labor-market claim: the residual 5 to 10 percent, the judgment layer, is where the career now lives, and seniority does not automatically confer it.

The through-line is a quiet inversion of the 2023 thesis. The bet then was that whoever had the best model would win. The bet now, according to the people actually shipping and allocating, is that the model is a commodity input and the scarce complement is knowing which problem is worth solving. Sumit Singh’s warning that “founders who AI existing workflows will lose” is the corollary. Automating a workflow that should not exist just produces the wrong thing faster.

Watch for the second-order effect nobody has priced yet. If taste and judgment are the binding constraint, the org chart inverts. The premium moves to the person who writes the brief, not the person who executes it, and to the investor who picks the objective, not the operator who hits it. Julien Bek’s line, “do not mistake an outlier operator for an outlier founder,” lands differently in that world. In a wisdom-work economy, outlier operators are exactly what AI is about to make cheap.