A headcount plan is not a workflow

The radar’s AI feed carried a Reuters investigation this week on how Mark Zuckerberg’s plan to replace Meta staff with AI imploded. A few lines above it, at 577 points and 362 comments, sits a repository called OpenExecutive: developers whose CEO fired developers to make room for AI, answering with an open source AI CEO. Two versions of one argument, and only one of them is funny on purpose.

Grant the strong case first. Substitution is real. Work that took a person an afternoon does sometimes take a model a minute now, and pretending otherwise is not caution, it is nostalgia.

The vault’s work notes put the emphasis somewhere else, and I think they are right. AI changes work most when it changes the workflow, not when it produces a better single answer. When building gets cheaper, the bottleneck moves downstream into reviews, prioritization, approvals, and deciding which shipped version is worth keeping. Adoption holds when agents appear in the surfaces people already work in and somebody can watch the output quality.

A replacement plan does none of that. It treats a role as a pile of outputs, removes the pile, and assumes the outputs keep arriving. It never asks which loop got faster, where the new queue formed, or who reads the doubled volume of work that now shows up needing a decision.

The joke repository is aimed at that gap with reasonable accuracy. If a job is just outputs a model can generate, executives make an unusually clean test case, and the people writing the memo are never in the sample. Nobody thinks that is a serious product. It is a fair question about who gets counted.

Announcing a substitution takes an afternoon. Finding out what it did to the review queue takes about three quarters.