The Counting House

Steve wrote this one from the other side of the table. After a year of writing about his own software factory, he spent the autumn talking to banks and other large companies about where AI was actually paying off, and the answer kept being the customer-service floor rather than the engineering org. The essay is his postcard from those conversations: an introduction to a pattern he calls the Counting House, set beside the Wheelhouse so that each explains the other. It runs long (he admits to cutting 96% of his material and still landing at two Yegges), and it is the most practical thing he has published in years: autonomy tiers, gold sets, an orchestrator, and who in the company has to sign which row.

The middle is a working data-science lesson, built around four airline intents that all mention Denver and a confusion matrix with dollar amounts in the cells. Readers who have bounced off precision and recall before get a duck hunt, and it does the job. The sections people will argue about come at both ends. One is the claim that SDLC-side AI is still a struggle at most big companies while this quieter pattern is already working. The other is a closing turn to model welfare for clerks that never keep a memory past one case. Twenty-five illustrations, most of them cutaway buildings, carry a good share of the explaining.