The most expensive line in enterprise AI is the human

Enterprise AI's most expensive line item is no longer the model license. It is the human hours spent supervising the model. The shift shows up first in procurement math: cost per task, tokens per workflow, minutes of worker correction per hour of agent output. Raw model quality, the score every vendor chased for two years, is becoming a secondary input.

The new weight on cost and supervision gives vendors with a pre-indexed view of company data a structural edge. That is the doorway Glean's Tau launch and its 81% token-cost claim against Claude Cowork on Claude Sonnet 5 are walking through, as reported by ChannelLife. The Glean-published numbers, a 78% grader preference and roughly 70% fewer tokens, are vendor-run, and the chosen competitor is Glean-defined. The labor anchor is Glean Work AI Institute's survey of 6,000 full-time digital workers: 6.4 hours a week supervising and correcting AI, more time than workers spend using it productively.

That is the reusable mechanism. Pre-indexed data plus automatic model routing compresses tokens and shortens supervision loops. The benchmark can be matched or beaten. The procurement lens it teaches cannot be uninvented.