Executive Briefing: 95% Adoption, and You Still Can’t Prove One Token Helped

Source: Nate’s Newsletter Date: 2026-06-07 URL: https://natesnewsletter.substack.com/p/ai-token-cost-management

Summary

A Nate’s Newsletter piece framing AI cost management as a labor management problem, not a spending problem. The central case is Uber: 95% of engineers used AI tools monthly, an internal agent produced ~1,800 code changes per week, and the company burned its entire annual AI budget ahead of schedule — yet the CTO could not connect any of it to better customer outcomes. The article argues this failure is not about spending too much; it’s about the absence of operational systems to route work intelligently and attribute output to value.

Implications

• AI-adoption-ROI discourse. The Uber figure is the loudest public data point yet that adoption rate and outcome accountability are orthogonal. 95% monthly active usage with zero provable customer impact is the reductio ad absurdum of “AI transformation” as a headcount metric. Expect this stat to circulate widely in enterprise procurement and CFO circles.
• Agentic-engineering patterns. The “minimum effective intelligence” routing principle — frontier model for hard tasks, open model for routine tasks, no model where humans suffice — is the operational response the article proposes. This is the same cost-tiering logic driving model-router products; seeing it framed for a general management audience signals the concept is crossing into mainstream enterprise thinking.
• Small-model economy. The implicit argument is that unconstrained access to frontier-model compute is the source of the budget problem, and structured routing to cheaper models is the fix — a direct tailwind for smaller, cheaper, task-specific models.