A Box sourcing manager calls for a new procurement discipline to govern AI token spending

The mechanism here is a missing denominator. A cloud bill and a token bill both grow with use, but cloud consumption attaches to infrastructure somebody provisioned and can point at, while token consumption attaches to an individual deciding to ask a question. Waditwar's formulation is that millions of everyday interactions drive the increase, not any single event ^[5], which locates the cost in behaviour rather than in an approved deployment. The approval gate sits at the contract, and the purchasing happens downstream of it, daily, by people who have never been shown a price.

The tradeoff is worth naming before anyone volunteers to own it. Governing consumption means metering it, per team and per workflow, or capping it. Metering costs instrumentation work and hands cost visibility to staff who have no basis yet for reading it; capping suppresses the process and operating-model redesign that McKinsey's State of AI associates with the largest value, on the author's summary of it ^[9]. Whoever owns the token line inherits two measures that pull against each other, and the org chart has to say which one wins.

This is a sourcing manager arguing that sourcing matters more than it used to, on a council platform that prints her employer beside a note that the views are her own ^{[1] [10]}. Point taken. The mechanism does not depend on her motive, though; it depends on whether your own AI line has a cost per unit of work that someone can quote. The evidence in the post is thinner than the diagnosis: it points to Deloitte and McKinsey directionally and puts no figure on how large a typical overrun is ^[13]. We do not know the magnitude, and treating an account drawn from conversations as a measured finding would repeat the error the post is complaining about.

The board-deck version - spend is up because adoption is up, and adoption is what was asked for - leaves out the one number that would explain the bill: a denominator. Deloitte's finding, as the post reports it, is that moving from pilots to scaled deployment makes governance, risk and compliance and demonstrating business value the critical priorities ^[8], which is the polite phrasing of the same problem. At pilot scale nobody needs a cost per outcome; at production scale the absence of that number is what the finance review finds.

Sequencing is the part operators can act on. On the analogy the post itself uses, the discipline arrives after the bill: cloud consumption came first and FinOps came second ^[6], and the post credits only two earlier transitions with producing a named management practice at all ^[14]. That pattern plays out over a decade, and a budget cycle moves much faster than it does. What this budget cycle can settle is the smaller question of who gets measured on consumption, because the hard questions the post identifies (governance, cost, accountability, and whether a use case justifies continued investment ^[15]) each need a name attached before anyone can answer them. Leave the name blank and next year's planning opens with the same question in the same words ^[3].