Diagnose every AI overage case individually and apply the remedy that case calls for - without squashing AI usage

Situation

Facing an AI token run rate heading toward roughly 3 million dollars a year, with the top ten spenders accounting for about 95 percent of it, Peter refused both a spend cap and a single blanket fix. He directed each manager to work out why the overage is occurring for their own people, case by case, and to do whatever is appropriate for that specific cause - with the standing constraint that the answer must never be to suppress AI usage. He gave different answers for the cases he had already looked at himself: one engineer at roughly 60k a month is justified and was told to keep going; two others at roughly 100k a month between them are low-return and their manager was already digging in; Zorina team burned 12k in August purely because they were hitting the Claude API directly rather than the enterprise plan, so Justin Haynes action is a plan switch. Peter also declined Ryan Smith push to have the AI committee issue a company-wide best-practice policy on spend.

Reasoning

PETER CORRECTED MY HYPOTHESIS HERE. I had read this as a structural fix - move people to the enterprise plan instead of capping. Peter: it was figure out why the overage is occurring for each individual case, and do what is appropriate, but do not squash AI usage. The plan switch was the right remedy for one case, not the decision. The decision is the diagnostic rule plus the constraint. A cap or a uniform policy would answer all four cases identically when the causes are genuinely different - one is a correct spend, one is a performance problem, one is a billing-mechanism mistake - and a uniform answer would also throttle the usage Peter actively wants. Hence transparency and per-case diagnosis rather than a rule: I do not want to force tools, I want to force transparency and asking people to think.

Additional Context

This extends the 2026-08-24 decision to govern AI spend by notification and trust rather than a cap. Peter spent the week running the per-report walkthroughs that decision implied - Wolford, Nathan, Ryan and Justin each got the same dashboard and the same ask. A live data-integrity problem surfaced during it: Ryan Ramp figure showed 2,237 dollars for 30 days against 17,000 on Peter chart, while the numbers Moody and Wallace produce reconcile against the actual Anthropic bill.

Observed Evidence

Peter to Nathan about the justified spender: not asking it to stop, just asking it to be smart. Peter to Ryan about the same person: my strong feedback is keep going. Peter to Ryan about one of the low-return cases: that was worth zero dollars. Peter to Justin about Zorina: the cause is direct API use, the fix is the plan. Four cases, four different answers, one rule.

Matching Patterns

Confidence Breakdown

Reasoning Depth Analysis