Work backwards from the Amazon figure and the size of the blind spot becomes legible. A $1.8m bill at 860% over budget implies an approved budget near $187,500, since 860% over plan means 9.6 times plan [4][1]. That leaves roughly $1.61m unbudgeted [2], and if it accrued evenly across the five months before anyone noticed, about $322,000 a month cleared without a flag [5][3]. The Financial Times account, as relayed in the Forbes piece, does not separate the total from the overrun, so treat that split as an estimate rather than a finding [4].
Uber's numbers describe the same mechanism from the other end. Burning a full-year AI budget in four months is a run rate three times plan [6][4]. The response its CTO describes, a $1,500 monthly ceiling per employee per agentic coding tool, works out to $18,000 a year per person per tool [7][5], which is 75 times Anthropic's $20 monthly seat [6]. The usage line, not the seat price, is now the number worth fighting over.
The reason this is an accounting problem and not a procurement one is the identity itself. Seat-based cost was a function of headcount, a number finance already holds; consumption cost is seat fees plus model usage across users, applications and agents, none of which finance owns [8]. The author's sharpest observation is that a developer reaching for the most capable model on routine debugging, multiplied across thousands of users and millions of requests, turns a default setting into a seven-figure decision [9]. Vendor-side caps help, but they do not reconcile spend across providers, and each coding agent that opens its own provider connection arrives with its own bill [12][17].
The piece is by Viktoria Semaan, an AI educator and advisor at Databricks [10], and the gateway pattern it recommends is illustrated with Databricks' own Unity AI Gateway [11]. That is worth holding in view. The two loss cases, though, come from elsewhere: the FT for Amazon [4] and Uber's own CTO, in remarks the source notes are paywalled [6]. And the pricing change is not an argument, it is a published price: a $20 seat with no usage allowance, with Claude, Claude Code and Cowork billed at standard API rates [1][2]. The remedy comes from a vendor, but the exposure it addresses was documented by other people, at the FT and at Uber.
The cheap part of the fix is metadata rather than software. Tags for department, project and environment have to be standardized before usage grows, because they are difficult to reconstruct afterwards, and a usage report without them shows totals but not which project drove them [13]. The same logic applies to the value side: the argument is to measure at the workflow level against a pre-AI baseline for speed, quality and unit cost, then weigh the gain against the spend that produced it [14]. This quarter the choice is whether attribution exists at all, with visibility and alerts before hard limits [15]; next quarter the choice is whether anyone can defend the spend at renewal, and that defense depends on the attribution existing first. One thing the record does not give us is what Anthropic's previous fixed seat cost, which makes the size of the repricing, as distinct from its shape, unknown for now [18].
