A subtle change of mood is running through corporate AI, and it can be summed up in two words that vendors are suddenly hearing a lot: prove it. The first wave of enterprise adoption was powered by fear of missing out. Companies bought AI because a rival did, approved pilots on the strength of a slick demo, and cheered internal dashboards showing usage climbing. That era is closing. Executive patience is thinning, IT budgets are tightening, and the burden of proof has moved decisively from the buyer, who used to ask "can we try this," to the vendor, who now has to answer "what did it earn."
The clearest symptom is a backlash against measuring success by activity. Analysts have started using the word "tokenmaxxing" for the habit of treating raw AI consumption as an achievement, often pushed by adoption targets or leaderboard-style contests between teams. The problem is that tokens burned are an input, not an outcome. Amazon reportedly shut down an internal AI leaderboard, and Uber capped employee AI spending after staff blew through an annual budget at speed. Vercel's gateway data hints at the correction already underway: token volume rose 29 percent in June while spend rose 27 percent, with price per token flat, a sign that firms are routing cheap models to summarising and reserving expensive reasoning models for the decisions that justify them, rather than simply consuming more of everything.
In place of usage, the new currency is a defensible dollar figure. Consider Infinity Loop, a negotiation-intelligence platform that pitches itself squarely at this moment. It claims a financial-services client identified 40 million dollars in opportunities across 531 million in spend, and a pharmaceutical company found 11.6 million across 350 million in outsourced research. Whether those exact numbers hold, the framing is the point. Negotiation is a shrewd place to make an ROI claim precisely because the value is easy to bound: you know what you spent, what you paid, and what better information was worth. That auditability is exactly what is missing from fuzzier categories like AI-assisted coding or internal knowledge search, where causality is contested and credit is hard to assign.
There is a second cost the honeymoon glossed over, and it arrives every time a model is upgraded. A report from the services firm Straive argues the next phase of enterprise AI is "not about deploying models" but about managing their constant evolution. Swapping one model generation for the next is nothing like a routine software patch. A new release can improve reasoning and context windows while quietly changing pricing, output quality and behaviour, forcing teams to revisit prompts, evaluation datasets, guardrails and workflows. And consumption pricing keeps climbing: Straive notes that moving across recent Gemini Flash generations pushed output-token pricing from 40 cents per million to 2.50 dollars and then to 9 dollars. The model got better. So did the bill.
None of this signals that enterprise AI is faltering; spending overall is still growing, and the market is expanding rather than contracting. What is ending is the permission to spend without accounting for it. The organisations that thrive in this phase will not be the ones with the highest token counts or the flashiest pilots. They will be the ones that can point to a specific process, a specific number, and a specific reason the AI is still switched on. The technology has graduated from novelty to line item, and line items get audited.