AINEWS
Enterprise buyers are capping per employee AI budgets and forcing vendors onto usage based billing. The repricing tests whether AI vendors can defend per query unit economics.
The bill for the generative-AI experiment is arriving, and it is landing in the same place every enterprise software bill has landed: on the procurement line item.
For more than a year, generative-AI tools felt free to the people who actually used them. Employees opened ChatGPT, Claude, Cursor, and Copilot and let the meters run, often because their employers told them to. That cost is now showing up in corporate finance, and the response is the one every CIO recognizes from past software cycles: caps, dashboards, and a hard pivot from "use as much as possible" to "show me what we are actually buying." The question is not whether AI will survive the correction. It is whether the per-query price, the new unit of account underneath the whole market, was always the real product.
A leaked internal meeting at Accenture, reported by 404 Media's Joseph Cox, made the buyer-side concern explicit. Justice Kwak, Accenture's agentic AI strategy lead, told colleagues that internal data showed non-engineers, not engineers, were driving the company's token consumption, with PDF-to-slide-deck conversions among the most expensive workflows. The leak confirmed what finance teams had been sensing: the bill is not coming from the developers the tools were built for. It is coming from everyone else.
A "token" is the small chunk of text, usually a few letters or a partial word, that an AI model reads and writes. Vendors price against it: every prompt and every reply is metered, and the bill scales with how much the model has to read and generate. When a tool converts a fifty-page PDF into a slide deck, the model reads the entire PDF, reasons over it, and writes the deck. That single task can burn through more tokens than a week of ordinary chat.
The pricing surface is moving with the spend. GitHub announced in April 2026 that all Copilot plans would switch to usage-based billing on June 1, 2026, replacing the old "premium request unit" (PRU) with GitHub AI Credits that meter input tokens, output tokens, and the cheaper cached tokens the model does not have to re-read. Plan prices stay the same: Pro at $10, Pro+ at $39, Business at $19, Enterprise at $39. Heavy users now pay a variable surcharge on top. GitHub cited the same agentic, multi-hour coding sessions the Accenture data described as the reason flat-rate plans no longer pencil out, and added admin budget controls at the enterprise, cost-center, and user level.
Uber is the cleanest case study of what happens when those controls arrive too late. According to TechCrunch, the company burned through its entire annual AI budget in roughly four months after a "use AI as much as possible" mandate and internal leaderboards pushed employees to run Claude Code, Cursor, and similar tools without limits. The fix was a hard cap of about $1,500 per employee, per agentic tool, per month, tracked on a dashboard the CFO could read in real time. Uber COO Andrew Macdonald then publicly questioned whether the usage had produced any new consumer features, an unusually blunt admission from a buyer who had just spent the budget.
TechCrunch and The Information describe the same pattern spreading across Meta and other large enterprises through the second quarter of 2026: from "tokenmaxxing" inside teams to "token rationing" from the top. The vendors are not passive in this. GitHub's redesign is the vendor version of the same correction: if the unit of account is the token, then the product has to be priced against it, and the admin controls have to exist for the buyer who is about to receive a larger bill.
The honest read is that the AI industry underpriced its own product. Early users got free or cheap access in exchange for behavior data, the same way cloud credits subsidized the last generation of infrastructure startups. The bill was always going to arrive. What is being tested now is whether vendors can defend the unit economics of a per-query product against buyers who have learned, very quickly, what every previous enterprise software cycle taught: the procurement surface is where the price is really set.
Watch item: GitHub's June 1 switch from flat-rate plans to usage-based billing is the first major vendor repricing of the cycle. The next test is whether the per-employee caps at Uber and Accenture hold, or whether the dashboards become a budgeting tool that gets re-flooded the next time a CTO promises that more usage will produce more product.