Generated by Codex with GPT-5

This summary covers The Economist’s June 20th, 2026 Business article listed in the contents as Big bills for tokenmaxxing and published under the headline Token reckoning.

Artificial intelligence is creating a new corporate headache: the more useful AI agents become, the harder their costs are to control. These systems do more than answer a prompt. They reason through several steps, call tools and sometimes create other agents, consuming large numbers of tokens - the units of text that AI companies commonly use for billing. As agents spread across the many software products used by a large company, small charges can compound into an enormous bill.

From Experimentation To Expense

Until recently, many companies encouraged employees to use AI as heavily as possible. Internal leaderboards celebrated the biggest users, and “tokenmaxxing” became a sign that a workforce was embracing the technology. That enthusiasm, combined with the rise of token-hungry reasoning models and agents, has driven a sharp increase in spending.

Data from Ramp, a corporate-card provider, suggest that its clients’ AI spending has risen thirteen-fold in a year. Uber said in April that it had exhausted its annual AI budget in just four months, while another company reportedly spent \$500m on tokens in a single month. The burden is uneven: among Ramp’s clients, the top 1% of spenders average roughly \$7,450 per employee each month, compared with a median of only \$11. For now, technology firms face the greatest exposure because they adopted AI early and use it heavily for software development.

A More Selective Approach

The heaviest users are replacing enthusiasm without limits with basic cost discipline. Meta and Amazon have dropped token-use leaderboards. Companies are choosing models according to the difficulty and value of each task instead of defaulting to the newest system. A less advanced Anthropic model can cost one-twentieth as much as a frontier model for some jobs, and an open-source Chinese alternative can cost one-twentieth as much again.

Spending caps are another tool. Uber limits employees to \$1,500 of tokens per month for each coding service. Other firms are directing larger allowances toward roles where AI contributes most directly to the core business, such as engineering at a technology company. Software suppliers are also adjusting. Intercom charges for customer-service questions only when its AI agent resolves them, while cloud providers now offer budgeting tools and systems that route requests to an appropriately priced model.

The Bill Has Not Settled

These measures do not eliminate the underlying tension. AI labs want customers to use more tokens, but their services are still subsidised as they compete for market share. OpenAI may cut prices sharply to win business from Anthropic, yet both firms will eventually need sustainable profits - especially if they proceed with public listings. That could force prices back up just as agents become embedded in ordinary corporate workflows.

The larger lesson is that adoption is moving from novelty to accounting. AI can still be worthwhile even when its token bill looks large, particularly if it replaces expensive human effort. But companies now need to measure the value of each task, match it to the cheapest capable model and decide where automation genuinely pays. The next phase of enterprise AI will be shaped not only by what agents can do, but by whether businesses can afford to let them do it at scale.