Tokenmaxxer

A Proposal for the Corporate AI Economy

Uber maxed out its full-year AI budget in four months. Somebody had to approve the yearly budget for tokens. But when a year’s credit vanishes in Q1, who is to blame for that?

This is the world corporations built when they decided that firing people and replacing them with AI would save money. An Axios report confirmed that companies were spending more on AI compute than on the employees the AI was meant to displace. The layoff had become more expensive than the salary.

Sniper got sniped

Historically, the natural corporate response to a spending problem is not to spend less; usually, we make committees for oversight.

Consider the following scenario:

Every company, beyond a certain headcount, would designate a team whose sole function is to decide which tasks are worth automating with AI. Before an employee feeds a prompt into an AI, they should be able to justify the use of AI for that task.

Only three out of ten companies are using AI in proportion to what it costs them. The other seven are wasting compute on tasks that are not worth it, run by employees who have been told, in no uncertain terms, that their AI usage is being monitored.

Andrew Bosworth, CTO of Meta, put it plainly: “Nobody should be using AI tools just for the sake of using them.” I find it to be a reasonable position. Might have been more useful before Meta built a “Claudeonomics” dashboard ranking all 85,000 of its employees by token consumption, burning through 60 trillion tokens in a single month.

The Auditor’s task classification system would work as follows. Tasks are divided into three tiers.

Tier One: AI-eligible. Complex synthesis, code generation, and large-scale analysis. These are the tasks AI was built for. Approved. Proceed. The Auditor also specifies which model. A small, cheap one for the calendar invite. A large, expensive one only when the task has earned it.

Tier Two: Human-eligible. Replying to an email that requires a human voice. Deciding whether the company picnic should have a theme. Reading the room. These are tasks where, despite everything, a person remains the better instrument.

Tier Three: The grey zone. Writing a meeting summary for a meeting that should not have happened. Generating five variations of a sentence that was fine as written. Asking an AI to explain a document that, had anyone read it, would have taken four minutes. These require a judgment call. The Auditor team earns their salary here.

This is human behavior, which should surprise no one. Give people a metric, and they will optimize for the metric.

Jensen Huang declared that Nvidia engineers should consume AI tokens worth at least half their annual salary each year to be fully productive. JPMorgan now has employees spending more on tokens than their own salaries.

The deeper issue is that there is a lack of a serious tokenomics governance framework for deciding what AI is actually for inside a company.

Gartner projects that even as individual token costs fall by 90% by 2030, enterprise AI bills will keep rising, because agentic models require far more tokens per task, and increased consumption outpaces falling unit costs.

Cheaper per token does not mean cheaper in total. The bill gets larger as the machine gets faster.

However, bureaucracy applied to computing is still bureaucracy, and the overhead of governance typically costs more than the waste it prevents.

These are management problems, and management problems, historically, resort to a committee as a solution.