Uber burned through its entire 2026 AI coding budget in four months. Microsoft is canceling most of its direct Claude Code licenses six months after rolling them out. Meta took down its internal "Claudeonomics" leaderboard within 48 hours of it leaking to the press. The official story in each case is that AI adoption exceeded forecasts. The real story is that companies built incentive systems that reward token consumption, and got exactly what they asked for.

I have spoken with engineers at Uber, Meta, and Amazon over the past few weeks. They describe a culture of openly gaming internal AI metrics. The most striking version is two agents set up to talk to each other all day, producing nothing but token consumption visible on a manager's dashboard. Less elaborate variants are everywhere: running AI on tasks the engineer could do faster by hand, asking it questions about code that is already documented, prompting it to prototype features no one intends to ship. My estimate, based on those conversations, is that close to half of the token spend at these companies right now is waste.

I have started calling this tokentaxxing. It is the counterpart to tokenmaxxing, and it names what the gaming is actually doing: levying a private tax on the company. Every wasted token is a payment the business is making to its own broken incentive structure. Anthropic and OpenAI receive the money. The company receives a dashboard.

The CFO needs to own this

Token spend has been treated as an engineering problem. It is a finance problem. If a company spends $500,000 on tokens, there needs to be at least $500,000 in proven value on the other side. That is how every other line item in the budget already works. Cloud spend gets justified per service. Software licenses get reviewed against utilization. Contractor invoices require deliverables. Token spend is the only category where companies have allowed themselves to skip the justification step, on the theory that AI adoption is so strategic that any spend is good spend. The Uber and Microsoft bills show how that theory ends.

The finance team has to get creative here, because tokens do not come with traditional accounting hooks. Below are practical mechanisms that work. None of them require sophisticated tooling. They require the will to actually do the accounting.

Tie token spend to specific projects

Every project above some threshold (say $10,000 in expected agent costs) gets a token budget at kickoff. The budget is approved alongside the project plan, the same way a contractor budget would be. The project owner specifies, in writing, what outcome the spend is supposed to produce. Faster delivery by how many weeks. Fewer engineering hours by how many. A feature shipped that otherwise would not have. At the end of the project, finance reviews actual token spend against the claimed outcome.

The mechanism works because it forces the conversation to happen before the money is spent, when the outcome is still a prediction. Engineers who cannot articulate what their token budget will produce do not get the budget. Engineers who consistently deliver on their predictions get more.

Require a productivity baseline before approval

Before any team is given heavy agentic access, they should establish what their delivery looked like without it. Cycle time on a representative project. Defect rates. Engineering hours per shipped feature. The baseline becomes the comparison point. After three months of AI-assisted work, the same metrics get measured again. If cycle time dropped by 30% on a team spending $200,000 a month on tokens, and the loaded cost of that time savings exceeds $200,000, the spend is justified. If it didn't, the spend isn't.

This is straightforward measurement. The reason most companies do not do it is that the answer might be uncomfortable, not that it is hard.

Measure hours saved, then convert to dollars

The simplest test, and the one Indeed publicly uses, is hours saved per engineer per week. Indeed's reported figure is more than four hours. At a fully loaded cost of $250 per hour for a senior engineer, that is $1,000 per engineer per week, or roughly $50,000 per year. If a company is spending more than $50,000 per engineer per year on AI tools, the math is not working. If it is spending less and the engineers report real time savings, the math is working.

Time savings are measurable with simple surveys, time-tracking sampling, or workflow analytics. The number is imprecise, but imprecise honest measurement beats precise wrong measurement, which is what token leaderboards produce.

Charge teams for their own consumption

The most powerful mechanism is internal chargeback. Engineering teams get a token budget as part of their annual plan, the same way they get a headcount budget or a cloud budget. When the team spends tokens, the cost shows up on their P&L. Suddenly the team lead has a direct incentive to stop wasteful use, because it competes with other things the team wants to fund.

This is how cloud costs got under control at most large companies. For years, cloud spend was treated as a central infrastructure cost with no team-level accountability, and it ballooned. The fix was attribution. Once teams saw their own AWS bills, the bills came down. Token spend will follow the same pattern as soon as companies attribute it the same way.

Audit a sample of high-spend sessions

For any engineer in the top decile of monthly token consumption, finance should pull a random sample of their sessions and have a senior engineer review what was actually done. This does not need to be heavy or punitive. The point is that the audit exists. People behave differently when they know their work can be reviewed. The audit also produces a stream of qualitative information that no dashboard can: what tasks are producing real value, what tasks are theater, what patterns of use should be encouraged or discouraged.

A finance team that runs even quarterly audits on the top 10% of consumers will know more about the company's AI ROI than the dashboards will ever tell them.

The shift that has to happen

Kill the consumption leaderboards immediately. They were a marketing tactic dressed as a management one, and they produced the predictable result. Replace them with project budgets, baselines, time-savings measurement, internal chargeback, and audits. Make engineering accountable to finance for AI spend the same way they are accountable for every other major cost.

The technology is genuinely useful. The bills are real. Companies that learn to attribute the spend to outcomes will keep buying more of it because the math will keep working. Companies that do not will keep paying the tokentax until they cancel the tools entirely, like Microsoft just did, and lose the value along with the waste.

The CFO does not need to understand prompts. They need to insist that every dollar of token spend can point to a dollar of value. That is the job.