The Shadow Token Problem in Enterprise AI

Uber’s experience with Claude Code serves as a cautionary tale for enterprises embracing generative AI. After just four months, the company reportedly burned through its entire annual budget — a clear case of what I call “shadow tokens”: AI consumption that occurs outside established governance frameworks.

This phenomenon is widespread. Microsoft is now reportedly curtailing internal licenses across engineering teams, and one in five organizations miss their AI spend forecasts by over 50%. By 2028, Gartner predicts AI coding costs will equal developer salaries — a significant shift from traditional software investments.

The core issue isn’t recklessness but rather inadequate oversight. When engineers have the final say on tool usage without clear return-on-investment expectations, consumption can quickly spiral out of control.

The New Dimension of AI Costs

Unlike SaaS purchases where costs are predictable upfront, AI expenses behave differently:

  • Consumption is based on behavior: Instead of fixed seats or contracts, you pay for each API call, prompt, or agent interaction
  • Costs scale exponentially: A tool that seems inexpensive in pilot can balloon with increased usage and complexity
  • Lack of visibility: Many organizations approved AI tools without establishing consumption guardrails

This creates a disconnect where teams adopt new technologies enthusiastically while finance struggles to model the true costs — particularly when decisions rest with engineers who may not be incentivized to optimize spending.

From Adoption Metrics to Financial Accountability

Uber’s experience highlights this tension. The company created leaderboards that rewarded token consumption, inadvertently encouraging “tokenmaxxing” behavior where quantity trumped quality. When teams are praised for using more AI rather than delivering better outcomes, shadow tokens become inevitable.

The solution? Shift the focus from adoption to yield — connecting AI usage directly to business results. Instead of rewarding engineers for consuming more tokens, recognize them for solving problems and creating value with those tools.