AI Costs Are Variable, Not Fixed—and Most Companies Aren't Tracking Them

The subscription model doesn't apply to AI

For decades, enterprise software followed a predictable pattern: buy a seat, pay a fixed monthly fee, use it as much as you want. That mental model is leading companies into a costly trap with generative AI.

Unlike traditional SaaS, generative AI charges by the token—roughly three-quarters of a word. Every output the system generates costs money, whether you keep it, edit it, or discard it entirely. What appears to be a subscription is actually a metered utility bill that scales with use.

The financial consequences are already visible. According to reporting first published by Inc., Axios documented one unnamed company that accumulated $500 million in charges from Anthropic's Claude in a single month because no usage limits were in place. Uber exhausted its entire 2026 budget for AI coding tools in just four months.

Why it matters

These aren't implementation failures at poorly managed companies—they're what happens when executives apply fixed-cost thinking to variable-cost technology. AI spend belongs on the CFO's radar as an operating expense, not buried in IT's software budget. Without active management, AI costs can spiral faster than any traditional software contract.

The waste rate metric companies ignore

The most critical number most organizations aren't tracking is their AI waste rate: the percentage of generated output they pay for but never use. Companies are billed for every token the model produces, regardless of whether the result makes it into production, gets heavily edited, or goes straight to the trash.

This creates a fundamentally different cost structure than seat-based software. A Slack license costs the same whether an employee sends one message or a thousand. An AI tool's cost varies with every query, every regeneration, every discarded draft.

Three operational controls

Managing AI as an operating cost requires three distinct practices:

Track usage at the source. Instrument systems to measure token consumption by team, use case, and individual workflow. Without granular visibility, you're managing in the dark.

Right-size access. Not every employee needs access to the most capable (and expensive) models for every task. Match model capability to actual requirements.

Implement governance. Set usage limits, approval thresholds, and review cycles. Treat AI spend like cloud infrastructure costs, not like Microsoft Office licenses.

The shift from subscription to metered pricing changes the fundamental question executives should ask. It's not "which AI tool should we buy?" but rather "how much of what we're generating are we actually keeping?"

These details were first reported by Pam Didner, a B2B marketing consultant and AI strategist, writing in Inc.