For CFOs and CTOs: How Tokens Quietly Eat Your Company's Budget

The AI bill does not arrive as one shock. It seeps in. Unlike software licences bought once per seat, language model usage is billed by consumption: every word that goes into the model and every word that comes out carries a price. The result is a line item that did not exist a year ago and is now among the fastest growing in many technology budgets, without ever passing through the usual procurement cycle.

Why spend grows faster than usage

The paradox is that headcount can stay flat while the invoice doubles. Four mechanisms explain it:

• Output costs more than input. Most providers price the tokens a model produces at several times the tokens it consumes. A feature that writes long text, an extended summary or a drafted report, costs far more than one that answers in a line, even at identical call volume.
• Conversations resend themselves. The model remembers nothing between requests, so each new message carries the entire prior exchange with it. Ten messages are not ten times one message. They are more, because every step carries everything before it.
• Grounding in documents multiplies the input. When you feed internal documents to a model so it can answer from them, those documents count as input on every call, not once.
• Agents turn one request into ten. Agentic features plan, act and review, which means several model calls behind a single button press. Failed attempts and retries are billed too.

Where the cost usually hides

The management problem is not the size of the number but its absence from the reports. The invoice arrives from the provider as a single block, while the decisions that produced it are spread across teams and products. Without tying each call to the feature that triggered it, neither the CFO nor the CTO can answer a simple question: which feature costs more than it returns?

Five practical steps

• Attribute cost to the feature, not to the invoice. Log the feature name, the model and the token counts with every call. That one line in the log is the difference between managing spend and watching it.
• Measure cost per successful outcome. A call repeated because its output was rejected cost twice and delivered once. Cost per accepted output is a truer metric than cost per call.
• Tier your models. Not every step needs the strongest model. Classification, extraction and routing are handled well enough by smaller, cheaper ones, keeping the expensive model for the step that earns it.
• Use caching and batch processing. Most platforms discount repeated context (prompt caching) and non urgent work (batch). Anything no user is waiting on in real time is a candidate.
• Alert on rate of change, not totals. The useful alert is not "we are over budget" but "this feature's spend doubled in a week". The first arrives late. The second arrives in time.

Model spend is not a fixed line approved once a year. It is an operating variable that moves with every feature release. The companies that treat it that way, as a unit cost to be measured and improved, are the ones that can scale their use of AI without being surprised by the bill.