Token costs you can actually explain to your CFO

The most common question from finance is not “how much did AI cost last month” — the invoice answers that. It is “what did we get for it”. Without task-level attribution, nobody can answer.

Attribute spend to outcomes, not API keys

Agentis records token usage per run, rolls it up per task, and aggregates by team, project, model and label. Cache reads and writes are tracked separately, because a healthy cache ratio is the single biggest lever on cost for long-context coding work.

• Per-task cost makes “was this worth automating” a data question.
• Per-model splits reveal when a cheaper model would do the job.
• Cache hit ratios expose context-engineering problems early.
• Per-team budgets turn AI spend into a normal operating line.

Once costs are attributed to tasks, an interesting thing happens: teams stop arguing about whether agents are “too expensive” and start comparing the cost of a run against the cost of the engineer-hours it replaced. That is a conversation finance is happy to have.