Quick answer
Analyse AI coding cost in three lines: fixed seats, variable usage such as credits and API tokens, and the cost of blocked time. Report each separately by team and project, and state whether figures are reported or estimated. Without that separation, a budget review cannot tell growth from a plan change.
The three lines
| Line | Source | Behaviour |
|---|---|---|
| Fixed seats | Contracts and invoices | Predictable |
| Variable usage | Credits, overage, API | Follows work |
| Blocked time | Incident and delay notes | Hidden, often largest |
Blocked time is the line nobody invoices. Count the hours lost when a team cannot work and the cost of the fallback, then compare it with the price of the next tier.
A defensible method
- Define the period and the population.
- Report seats and usage from billing.
- Attribute usage to projects using local session data.
- Divide by completed tasks, not prompts.
- State the estimation method and the known gaps.
Executives accept estimates that are labelled. They distrust precise numbers with no method, and they are right to.
What the analysis should decide
- Seat count and tier mix for the next period.
- Which tools to renew.
- Where to set spend caps.
- Whether to invest in cheaper models for routine work.
- Which projects need a dedicated budget.
UsageAtlas provides the local attribution layer for 14 providers without collecting prompts. See cost per developer and ROI.
Questions, answered
How do enterprises measure AI coding cost?
Separate fixed seats, variable usage, and blocked time; attribute usage to projects; and divide by completed tasks with a stated method.
What is the biggest hidden cost?
Blocked time and retries. Neither appears on an invoice, and both grow when plans are undersized or agent runs loop.
Can this be done without collecting prompts?
Yes. Tokens, model, project, and date are enough. Prompt content adds risk and no analytical value.