Use this on-page worksheet to document a monthly AI-agent budget. It is a copyable planning table, not a downloadable Excel file. Put the provider’s current quote or pricing page and the date you checked it beside each rate; the sample below uses fictional values only.
Line items
Separate usage that changes with demand from recurring fixed charges and one-time implementation. Keep labor visible as hours × an organization-specific loaded rate. Capture unit, volume, rate, currency, source, rate date, owner, and whether a line is measured, quoted, estimated, or allocated. Use the chosen provider’s official pricing page—such as OpenAI API pricing where applicable—and verify the model, service mode, region, tool charges, and contract that actually apply. A rate page is evidence for a rate, not a complete workload forecast.
Template
Copy the table into your own spreadsheet or budgeting system. Add rows when a workflow has another billable unit; keep implementation separate from the recurring monthly total.
| category | unit and volume | rate / formula | monthly estimate | source, date, owner, status |
|---|---|---|---|---|
| Model input tokens | tasks × steps × input tokens | input tokens ÷ 1,000,000 × rate per million | enter amount | official rate or invoice; checked date; owner; measured/quoted/estimated |
| Model output tokens | tasks × steps × output tokens | output tokens ÷ 1,000,000 × rate per million | enter amount | same evidence fields |
| Cache / batch / other model modes | eligible units only | provider-specific billable unit × entered rate | enter amount | verify eligibility and terms; do not assume a discount |
| Tools and orchestration | calls, runtime, storage, or fixed plan | measured quantity × applicable rate | enter amount | service pricing or invoice; avoid double counting |
| Infrastructure and observability | monthly fixed or usage quantity | documented charge or allocation basis | enter amount | invoice / allocation owner |
| People, evaluation, and review | hours and review volume | hours × loaded rate; label internal allocation | enter amount | team estimate / owner |
| One-time implementation | discovery, integration, release | hours × rate + known external charges | keep outside monthly run total | estimate owner and date |
For each period, calculate recurring estimate = variable usage + recurring fixed charges + recurring people/operations. Show any one-time implementation separately. Keep direct provider charges distinct from allocated internal cost.
Example
Hypothetical monthly scenario: 1,000 tasks, 2 model steps per task, 1,000 input tokens and 500 output tokens per step. Entered illustrative rates are $2 per million input tokens and $8 per million output tokens; a tool call costs an assumed $0.001 per step. The arithmetic is: input = 2,000,000 ÷ 1,000,000 × $2 = $4; output = 1,000,000 ÷ 1,000,000 × $8 = $8; tools = 2,000 × $0.001 = $2. Add an assumed $25 fixed monthly allocation and $40 of review/support effort, for a hypothetical monthly total of $79. All rates and fixed amounts here are fabricated user-entered examples, not current provider prices. One-time setup is excluded.
Replace each assumption with an invoice, measured usage, current official price, or clearly labeled estimate. Add retries, longer-context cases, cache treatment, or additional tool products only when the source data and applicable billing terms support them. The AI agent cost calculator can scenario-test task, step, token, tool, fixed, and human-cost entries; the LLM API cost calculator and token cost calculator help isolate token assumptions. The local tools calculate entered scenarios and export result CSV; they do not retrieve pricing, create a workbook, or approve spending.
Review cadence
Choose a review interval that matches workload volatility and your billing period. Compare forecast and actual for the same scope and dates. Record variance in tasks, steps, token mix, rates, tools, and review effort; assign an owner and action. Recheck after a model, prompt/context, tool, contract, or workflow change. Keep the prior approved base case so revisions remain explainable. The FinOps Foundation’s Planning & Estimating and Forecasting guidance treats estimates as inputs to forecasts that should be revisited as usage and plans change; use your organization’s own approval process for budget decisions.