AI Budget Planner
Plan, estimate and monitor token spend for AI agents and workflows. Describe the work in plain terms, see the monthly bill, the 12-month trend and the levers that bring it down. Every price and assumption is in plain view.
Step 1 · Workloads
A workload is one repeatable job, such as triaging invoices or summarizing tickets. We filled in two examples. Change any number and the results update as you type.
Words to tokens
Rule of thumb: about 750 words is about 1,000 tokens. One page is about 500 to 700 tokens. Use “Fill from pages” on any workload to skip the math.
What this means
12-month projection
Savings levers
Monthly savings on today's plan. Each one has a trade-off, so read the sentence before you pull it.
Same workload, other models
Same volume, tokens, caching and batching. Only the model changes. Price is one input: test quality on your own cases before you switch.
AI costs are unpredictable.
They are metered and forecastable. Once you know volume, tokens per call and the model, the bill is arithmetic. What surprises people is unmetered growth, not the price.
Step 2 · Monitor
Enter what you were billed each month, or paste an export from your billing console. You will see the variance, the burn to date and when you would cross the budget if the trend holds.
One line per month, as YYYY-MM,amount. Headers, dollar signs and thousands separators are fine. Parsed in your browser; nothing is uploaded anywhere.
Guardrail checklist
Spend is not the only thing to monitor. Keep people in the loop for decisions that carry financial, legal or customer risk, and cap automation around 80% of a task. A cheap agent that approves the wrong payment is the most expensive line in the plan.
Prices used
This is an estimate from public list prices. It does not include enterprise agreements or committed-use discounts, taxes, cache-write and storage fees, or long-prompt surcharges. Batch pricing is 50% off input and output at all three providers. Check your own contract and billing console before you commit a budget.
Field guide
What finance and operations leaders need to know before the first invoice arrives.
A chat answer is one model call. An agent works in a loop: read the task, plan, call a tool, read the result, decide again. Each step resends the instructions and everything gathered so far, so input grows with every call. Ten steps on a long document can cost more than a hundred simple summaries.
Rule of thumbBudget agents at 4 to 8 calls per task until you have measured the real number.
Model choice is the biggest: the same work can differ by a factor of ten or more between tiers. Caching cuts the cost of instructions and reference material you resend. Batching halves the price of anything that can wait. Use model routing to send easy steps to a small model.
Rule of thumbStart on a mid-tier model, cache the fixed prompt, batch what can wait, then measure.
AI is billed like electricity: by use, every month. Forecast from volume and tokens, set alerts before the meter runs, and compare plan to actual on a fixed day each month. When usage grows because the agent is working, raise the budget on purpose rather than by surprise.
Rule of thumbForecast, monitor, review monthly. One owner, one meeting, one page.
We build custom assessments, skills and internal apps as part of our engagements. Tell us what you need.