A forecast reduces both risks. For one team, it can take a few days. You need volume reports from your current systems, a short trial and a spreadsheet. All figures in this article are illustrative. Prices are before tax and in no particular currency.
List each process that will use units
Start with the work the software will do. For each process, write down four things:
- What it does. For example, an AI agent drafts and sends replies to support emails.
- Who owns it. Use a role, such as "Support manager", so the list survives staff changes.
- What starts it. For example, a new email arrives in the support inbox.
- How it is billed. Token-based usage, fixed credits per action or a charge per outcome. How usage-based pricing works explains each model.
The example uses three processes. An AI agent answers support emails, metered by tokens and converted into credits. A second feature summarises recorded sales calls at a fixed 8 credits each. A workflow enriches each new contact with company details at a fixed 1 credit each.
Count the billable events
Volume is the number of billable events each process will create. Your current systems already hold most of this.
- Run a report of monthly totals for the last 12 months. Count emails received, calls held or contacts created.
- Find the average month and the busiest month. Note why the busiest month was busy, such as a seasonal rush or an event.
- Ask each process owner about planned changes for the next year. New markets, products, campaigns and hires all change volume.
Count only the events the vendor bills. Not every email, contact or call necessarily incurs a charge. For example, spam filtered out before the agent runs may not count. A contact that already holds company details may be skipped. Ask the vendor which events are eligible, and adjust your counts.
Use reports of counts only. Keep customer records in the systems that hold them.
Measure the units for one action
For token-metered features, input length, output length and model calls affect usage. For fixed-credit features, use the vendor's published charge per billable event. Test the token-metered features before you rely on a vendor's average.
- Ask for a trial or test environment with the same settings you will use in production.
- Start with a varied sample of short, long and difficult cases. Use sample or anonymised content where you can.
- Expand the sample until the estimate reflects normal and difficult cases. For a first check, 20 to 50 examples may be enough.
- Record the average units per action and the highest you saw.
- Ask whether trial usage is measured the same way as live usage, and whether retries and failed runs are billed.
In the example, the support agent averaged 5 credits per email in testing. The highest single result was 9 credits.
Build the forecast for each kind of month
Forecast four kinds of month. A normal month uses your average volume. A peak month uses your busiest volume. A growth month shows your expected volume in 12 months. A growth peak month applies that growth to the peak.
| Process | Credits each | Normal volume | Peak volume | Growth volume | Growth peak volume |
|---|---|---|---|---|---|
| Support emails answered | 5 | 2,000 | 3,000 | 2,500 | 3,750 |
| Sales calls summarised | 8 | 400 | 500 | 500 | 500 |
| Contacts enriched | 1 | 1,600 | 3,200 | 2,000 | 4,000 |
Multiply each volume by its credits per action. The subtotal is the expected operating usage. Then add a buffer for testing, build work and changes you did not expect. Include retries and failed runs only if the vendor bills them. The total is the buffered budget.
This example uses a 10% buffer. Choose your own from the spread in your test results and your volume history.
| Credits | Normal month | Peak month | Growth month | Growth peak month |
|---|---|---|---|---|
| Support emails | 10,000 | 15,000 | 12,500 | 18,750 |
| Call summaries | 3,200 | 4,000 | 4,000 | 4,000 |
| Contact enrichment | 1,600 | 3,200 | 2,000 | 4,000 |
| Expected operating usage | 14,800 | 22,200 | 18,500 | 26,750 |
| Buffer (10%) | 1,480 | 2,220 | 1,850 | 2,675 |
| Buffered budget | 16,280 | 24,420 | 20,350 | 29,425 |
The peak month reflects a seasonal rush in support and a large event list in the same month. The growth month assumes 25% more emails and contacts. It also assumes two new sales reps lift calls from 400 to 500. In the growth peak month, calls stay at 500, because they depend on the number of reps.
Compare the forecast with the allowance
Suppose the vendor offers a monthly plan: 20,000 credits for $700 a month. Unused credits reset each month, and overage costs $0.04 per credit. Alerts fire at 75% and 90% of the allowance. A hard limit stops usage at 22,000 credits.
| Month | Buffered budget | Allowance | Difference | Overage cost |
|---|---|---|---|---|
| Normal | 16,280 | 20,000 | 3,720 under | $0.00 |
| Peak | 24,420 | 20,000 | 4,420 over | $176.80 |
| Growth | 20,350 | 20,000 | 350 over | $14.00 |
| Growth peak | 29,425 | 20,000 | 9,425 over | $377.00 |
The overage costs assume the whole buffer is used, so they show a cautious case. The peak and growth peak budgets also pass the 22,000-credit hard limit. Their overage costs assume someone raises the limit for those months.
Decide what happens if usage reaches the hard limit. In the example, the support agent would stop and new emails would wait. Write down the fallback, such as a staff inbox, and who can approve more units.
Now suppose the vendor also offers an annual pool of 240,000 credits for $9,000 a year, with the same overage rate. A base year with 10 normal months and 2 peak months uses 211,640 credits. The pool covers both peaks, with 28,360 credits left over.
If every month reached the growth budget, a year would use 244,200 credits. That is 4,200 more than the pool. A growth year with 10 growth months and 2 growth peak months uses 262,350 credits.
Add the plan prices before you compare options.
| Option | Price a year | Base year cost | Growth year cost |
|---|---|---|---|
| Monthly plan, 20,000 credits | $8,400.00 | $8,753.60 | $9,294.00 |
| Annual pool, 240,000 credits | $9,000.00 | $9,000.00 | $9,894.00 |
In this example, the monthly plan costs less in both years, even though the annual pool covers the base-year peaks. Other prices or overage rates can reverse the result.
Copy the forecast sheet
Use one row per process in a spreadsheet. The examples show the support agent.
| Field | What to enter | Example (illustrative) |
|---|---|---|
| Process | What the automation does | AI agent answers support emails |
| Owner | Role responsible for the process | Support manager |
| Billing model | Token-based, fixed credits per action or outcome-based | Token-based, converted into credits |
| Billable event | What the vendor counts | Each email the agent processes |
| Units per action | Average from your test, or the published charge | 5 credits |
| Highest in test | Largest single result | 9 credits |
| Normal monthly volume | Average billable events over the last 12 months | 2,000 emails |
| Volume source | The report you used | Support inbox monthly totals |
| Peak volume | Busiest month and the reason | 3,000 emails, seasonal rush |
| Growth assumption | Change expected over 12 months | 25% more emails |
| Expected operating usage | Normal volume times units per action | 10,000 credits |
| Buffer | Share added for testing and unplanned change | 10% in this example |
| Limits and fallback | Alerts, hard limit and what happens at it | Alerts at 75% and 90%, hard limit 22,000 credits, staff inbox at the limit |
| Approver | Named person who approves increases | Budget owner, such as the head of finance, by name |
Agree the forecast before you sign
- Ask the vendor about reset dates, rollover, test usage and how limits behave.
- Ask whether the units charged for each action can change during the contract term.
- Name the person in your business who approves increases, and get their sign-off on the forecast.