The bill moves with your volumes. Under some billing models, it also moves with how much work the software does for each task. All figures in this article are illustrative. Prices are before tax and in no particular currency.
Learn the three ways vendors bill usage
Most AI and automation features use one of three billing models. One product can use a different model for each feature.
- Token-based usage. A token is a small unit of text that an AI model reads or writes. A short word is often one token, and a long word can be several. The bill counts input tokens, which the model reads, and output tokens, which it writes. Some vendors convert tokens into credits before they bill them.
- Fixed credits per action. The vendor publishes a charge for each billable event, such as 8 credits per call summary. The charge stays the same however much model work sits behind it.
- Outcome-based charges. The vendor bills a completed result, such as a resolved conversation. Attempts that do not reach that result may not be billed.
Under token-based billing, input tokens include more than the customer's message. They also include your instructions to the model and any articles or records it looks up. Many model providers price output tokens higher than input tokens.
Define the billable event
Each billing model counts a specific event. Find the vendor's definition before you estimate anything.
- Message. One request or one reply.
- Conversation. A thread that can hold many messages.
- Resolution. A conversation that the vendor's rules mark as solved.
These counts differ. For example, 2,000 conversations in a month might hold 6,000 messages and reach 1,500 resolutions. Check what the vendor counts and when it makes the count, such as when a conversation closes.
Understand allowances, overage and limits
Most usage-based plans combine five parts. Read each one in your order form or price sheet.
- Minimum commitment. The amount you pay each period, whatever you use. It often buys the included allowance.
- Included allowance. The units your subscription includes for a period, usually a month or a year. Check whether unused units roll over or reset at the end of the period.
- Overage. Usage above the allowance. It is billed at a set rate per unit, often in arrears. Some vendors ask you to buy an extra pack of credits instead.
- Alert. A notice sent when usage reaches a set share of the allowance, such as 75%. Work carries on as normal.
- Hard limit. A cap that stops usage when it is reached. Any work that depends on the feature stops at the same time.
Limits, automatic upgrades, resets and handoffs vary by product. Some plans move you to a higher tier when you pass the allowance. Some features hand work to a person at the limit, and others stop. Check how each feature behaves before you rely on it.
A hard limit protects your budget, but it can pause a business process. If an AI agent stops at its limit, customer emails wait unanswered. Decide the fallback before you set the limit, such as sending emails to a staff inbox.
See what drives the bill
Usage grows for five main reasons. How much each one matters depends on the billing model.
- Volume. More billable events mean more charges under every model.
- Size of each action. Longer inputs and longer replies use more tokens. This changes the bill under token-based billing only.
- Steps per action. An agent may call a model several times to finish one task. Under token-based billing, each call adds usage. Under fixed-credit or outcome billing, extra model work does not necessarily change the bill.
- Retries and testing. Failed runs and test runs can count against the same allowance as live work. Check whether the vendor bills them.
- New features. Someone switches on a feature that uses units, and nobody updates the budget.
Calculate a month's usage for one AI agent
Example (illustrative): an AI agent answers support emails and averages 5 credits per email. The vendor meters the agent by tokens and converts them into credits. Longer emails therefore use more credits.
The agent shares a plan with two fixed-credit features: call summaries and contact enrichment. Together they use 4,800 credits in a normal month and 7,200 in a peak month.
The plan includes 20,000 credits a month, with no rollover. Overage costs $0.04 per credit. Alerts fire at 75% and 90% of the allowance, and a hard limit stops usage at 22,000 credits.
| Month | Emails | Credits per email | Agent credits | Other features | Credits needed | Overage billed | Overage cost |
|---|---|---|---|---|---|---|---|
| Normal | 2,000 | 5 | 10,000 | 4,800 | 14,800 | 0 | $0.00 |
| Peak | 3,000 | 5 | 15,000 | 7,200 | 22,200 | 2,000 | $80.00 |
| Longer threads | 2,000 | 6 | 12,000 | 4,800 | 16,800 | 0 | $0.00 |
In the normal month, the plan needs 14,800 credits. This is the expected operating usage, and it stays inside the allowance.
In the peak month, a seasonal rush lifts email volume by 50%. The features need 22,200 credits, but the hard limit stops them at 22,000. The plan bills 2,000 credits of overage, which costs $80.00. The last 200 credits of work go to the fallback.
In the longer-threads month, volume is flat, but customers send longer email threads. The average rises from 5 to 6 credits per email. That adds 2,000 credits, the same as 400 extra emails. The 75% alert fires at 15,000 credits.
This rise happens only under token-based billing. At a fixed 5 credits per email, the longer-threads month would cost the same as the normal month. Track volume and credits per action each month.
To build these figures for your own processes, see how to forecast AI and automation usage. To check them each month, see how to run a monthly usage review.
Record the pricing terms for each product
Copy this table into a spreadsheet and complete one for each product that charges by use. The examples match the worked example above.
| Term | What to record | Example (illustrative) |
|---|---|---|
| Billing model | Token-based, fixed credits per action or outcome-based | Token-based, converted into credits |
| Billable event | What the vendor counts, and when | Each email the agent processes |
| Units per action | Average units for one billable event | 5 credits per email |
| Minimum commitment | Amount payable whatever you use | $700 a month, which buys the allowance |
| Included allowance | Units included in each period | 20,000 credits a month |
| Allowance period | Monthly or annual, and the reset date | Monthly, resets on the 1st |
| Rollover | Whether unused units carry over | No rollover |
| Overage rate | Price per unit above the allowance | $0.04 per credit |
| Alerts | Thresholds and who receives them | 75% and 90%, to the platform admin and budget owner |
| Hard limit | Where usage stops, if anywhere | 22,000 credits a month across all features |
| Behaviour at the limit | Whether the feature stops, hands off or upgrades | Stops. Emails go to the staff support inbox. |
| Test usage | Whether test runs count against the allowance | Yes, they count |
| Approver | Named person who approves increases | Budget owner, such as the head of finance, by name |
Track usage as counts and totals only. Keep customer records, emails and AI prompts out of the usage report.
Avoid these common mistakes
- Reading only the allowance. The minimum commitment, overage rate, reset date and rollover rules decide what you pay in a busy month.
- Assuming every action costs the same. Longer inputs or extra steps can increase token-metered usage. Check whether the vendor charges fixed credits per action or bills a completed outcome.
- Forgetting the build phase. Building and testing an agent can use units before it goes live.