The clearest usage-based pricing examples are published price lists: Twilio charges per text message segment, AWS Lambda per request and per GB-second of compute, Snowflake per credit of warehouse time, Zapier per task, and Intercom per resolved support outcome. Each one charges for consumption rather than access. What separates them is what the unit actually counts: raw activity, reserved capacity, or a finished result.
Most lists of usage-based pricing examples sort companies by industry: cloud, APIs, AI models, automation. That tells you who uses the model. It does not tell you which unit to pick, and it misses the change that matters most for software companies in 2026. More and more usage is generated by AI agents, not people, and an agent does not consume a product the way a person does.
This post groups real examples by what the unit counts, then tests each group against one question: what happens to the bill when an agent, not a person, drives the usage? The short answer is that activity units stop tracking what the customer needed and start tracking how the agent was built. Several well-known vendors have already changed their units because of it.
Three kinds of usage unit
Every usage-based price list charges for one of three things.
- Activity. Something happened: a message was sent, a request was made, an action ran. The vendor counts events.
- Capacity. The customer buys an amount of resource or a budget in advance, then draws it down: credits, compute time, a pool of dollars.
- Result. The vendor charges only when a defined outcome occurs, such as a resolved support conversation.
Stripe's guide to usage-based pricing examples makes the point that the billable unit shapes the customer's experience as much as the price does. We agree, and we would add that the unit also decides who carries the risk when usage becomes unpredictable.
Activity units: price per event
These are the oldest and most familiar usage-based pricing examples. The unit is easy to count and easy to explain.
Twilio, per SMS segment. Twilio's US SMS pricing sets a base rate of $0.0083 "to send or receive standard outbound or inbound SMS segments in the US," per 160-character segment, with carrier fees added on top.
AWS Lambda, per request and per GB-second. AWS Lambda pricing charges $0.20 per one million requests plus $0.0000166667 per GB-second of compute in its worked examples, and the free tier includes one million requests and 400,000 GB-seconds per month. Two meters run at once: how often the function is called, and how much memory it holds for how long.
Zapier, per task. Zapier defines a task as "any successful action that runs in Zapier." Triggers never use tasks, and neither do filter or path steps. The interesting line is newer: Zapier MCP uses 2 tasks per successful tool call. When an AI agent calls Zapier as a tool, every step it takes is metered.
Salesforce Agentforce, per action. Agentforce pricing sells Flex Credits at $500 per 100,000 credits, and "Agentforce actions are 20 Flex Credits." That works out to ten cents per action the agent takes, whatever the action achieved.
Activity units work well when a person decides how many events happen. A marketer chooses how many texts to send. A developer chooses how often a function runs. The bill follows choices the customer can see.
Capacity units: prepay and draw down
Capacity units put a budget around activity. The customer still pays for consumption, but buys it in advance or within an allowance.
Snowflake, credits by the second. Snowflake's pricing guide bills warehouse credits "by the second with a one-minute minimum," and an X-Small warehouse uses 1 credit per hour. Customers can also pre-purchase capacity: a fixed dollar commitment drawn down month by month.
GitHub Copilot, premium requests. GitHub's Copilot billing docs give Copilot Pro 300 premium requests a month and Pro+ 1,500, with additional requests at $0.04 each. Code review is priced separately: from June 1, 2026, each pull request review consumes 13 premium requests.
Cursor, a monthly dollar pool. In June 2025 Cursor moved its Pro plan from a count of requests to what its pricing update calls "a usage credit pool for the month," worth $20 of frontier model usage at API rates, with a spend limit for anything beyond it.
Capacity units are the usual answer to the main complaint about pure usage pricing, the surprise bill. They keep the meter but put a ceiling on it.
Result units: price per outcome
The newest group charges only when something defined as done happens.
Intercom Fin, per outcome. Intercom prices Fin at $0.99 per outcome. An outcome counts when a customer confirms the issue is resolved, when they make no further request for help after Fin replies, or when Fin completes a workflow, and each conversation is charged once.
Salesforce also offers a per-conversation price for Agentforce, $2 per conversation. That sits between the groups: a conversation is a container for activity, not a guarantee of a result.
Result units move the variance to the vendor. If the agent takes ten steps instead of two, the customer pays the same. The price is only as good as the definition of done, and the vendor writes that definition.
What changes when an agent drives the usage
Here is the part the usual lists leave out. Activity units were designed around people, and people generate roughly predictable amounts of activity. Agents do not.
Anthropic's engineering team measured this on its own research system: "agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats." See How we built our multi-agent research system. The same request from a customer can cost very different amounts depending on how the agent behind it is built.
Three of the examples above show vendors reacting to that.
- Cursor dropped the request as its unit. Its update explains that newer models can use many more tokens on longer tasks, so the hardest requests cost far more than simple ones and a request count no longer reflected real cost. It switched to a dollar pool at API rates. The rollout was rough enough that Cursor apologized and offered refunds for unexpected charges between June 16 and July 4.
- GitHub kept a person-shaped unit on purpose. In agent mode, Copilot charges one premium request per prompt you send, and its cloud agent charges one per session plus one per steering comment. Tool calls the agent makes on its own do not count. GitHub prices what the person asked for and absorbs the variance in what the agent does to deliver it.
- Zapier meters the agent's steps. Zapier MCP charges per tool call, so a customer's agent that takes twelve steps to finish a job pays for twelve calls. That is honest when the customer built the agent and chose those steps.
The pattern: when the customer controls the agent, an activity unit is fair, because the customer controls how many steps it takes. When the vendor's own AI does the work, an activity unit bills the customer for the vendor's design choices. We made the broader version of that argument in usage-based vs outcome pricing: who controls the meter?. This post is about what it means for the unit itself.
A test for any usage unit
Before you copy a unit from a price list, put it through four questions.
- Who generates it, a person or an agent? If an agent generates most of it, expect the count to depend on how the agent works, not on what the customer needed.
- Can the buyer predict it before the month starts? Twilio texts are predictable for a known campaign. Agent tool calls usually are not.
- Does it grow with what the customer got, or with how the work was done? A resolved conversation grows with value. A token count grows with effort.
- Is there a ceiling? Capacity units and allowances turn an unpredictable meter into a budget. If you keep an activity unit, add one.
| Example | Unit | Counts | When an agent drives usage |
|---|---|---|---|
| Twilio | SMS segment | Activity | Fine if the customer decides what gets sent |
| AWS Lambda | Request, GB-second | Activity | Fine for the customer's own code and agents |
| Zapier MCP | Tool call | Activity | Bills every step the agent takes |
| Agentforce | Action | Activity | Bills every step the agent takes |
| Snowflake | Prepaid credits | Capacity | Variance capped by the commitment |
| Cursor | Dollar pool | Capacity | Variance capped by the pool and spend limit |
| GitHub Copilot | Premium request per prompt | Capacity, person-shaped | Vendor absorbs the agent's variance |
| Intercom Fin | Outcome | Result | Vendor absorbs the variance, defines done |
The unit for work your AI does for customers
If you run a software product, you will likely need two units at once. Usage on your API or MCP server, where customers and their own agents call you, suits activity or capacity pricing, because the customer decides how much happens.
Work your own AI does for each customer is different. Much of it is not a single event at all. It is an ongoing job, like getting every new hire through onboarding or keeping every location compliant, and it holds a state that needs attention whenever it slips. For that, neither a per-step price nor a per-event outcome fits well. A subscription for the job does: the customer pays a monthly price for the job kept done, and the vendor carries the cost of the steps.
An illustrative example: an LMS like Lessonfield could sell its customer Pellmark Training "get every new hire through onboarding" for a monthly price. Whether that month takes forty reminders or four is Lessonfield's cost to manage, not a line on Pellmark's bill. We covered why per-unit job pricing keeps subscription predictability in seat-based pricing after AI agents.
Where Aimdoc fits
Aimdoc connects to your product and your systems and delivers the work your customers need, so you can sell that work as a service. What you charge your customers for a job is your price, billed through your own Stripe account, usually as a subscription for the job. See selling services.
What you pay Aimdoc is a capacity unit: prepaid credits for the AI's work, with spend caps and per-job budgets you set. You decide which jobs run, so you control the meter. Aimdoc never charges per outcome or per resolution.
FAQ
What are examples of usage-based pricing?
Twilio charges per SMS segment, AWS Lambda per request and per GB-second of compute, Snowflake per credit of warehouse time, Zapier per successful task, GitHub Copilot per premium request beyond a monthly allowance, and Intercom per resolved outcome.
What units do usage-based pricing models use?
Three kinds: activity units that count events (messages, requests, actions), capacity units that are bought in advance and drawn down (credits, dollar pools), and result units that charge only when a defined outcome happens.
Does usage-based pricing work for AI agents?
It works when the customer controls the agent, because the customer controls how many steps it takes. When the vendor's own AI does the work, activity units bill the customer for the agent's design, so capacity limits, outcome prices, or a subscription for the job fit better.
What is the difference between usage-based and credit-based pricing?
Credit-based pricing is a form of usage-based pricing. The customer buys credits in advance and spends them as they use the product, which caps the bill. Snowflake and Salesforce Agentforce both price this way.
Your customers want the work done, not a meter to watch. Aimdoc delivers that work through your product and lets you sell it as a service at a price you set. Start free, or book a demo.