BYD TechThe short answer
Price an AI product by a unit your customer already counts and your costs move with. Seats are easy to buy but fit badly when the software does the work; usage follows your costs but is hard for buyers to budget; outcomes match value but only work when both sides count the result the same way. Most AI companies now combine a base subscription with usage on top.
Key points
- Know what one finished task costs you before you choose a pricing model.
- Seats are the easiest to sell and the weakest fit when an agent does the work.
- Usage pricing follows your costs, but buyers need a way to cap it.
- Charge per outcome only when you and the customer can count the outcome the same way.
- A base subscription plus usage is the most common choice among B2B software companies in 2026.
Why AI products are harder to price
Most software costs about the same to run whether a customer uses it a little or a lot, so a price per user works. An AI feature is different: every request calls a model, and you are billed for it.
One click can mean several model calls, a search, a few tool calls and a retry. A heavy user can cost you many times more than a light one on the same plan. If the price does not move with that cost, your best customers can become your least profitable ones.
The other change is what the customer is buying. When an agent answers support tickets or reads documents, it does work a person used to do. Charging per seat then works against you: the more work the agent does, the fewer seats the customer needs.
The five ways to price it
| Model | The customer is charged for | Works when | Breaks when |
|---|---|---|---|
| Per seat | Each person with a login | People use the product as a tool and usage per person is similar | An agent replaces work people did, so value grows while seats shrink |
| Flat subscription with limits | A plan with a usage allowance | Usage is predictable and the allowance covers most customers | A few heavy users consume the margin of everyone else |
| Usage | Each request, document, minute or task | The unit is easy to understand and your cost follows it | Buyers cannot predict the bill and slow down to save money |
| Credits | A prepaid balance spent across different actions | The product has several AI features with different costs | Customers cannot tell what a credit buys, so it feels like hidden pricing |
| Per outcome | A finished result, such as a resolved ticket | The result is clear, easy to count and agreed in advance | Nobody can agree whether the result happened |
Most products end up with a mix: a base subscription for access and support, plus usage, credits or outcomes on top. That is usually called hybrid pricing.
What AI companies are doing now
- Hybrid pricing is the most common model. In Kyle Poyar’s 2026 survey of more than 230 B2B software and AI companies, 37% used it, up from 25% a year earlier.
- Credits are spreading. In the same survey, 29% of companies already used AI credits and another 33% planned to introduce them.
- Intercom prices its Fin support agent per outcome. A resolution counts when no further help is requested after Fin’s last answer, and a conversation simply passed to a person without an outcome is not charged.
- Salesforce offers its Agentforce agents three ways: credits spent per action, a charge per conversation, or a license per user. One vendor selling the same product by three different units shows how unsettled this still is.
How to choose
- Measure what one finished task costs you: every model call, tool call, retry and any human review. Use real usage, not a demo. How to reduce LLM costs covers where that cost comes from.
- Find the unit your customer already counts. Support teams count tickets, finance teams count documents, sales teams count leads. A price in their unit is easier to approve than a price in yours.
- Decide who carries the risk of a heavy month. A cap, a usage allowance with a clear overage, or prepaid credits each put that risk in a different place. Say which one in plain words on the pricing page.
- If you charge per outcome, write down what counts as one with your first customers before launch, and show them a record of every outcome you count.
- Start with one simple model, and measure usage per customer from the first release even if you do not charge by it yet. That data is what lets you change the price later without guessing.
Common mistakes
- Unlimited AI use on a flat plan. It feels generous until one customer runs an automated job against it.
- A price that only makes sense with your current model costs. Model prices change, and so does the model you use.
- Outcome definitions the customer cannot check. If they cannot see why something counted, every bill becomes a dispute.
- Hiding usage. Customers who can see their own usage, with alerts before a limit, complain less and budget more.
- Copying a large company’s pricing page. Their units suit their customers and their scale, not necessarily yours.
What the product needs to support the price
A pricing model is also a set of features. Whichever one you choose, the product usually needs to record usage per customer, show it to them, enforce limits, send alerts before a limit is reached, and keep a record of every billed outcome or credit that the customer can check.
These are far easier to build into the first version than to add once customers are on a plan. If you are planning an AI feature or agent, include them in the scope from the start; AI agent development describes how we approach the rest of the build.
Sources
Checked when this guide was written.
