·5 min read
AI Agent Pricing: Per Seat, Usage, Credits or Outcome?
How to price an AI product when every call costs you: four models compared, the margin math that decides between them, and what outcome pricing requires.
- AI pricing
- Pricing models
- SaaS
There are four ways to price an AI agent: per seat, per usage, per credit, or per outcome. The right one is decided by two numbers, not by fashion: your variable cost per successful task, and how reliably you can attribute a result to your agent. If cost per task is meaningful and results are countable, outcome pricing is available to you. If not, credits or usage protect your margin, and per-seat quietly destroys it.
This is the one pricing question where copying the market leader is genuinely dangerous, because their cost structure is not yours.
Why AI breaks the SaaS default
Classic SaaS has near-zero marginal cost, which is why per-seat pricing worked for two decades: one more user cost you almost nothing. An AI agent inverts that. Every request consumes inference, and a heavy user can consume a multiple of a light one on the same plan.
The consequence is blunt: under per-seat pricing, your best customers are your least profitable ones, and you cannot see it in aggregate revenue until the cloud bill arrives.
The four models
| Model | Customer pays for | Protects your margin | Main risk |
|---|---|---|---|
| Per seat | each user | No | Heavy users are subsidised by light ones |
| Usage | consumption (tokens, calls, documents) | Yes | Unpredictable bills, budget objections |
| Credits | a prepaid allowance spent per action | Yes | Opaque if the conversion is unclear |
| Outcome | a successful result | Only if the success rate is known | You carry the failure cost |
Per seat
Still defensible when the agent is a bounded assistant inside a larger product and usage per user is stable. It stops working the moment usage varies by an order of magnitude across your base.
Usage
The purest alignment between what you spend and what you charge. The cost is predictability on both sides: your customer cannot budget, and your revenue moves with their activity. If you go here, give buyers a spend cap or an alert, because the single most common complaint about usage pricing is the surprise invoice.
Credits
The current default for AI features, because credits absorb the fact that different actions cost you different amounts while presenting one number to the customer. Two rules decide whether they are accepted: the customer must know before acting what an action costs, and the credit-to-cost ratio has to hold for your heaviest decile of users, not the average one.
Outcome
The most aligned and the most demanding. The customer pays when the agent delivers. The public reference point is Intercom's Fin, which charges $0.99 per resolution and defines a resolution precisely: the customer confirms the answer helped, or leaves without asking again. Intercom also publishes a platform variant with a base fee plus per-outcome charges, and distinguishes several outcome types at different prices.
That precision is the actual lesson. Outcome pricing is not "charge for results", it is "define, in writing, the single event that triggers a charge, in a way a customer will not dispute on an invoice". If you cannot write that sentence, you are not ready for it.
The margin math that decides
Before choosing, compute two numbers honestly:
- Cost per successful task. Include retries, failed attempts and the calls you make before a result is produced. If your agent succeeds on roughly half of attempts, your cost per success is double your cost per attempt.
- Task volume for the heaviest decile of customers. Not the average. The average never breaks your margin; the top ten percent does.
Multiply and compare against the plan price. If the heaviest decile consumes a serious share of what they pay, per-seat is off the table regardless of how simple it feels. Note also that success rates in production are lower than in demos: Intercom's own resolution rates in the field are reported in the 42 to 50 percent range, which is the number you should model your invoice against, not a benchmark from a launch post.
The free break-even calculator does the floor part of this arithmetic in two minutes.
A practical decision path
- Can you name the single event that counts as success, in one unambiguous sentence? If yes, and you can absorb failures, consider outcome.
- If not, does usage vary strongly across customers? If yes, use credits with a published conversion, or usage with a spend cap.
- If usage is stable and the agent is one feature among many, per seat is still fine, and simplest to sell.
- If buyers demand predictability but you need cost coverage, use a hybrid: a base fee for the relationship plus an allowance, with overage priced transparently.
What ages badly here
Two things in AI pricing change faster than the rest of your strategy. Inference costs have fallen repeatedly, which means a credit ratio set today may be needlessly expensive for customers in a year and should be revisited on a schedule rather than left to drift. And the vocabulary is unstable: "agent", "resolution" and "task" mean different things across vendors, so define your own terms in your contract instead of borrowing theirs.
The wider model landscape, including flat, tiered and freemium, is covered in the 9-step guide, where model selection is step 8 and rests on the cost work done in step 5. If you want to test what your buyers actually consider acceptable before committing, the free Van Westendorp tool gives you a price corridor from your own respondents.
Frequently asked questions
Should I show prices publicly for an AI product? Yes for self-serve tiers, and yes for the unit price of usage or credits. Hiding the unit price is read as a warning sign by exactly the technical buyers who evaluate AI tools.
How do I move from per-seat to usage without losing customers? Grandfather existing contracts through their current term, publish the new model with a calculator showing what a typical account pays, and offer a cap for the first period. The mechanics of announcing that change are the same as for any price increase.
What if my costs drop sharply after I set the price? Then you have a choice to make deliberately: keep the margin, or pass some of it on and say so publicly. Passing it on is a strong retention story, but only if you announce it. A silent margin increase is a decision too, just not one your customers will thank you for later.