How to Price an AI Product: Seats, Usage or Outcomes

To price an AI product, choose a unit that tracks both the value the customer gets and the compute you consume, then test the price against your heaviest users, not your average ones. Most AI products now combine a base fee with metered usage, because flat seats lose money when customers run agents and outcome pricing works only when your product succeeds reliably. The right answer depends on who should carry the variance in compute cost: you, your customer, or both.

The industry learned this through expensive mistakes. In December 2024 OpenAI launched ChatGPT Pro at $200 a month, ten times the price of ChatGPT Plus. In January 2025 Sam Altman said OpenAI was losing money on it, because people used it more than the company expected (company statement). The most expensive consumer plan on the market was unprofitable within weeks, because the customers who choose a $200 plan are the ones who use it most.

Why is AI pricing different from SaaS pricing?

Software-as-a-service settled in the 2000s on charging per user per month at gross margins near 75–85%. Generative AI breaks that formula in three ways. Its marginal cost is real and can vary a hundredfold between users. Its unit of value is moving from a tool that helps a person to work that replaces one. And its cost floor falls so fast, under such competition, that a price set this year may be wrong next year. The inference cost explainer covers the cost side in detail.

Every AI price therefore answers four questions, which The Intelligence Equation sets out in Chapter 13:

1. What is the unit of value?
2. Who carries the variance in compute cost?
3. How much of the value does the price capture?
4. What happens to all three when model prices fall, a better model costs more per attempt, or your supplier enters your market?

Step 1: Choose your pricing model

The industry has tried six models, usually in combination.

The model allocates risk. Tokens and usage pass the volatility of compute cost to the customer. Seats and subscriptions keep it with you. Outcomes keep it with you and also make you answerable for quality. Most large companies sell several at once: about 80% of Anthropic's first-half 2026 revenue came from usage-based billing and about 20% from subscriptions (S-1 coverage).

Step 2: Model your margin at real usage

Before you set a price, model at least three users: light, typical and heavy. The book's stylized example uses a coding-assistant seat at $40 a month, a frontier model at $9 per million blended tokens, $3 of hosting, $1 of support and 3% payment fees.

A light user is a good software customer. A typical user already makes the seat unprofitable on a frontier model. Routing traffic to a cheaper model rescues the typical user but not the heavy one. Only a change in the price unit does: if the seat includes five million tokens and the rest is billed at 1.25 times model cost, the heavy user pays $265 a month and the seat earns 11% on the frontier model, or 53% with routing.

That is the logic behind Cursor's June 2025 pricing change. It replaced a fixed allowance of 500 "fast requests" on its $20 Pro plan with usage billed at API rates. Heavy users' bills rose, complaints followed and in early July Cursor apologized and refunded unexpected charges. Its gross margin had been reported at about −23% in early 2025 (The Information, reported). By April 2026, after repricing and shifting work to its own models, it reached "slight" gross-margin profitability (TechCrunch, reported).

The rule: check that your price covers cost at the heaviest decile of usage, not the median.

Step 3: Decide whether seats will shrink your revenue

Seats remain the largest enterprise model because they fit existing budgets. Microsoft 365 Copilot launched at $30 per user per month in November 2023 and reported "over 30 million" paid seats in July 2026 (company-reported). But seats have two structural weaknesses. Usage is uncapped while revenue is fixed. And successful automation reduces seats: if AI lets a customer cut its support team by 30%, seat revenue falls by 30% even though you are doing more work.

Salesforce faced exactly this. Its revenue came from seats, and its Agentforce agents would shrink the number its customers needed. It launched Agentforce at $2 per conversation in October 2024, then on 15 May 2025 moved to Flex Credits at $0.10 per action, charged only when an action runs. A "conversation" could be one question or a long session, so it tracked neither cost nor value. Agentforce annual recurring revenue rose from about $100 million in May 2025 to $1.2 billion in May 2026 (company). Salesforce then bought Fin, formerly Intercom, the per-resolution rival, for about $3.6 billion in September 2026.

If you sell automation on seats, change your metric before your customers change their headcount. Incumbents have one extra option: fold AI into a suite and raise the base price, as Microsoft did with a $3-a-month US increase for consumer Microsoft 365 plans in January 2025.

Step 4: Use usage pricing, but control bill shock

Usage pricing ties revenue to activity that tracks cost, and it spread from developer tools into business software in 2025. GitHub added metered "premium requests" for advanced models, and Cursor moved to usage pools. It has three known problems:

• Unpredictable bills. Customers cannot forecast spend, so add prepaid pools, caps and alerts.
• Paying for failure. The customer pays for actions or tokens even when the agent fails.
• Opaque credits. Credits with conversion rates obscure the real price, which helps in the short run and costs trust later.

Cursor's refunds showed that the reputational cost of an unclear change can exceed its financial benefit. Announce changes early, show customers their usage, and grandfather where you can.

Step 5: Consider outcome pricing if you can measure success

Sierra, founded in February 2023 by Bret Taylor and Clay Bavor, sells AI customer-service agents to large consumer brands and charges when an agent resolves a case or achieves another agreed result, such as a saved cancellation. The logic is that help-desk software priced per human agent punishes the vendor for automating, while the larger pool of money is the labor cost of each contact. Sierra said it reached about $100 million of annual recurring revenue in seven quarters and "$400M+" by 30 October 2026 (company claims, unaudited).

Outcome pricing made measurement part of the product. Sierra built analytics and testing so both sides could agree what had happened in each conversation. It also leaves the vendor paying for every conversation it does not get paid for. Its rival Decagon argued in December 2024 that outcome pricing invites disputes over what counts as "resolved"; it offers both meters and says most customers choose a fixed rate per conversation because it is predictable.

The book's stylized three-year comparison shows the trade-off. A brand sends 1 million support conversations a year to an AI agent; vendor A charges $0.80 per conversation, vendor B $1.50 per resolution. Resolution rates rise from 45% to 70% while compute cost per conversation halves each year from $0.50.

Both vendors keep the falling cost of compute as margin. Only the per-resolution vendor also keeps the gain in capability: its revenue rises by more than half as the agent improves. Outcome pricing is a bet on your own capability. Choose it only where resolution rates are high and rising.

Step 6: Decide who keeps the cost curve

Model prices fall fast for any fixed capability. If your price is tied to tokens, competitors will pass those savings to customers and your revenue per task will fall. If it is tied to actions or outcomes, you keep the savings as margin until a rival undercuts you. Decide deliberately, and watch whether competitors force the answer.

A pricing checklist

• Name the unit of value. Price as close as possible to what the customer would pay a person to do.
• Test the heaviest decile. The top few percent of users, often running agents, set the margin of any flat price.
• Allocate variance on purpose. Know who pays when usage spikes.
• Check whether success shrinks revenue. Seats plus automation is a slow leak.
• Write tight definitions. Outcome contracts are only as good as the definition of an outcome.
• Report honestly. Run-rates in usage-priced businesses can overstate the next twelve months.

The business models and pricing study collects the evidence behind each point, and our guide to how AI companies make money shows how pricing at the top of the stack flows down to labs, clouds and chips.

The takeaway

Pricing is where the second factor of the intelligence equation, unit economics that hold at real usage, is won or lost. Flat prices attract the users who cost most, as ChatGPT Pro and Cursor both discovered; seats shrink as automation works; usage pricing protects margin but risks bill shock; and outcome pricing pays only when the product succeeds reliably. Start from the unit of value, test it against your heaviest users, and decide on purpose who carries the cost of compute. Chapter 13 of the book, Business Models and Pricing, works through each model with its arithmetic.