The thesis. Traditional software was costly to build and nearly free to run, so a flat price per seat worked. Artificial intelligence (AI) products incur a real compute cost every time they are used, which turns pricing from a marketing decision into a margin decision. The question for leadership is who bears the cost when usage rises, and who captures the value when outcomes improve.
Why AI breaks the seat model
For two decades, enterprise software economics rested on one fact: once a product was built, an extra user cost almost nothing to serve. A price per seat captured value well enough, and margins took care of themselves.
AI changes that. Each query, document or agent action consumes compute that someone pays for. Unit prices are falling fast. Stanford’s 2025 AI Index reports that the inference cost of a system performing at the level of OpenAI’s GPT-3.5 fell more than 280-fold between November 2022 and October 2024 [1]. Andreessen Horowitz estimated that, for a model of equivalent capability, inference cost has been falling by roughly ten times per year [2].
Falling unit cost does not mean falling total cost, because usage is rising faster. Google said in May 2025 that it was processing more than 480 trillion tokens a month across its products and application programming interfaces, up from 9.7 trillion a year earlier, a fifty-fold increase [3]. (A token is a small unit of text that a model reads or writes.) A product that completes a task in several model steps pays for every step. The cost of serving a heavy user can therefore differ widely from the cost of serving a light one, and both can shift from quarter to quarter.
The margin effect is already visible in growth-stage companies. Bessemer Venture Partners found that the fastest-growing AI companies it studied averaged only about 25 percent gross margins, often accepting lower profit in exchange for faster distribution. A second group of high performers with more conventional growth averaged about 60 percent, which Bessemer describes as slightly below software peers because of model-related costs [4]. Pricing that ignores compute shows up as margin erosion, not as a line item anyone owns.
Three pricing models, three different bets
Each pricing model allocates two risks: cost risk (who pays when usage rises) and value risk (who loses when results disappoint).
| Model | Cost risk sits with | Value risk sits with | Works best when |
|---|---|---|---|
| Per seat | Seller | Buyer | Use per user is steady |
| Per use | Buyer | Buyer | Use tracks value closely |
| Per outcome | Seller | Seller | Outcomes are clear and attributable |
Per seat
Seat pricing is familiar, easy to budget and easy to sell. Microsoft priced Microsoft 365 Copilot at US$30 per user per month when it announced commercial pricing in July 2023 [5]. For buyers, the appeal is predictability. For sellers, the risk is that a small share of heavy users consumes most of the compute. Sellers usually respond with fair-use limits, tiered seats or premium plans for heavier workloads, which gradually turns seat pricing into usage pricing by another name.
Per use
Usage pricing aligns revenue with cost. It charges per call, per action, per document or per credit. Salesforce, which had priced its Agentforce agents per conversation, introduced Flex Credits in May 2025, with each agent action priced at US$0.10 [6]. For sellers, the appeal is margin protection. For buyers, the risk is an unpredictable bill, which slows procurement and invites usage caps that limit adoption. Usage pricing works best when each unit of use maps clearly to something the buyer values.
Per outcome
Outcome pricing charges only when the product delivers a defined result. In August 2024, Zendesk introduced outcome-based pricing for its AI agents, charging only for issues the AI resolves without a human [7]. For buyers, the appeal is clear: they pay for value, not activity. The seller takes on both the cost risk and the value risk. That requires three things many sellers do not yet have: an outcome definition both sides accept, instrumentation that measures it reliably, and enough confidence in performance to carry the downside.
Hybrid is the default, for good reason
Most companies will combine models rather than choose one. ICONIQ’s 2025 survey of 300 software executives found many using hybrid pricing, typically a base subscription combined with charges based on usage, and 37 percent planning to adjust their pricing within the next year [8]. The same report notes that as AI products mature, cloud, inference and governance costs become the largest share of spend [8].
A sensible hybrid has three parts:
- A platform or seat fee that covers fixed costs and secures the relationship.
- An included usage allowance sized to typical use, not heavy use.
- A usage or outcome component above the allowance, priced to protect margin and tied to a unit the buyer recognizes as value.
The go-to-market consequences
Pricing choices reach well beyond the price list.
- Sales compensation. If revenue depends on consumption, quota credit at signature rewards the wrong behaviour. Commission should reflect realized usage or outcomes.
- Forecasting. Usage and outcome revenue are less predictable than seat revenue. Finance teams need consumption data and cohort models, not only bookings.
- Customer success. Under usage or outcome pricing, adoption is revenue. Customer success becomes a growth function rather than a retention cost.
- Partnership structure. Where a partner resells or embeds the product, the split of compute cost and outcome risk must be written into the agreement, or margin leaks at the boundary.
Much of my work is translating technical capability into value that clients, investors and partners can price. In AI, the hardest part of that translation is usually agreeing on the unit of value. When both sides agree on what a resolved case, a processed claim or a qualified lead is worth, the pricing model follows. When they cannot agree, no model will hold.
Questions for the leadership team
- What is our fully loaded compute cost per unit of customer value, and how widely does it vary between light and heavy users?
- Which pricing model allocates cost and value risk in a way we could carry at three times today’s usage?
- Can we define and measure an outcome our customers would accept as the basis for a price?
- Do our sales incentives reward adoption and outcomes, or signature alone?
- If unit compute prices fall again next year, do we pass the savings on, keep them as margin or reinvest them in capability?
References
- N. Maslej et al., “The AI Index 2025 Annual Report,” AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, Apr. 2025.
- G. Appenzeller, “Welcome to LLMflation: LLM inference cost is going down fast,” Andreessen Horowitz, Nov. 2024.
- S. Pichai, “Google I/O 2025: From research to reality,” Google, May 20, 2025.
- K. Bennett et al., “The State of AI 2025,” Bessemer Venture Partners, Aug. 2025.
- Y. Mehdi and J. Spataro, “Furthering our AI ambitions: Announcing Bing Chat Enterprise and Microsoft 365 Copilot pricing,” Microsoft, Jul. 18, 2023.
- Salesforce, “Salesforce introduces new flexible Agentforce pricing to accelerate the digital labor revolution,” press release, May 15, 2025.
- Zendesk, “Zendesk first in the CX industry to offer outcome-based pricing for AI agents,” press release, Aug. 28, 2024.
- ICONIQ, “2025 State of AI Report: The Builder’s Playbook,” 2025.