Artificial intelligence companies are still borrowing prices from the software era they are disrupting. Per-seat subscriptions, token meters and vaguely defined credits may be convenient to invoice, but they often hide the economic question a buyer is trying to answer: what useful work did the system complete?
That gap is becoming a sales problem. AI changes both the cost of delivering software and the unit a customer values. A product that resolves support tickets, reviews contracts or creates campaign variations is not merely giving another employee access to a tool. It is supplying part of the work itself.
Founders should stop treating pricing as a packaging exercise performed after product-market fit. In AI, pricing is part of the product strategy because it defines what the company believes it can measure, promise and defend.
Seats and tokens describe inputs, not value
The per-seat model worked when software helped a known number of people do their jobs. Automation breaks that relationship. If an AI support system handles more conversations, a customer may need fewer human seats even as the product creates more value. Charging only by seat can punish the vendor for succeeding.
Token pricing has the opposite problem. It makes an application’s internal consumption visible to a customer who may have no reason to care. Tokens are a rational unit for raw model access. They are a poor proxy for the value of an application that also supplies data, integrations, orchestration, quality controls and workflow design.
Andreessen Horowitz reported on Aug. 27 that 27 of 50 technical AI buyers preferred credits tied to recognizable work, while 14 preferred tokens. Its recommendation is simple: price model access in tokens, useful work in recognizable units and attributable business results as outcomes.
Silicon Valley Bank’s 2026 enterprise software survey points in the same direction. Among more than 120 venture-backed companies, 37% used subscription-only pricing, but only 26% expected to remain subscription-only as companies considered usage and outcome-based models.
The best unit is visible on both sides
A good pricing unit is something the customer recognizes and the vendor can verify. That might be a resolved ticket, a processed invoice, a qualified research file, a completed compliance review or a published product description that passes approval. The unit should map closely enough to value that a buyer can forecast spend without learning the vendor’s architecture.
Outcome pricing goes further by tying payment to a business result such as revenue, savings or recovery. It can be powerful when attribution is clear. It can also create arguments when marketing, seasonality, staff performance and other systems contribute to the same result. Not every product can defend a pure outcome fee.
For many companies, the practical answer is a hybrid: a platform fee that supports integration, security and service, plus a variable charge for recognizable work. The fixed component pays for readiness; the variable component aligns expansion with use. A minimum commitment can protect the vendor’s capacity planning without forcing the customer into an oversized contract.
Credits can support that model if they remain intelligible. A credit representing one completed document or one minute of analyzed audio is useful. A credit that merely disguises a shifting number of tokens recreates the opacity buyers already dislike.
Pricing has to survive falling inference costs
Model costs continue to move, and competition can make them fall quickly. An application priced as a markup on tokens risks turning technical efficiency into a demand for an immediate discount. An application priced around completed work can share the benefit: the vendor improves its margin while the customer receives the same result at a predictable price.
That does not excuse vendors from monitoring unit economics. Some workflows trigger long reasoning chains, tool calls, retries and human review. Companies need contribution-margin data at the task level before promising an unlimited plan or a fixed price for an unpredictable outcome.
A separate a16z analysis of AI price wars warns that new entrants can use subsidized inference to undercut competitors. Matching every low offer can win a contract while destroying the business that must fulfill it. A founder should know the lowest sustainable price, the service level attached to it and which premium capabilities justify a higher tier.
Marketers can make the value legible
Marketing leaders have a central role because the pricing unit becomes the product story. “One million tokens” describes consumption. “One thousand compliant product pages ready for review” describes a business capability. The second message is easier for an executive to compare with current labor, agency and delay costs.
That requires honest boundaries. If the system produces drafts rather than approved work, the metric should say so. If human review remains essential, include it in the value calculation. Overstating automation may accelerate the first sale and undermine every renewal that follows.
Business-development teams should test pricing during pilots, not after them. Agree on the unit, observe variance and identify who owns the underlying business metric. Procurement can then compare options on a common basis, while the vendor learns whether its proposed unit remains stable across customers.
The right price will vary by product, but the principle is durable: charge at the highest layer of value the company can reliably measure. AI sellers that price the customer’s work instead of their own ingredients will make budgets clearer, sales conversations sharper and growth more defensible.