The Monetization Problem in AI

AI adoption is spreading quickly across industries, but turning those products into reliable revenue streams is another matter. Survey data from Stripe shows 63% of tech respondents and 31% of nontech respondents already sell AI products, with another 45% of the latter planning to do so. That growth, however, hides a hard reality: the costs of running AI models scale fast, and the value delivered is often difficult to quantify.

On top of that, AI companies face the same monetization pressures as any software business—predictable revenue, customer acquisition, and margin protection. The companies that handle this well tend to make deliberate choices around three questions: what value metric to charge for, how to structure billing to balance predictability and adoption, and what guardrails to put in place against usage risk.

Choosing the Right Charge Metric

A charge metric is the incremental unit that determines what a customer pays. It is not new—phone plans charged per minute, cloud storage per gigabyte—but AI makes the choice harder. The metric must line up with both the value customers receive and the incremental cost of delivering it, and neither is easy to pin down.

Value and usage are not the same thing. Intercom’s support agent Fin, for example, only provides real value when it resolves a customer’s ticket. A conversation that ends without a resolution still costs the same in model calls, but it delivers nothing worth charging for. Identifying these outcome-driven moments is the first step to setting a defensible price.

Cost is the other side of the equation. A task like “resolve a ticket” can take one long conversation or several short ones, and the number of model calls can vary wildly. The harder the outcome is to define, the harder it is to predict what it costs to deliver.

Three types of charge metrics dominate AI pricing today, each trading off cost alignment against value alignment:

  • Consumption-based (per API call, per token): Closely tied to infrastructure costs, so margins are predictable for the seller, but value can be opaque to the buyer. Most common with AI model providers, used when customers want granular control over what they consume.
  • Workflow-based (per task): More variable on cost, but easier to justify. A fixed price for a defined job—booking a meeting, analyzing a spreadsheet—has clear time-saving value that customers understand instantly.
  • Outcome-based (per success): Highest cost variability, but the strongest link to business results. Intercom chose this route, pricing per successfully resolved ticket despite the risk it places on cost to deliver.

Balancing Predictability Against Growth

Once the metric is set, the next question is how to bill. Pure usage billing is possible, but the data shows most companies do not go that way: 56% of AI company leaders in Stripe’s survey said they used a hybrid model combining subscriptions and usage-based fees, against 38% for pure usage pricing.

The reason is that hybrid models give companies multiple levers to manage two competing priorities:

Revenue predictability. Recurring subscription fees make forecasting easier, which supports long-term investment decisions. Enterprise buyers often prefer a stable monthly figure for the same reasons.

Customer acquisition and growth. Shifting more revenue into usage-based fees lowers the barrier to entry. Customers can try a product with minimal commitment and only start paying more as they get value from it. That suits early-stage companies or product-led growth motions.

Browserbase, which hosts headless browsers for AI agents, built its pricing around this tradeoff. Tiered subscriptions each bundle a set amount of usage aimed at a different developer persona, a free tier lets prospects test the product, and clearly posted overage rates allow customers to expand without immediately moving to a more expensive plan. Enterprise customers get a custom plan.

Guardrails Against Usage Risk

Any model with variable usage carries risk, and that risk is highest with pure pay-as-you-go pricing, though subscription-plus-overage models still face it. Malicious abuse is a bigger concern than honest overuse, since costs can spike sharply before anyone notices.

The right protections depend on where the risk lives:

  • For customers who might accidentally exceed their budget, usage thresholds with alerts prevent bill shock.
  • For fraudsters who run up charges they never pay, requiring upfront purchase of credits that draw down over time cuts the exposure.
  • For workloads that can spike unexpectedly, rate limits and usage caps keep spend controlled while the customer adjusts.

Communication matters as much as the mechanisms themselves. Customers need visibility into how usage converts into spend, and a way to respond quickly when that spend climbs faster than expected.

Pricing Is Never Final

Mature AI companies do not treat pricing as a one-time decision. Intercom, for instance, currently charges a flat rate per resolved ticket but is exploring tiered pricing based on ticket complexity.

That reflects broader behavior: among companies selling AI products with any usage-based component, 92% said they had adjusted their pricing after launch. Underlying model costs shift, markets mature, and customer expectations move quickly. Pricing structures need to keep up.