AI-native products are changing the monetization game for good

Look at how the most visible AI-native companies charge today and a pattern shows up quickly. Fin charges per outcome. Salesforce charges per agent action. Replit charges by the effort behind each request. Each of them prices the work the product does.

None of these models is perfect, and customers have pushed back on all of them at some point. They are still useful to study, because each one answers a question every AI company runs into: what exactly is the customer paying for, and how do you keep that price connected to what the work costs you?

Here are four patterns worth borrowing.

1. Charge for a unit the customer already counts

A support leader already tracks resolved conversations. A sales leader already tracks qualified leads. When your price uses the same unit, the buyer can compare it directly to what they spend on the same work today.

Fin, the AI agent originally built by Intercom, is the clearest example. It charges $0.99 per outcome, bills at most one outcome per conversation even when the agent takes several actions, and prices a qualified sales lead separately at $9.99.

The hard part of outcome pricing is the definition. Fin counts a resolution in two ways: the customer confirms the answer helped, or the customer leaves after the last answer without asking for more help. The second case is where most billing questions come from, because a customer who leaves may have given up. Any outcome price you set depends on a definition of "done" that both sides accept, and that your meter can record at the moment it happens.

What to borrow: pick a unit your buyer already reports on, write the definition of "done" before launch, and publish it.

2. Put credits between the price and the cost

An AI agent does different kinds of work, and each kind costs you a different amount to run. Credits let you sell one currency and decide separately how much of it each action consumes.

Salesforce moved Agentforce to this model in May 2025. The original price was $2 per conversation, which buyers found hard to forecast. Flex Credits are sold in packs of 100,000 for $500, and a standard action such as updating a record or resolving a case consumes 20 credits, or $0.10. Voice actions consume 30 credits. Customers track their balance in a Digital Wallet that shows consumption by agent.

The useful property is the rate card. When a model gets cheaper, or a new action type costs more to run, Salesforce can adjust how many credits that action consumes. The price of a credit stays the same, so the customer's contract stays the same.

What to borrow: sell credits at a stable price, keep the burn rate per action in a rate card you can update, and give customers a live view of their balance.

3. Give customers a predictable base and meter the rest

Pure usage pricing is fair and hard to budget for. Pure subscription pricing is easy to budget for and exposes you to your heaviest users. Most AI-native companies now combine the two.

Replit is a good example of how the pieces fit. Its Core and Pro plans are monthly subscriptions that include a set amount of credits, and Agent work beyond that draws on usage. In June 2025, Replit replaced its flat $0.25 per checkpoint with effort-based pricing: each request now produces one checkpoint, priced by the work the agent did to complete it. Small fixes cost a few cents. Large builds can cost several dollars.

Effort-based pricing tracks cost closely, and it has drawn criticism because customers see the price after the work runs. Replit's answer has been spending alerts, budget limits, and lighter agent modes for smaller edits. That is the lesson for anyone building a hybrid plan: the variable part needs controls the customer can set before the bill arrives.

What to borrow: bill the base in advance, bill usage in arrears, and ship spend alerts and limits with the first version of the plan.

4. Expect to change the model, and plan how you will

Most of the companies in this post have changed how they charge. Salesforce went from per-conversation to per-action. Replit went from flat checkpoints to effort-based checkpoints. Cursor moved its Pro plan from request limits to usage-based limits in mid-2025, then apologized and refunded unexpected charges within weeks.

The changes that landed well had a few things in common. Salesforce kept per-conversation pricing available for customers who preferred it, so nobody was forced across on day one. Replit explained the change in public, with example costs for small and large requests. The change that drew the most criticism reached existing customers as a surprise on their bill.

Pricing changes are a normal part of running an AI product, because model costs and product scope keep moving. The companies that handle them well treat a new pricing model like a product release: versioned, announced, and rolled out in stages.

What to borrow: version your plans, let existing customers keep the old one for a set period, and show them what their bill would look like under the new one before you move them.

What this asks of your billing system

Put the four patterns together and the requirements are concrete. You need a meter that records an outcome at the moment it happens, using a definition you can publish. You need a credit rate card that changes independently of the price. You need a plan that bills a base in advance and usage in arrears. You need plan versions, so existing customers can stay on the old model while new ones start on the new one.

Most teams start with a homegrown setup, and each of these adds another code path to it. That is usually when pricing work starts to slow down.

That's why we built Clocknext; to fit your billing, regardless of model. A single plan can combine credits, outcomes, units, wallets, and flat fees. Each component can be billed in advance or in arrears. Provider cost is recorded with every signal, making it possible to track margin per customer, per agent, and per model.

If you want to experiment with different models and find the right model for your product, talk to us.