How Should You Price an AI Feature?
AI features cost money every time someone clicks. Here's how to price them with credits, usage tiers, and outcome-based billing.
Shipping an AI feature has never been easier. One API key and you have summarization, support replies, or document generation inside your product. The hard part comes next: what do you charge for it? Traditional software had near-zero marginal cost — once built, every new user was almost free. AI features spend real money every time someone presses the button. That's why so many teams launch a flat monthly plan and discover six months later that their margin has quietly evaporated.
Why doesn't a flat monthly price work anymore?
Because usage variance between customers is extreme. Two users paying the same price can easily cost you an order of magnitude apart. Flat pricing has no mechanism to absorb that gap.
- Marginal cost isn't zero — every call burns tokens. As you grow, what grows isn't your server bill but your raw material bill.
- Heavy users eat the margin — a small fraction of accounts typically drives most of the consumption. Without a limiter, growth makes losses bigger, not smaller.
- Agents multiply calls per request — one user request can now trigger retrieval, tool calls, and retries, meaning dozens of model calls behind a single click.
- Your cost base keeps moving — switching models changes your unit economics overnight. Your price list needs to be able to respond.
Which pricing model should you choose?
There isn't one right answer — there's a right combination. In practice, a base fee plus usage is the safest default: the base fee gives you predictable revenue, the usage component protects you from cost blowouts.
- Per-seat / flat — works when consumption tracks headcount and variance is low. Easy to forecast, but defenseless against heavy users.
- Usage or credits — Zapier bills by tasks; developer tools like Cursor and Replit meter credits and usage. The underlying OpenAI and Anthropic APIs are themselves priced per token.
- Outcome-based — Intercom's AI support agent Fin charges per resolved conversation. It's the clearest value story for the buyer, but only works if both sides can agree on what "success" means.
- Hybrid — a monthly fee with included credits and paid overage. For solo builders and small SaaS teams, this is the realistic starting point.
What should you calculate before setting a price?
Start with your cost unit, not your price. Define the smallest customer action that spends money, then measure what one of those actually costs using your own logs.
- Pick a billing unit customers understand — one contract reviewed, one ticket resolved, one video generated. "Tokens" is not customer language.
- Measure real cost per unit — average across 100 real requests, including retries. Record the worst case (the longest document) separately.
- Set a margin guardrail — apply a multiple to your unit cost, but make sure even the worst-case request doesn't put you underwater.
- Define caps and overage — never say unlimited. Decide whether overage auto-charges or stops with a warning, and say so plainly.
- Cap the free trial by volume, not time — ten free runs is far more predictable than thirty unlimited days.
Say you're building a contract review tool. Each review triggers several model calls for summarization and clause comparison. Your logs show an average cost of about ₩300 per review, with the longest contracts hitting ₩1,200. A flat "₩1,000 per review" price is dangerous. A monthly base fee with 20 reviews included plus a per-review overage feels like a subscription to the average customer while capping your downside on the outliers.
Three landmines to avoid
- Opaque credits — if users can't see what an action will cost before running it, they get scared and stop using the product. Show an estimate up front.
- Charging for failures — burning credits on a bad output leads straight to refund requests and churn. Publish an automatic refund rule.
- Pricing from cost alone — cost is only your floor. Your ceiling is what the same job would cost the customer in human hours.
It comes down to one principle: sell in a unit your customer understands, in a structure that survives your costs changing. The first task isn't rewriting your pricing page — it's logging the true cost of one action in your product.