The Pricing Page Is the Product: A Field Guide to Charging for AI

Software used to have a handy economic quirk: once you'd built it, serving one more customer cost almost nothing. That's why per-seat pricing took over. You charged each user a monthly fee, and heavy users were simply happier customers.

AI breaks that quirk. Every question, draft and agent run costs real money in model usage. A customer who uses your product ten times as much now costs you roughly ten times as much to serve. Like it or not, your pricing page has become a statement about who carries that cost.

The four main models

Nearly every AI product uses one of four pricing models, or a blend of them.

No single answer is right for everyone. There is a right answer for your product, and it depends on how predictable your costs are and how clearly customers can see the value.

A worked example

Quillmate, a fictional tool that drafts proposals for marketing agencies, launched at $49 per seat per month. Most customers used it a few times a week. A handful of agencies used it to draft every proposal they wrote, plus pitch decks and case studies.

The founders ran the numbers after six months. Their median seat cost them about $6 a month in model usage. Their top 5% of seats cost more than $70, which was more than the subscription itself. They were, quite literally, paying their best customers to use the product.

They didn't punish the heavy users. They redesigned the plan:

• $39 per seat, including a generous monthly allowance of drafts.
• Credit packs for agencies that need more, priced with a healthy margin.
• A team plan with pooled credits, so one busy writer can borrow from quieter colleagues.

Revenue per customer rose slightly and margins on the heaviest accounts recovered. To the founders' surprise, cancellations also fell, because the cheaper entry price brought in smaller agencies that had balked at $49.

Five questions before you publish a price

1. What does your 95th-percentile customer cost you? Averages hide the customers who'll sink you, so model the heaviest users first.
2. What unit does your customer think in? Agencies think in proposals, support teams in tickets, lawyers in documents. Price in their unit, not yours.
3. Can the customer predict their bill? Nobody likes a surprise invoice. If you charge by credits or usage, show a running total and send a warning before limits are reached.
4. What happens when model costs fall? They often do. Decide whether you'll pass the savings on, add features or keep the margin, and be consistent about it.
5. Where's the value ceiling? If your tool saves an agency ten hours a month, that's the most they'll pay, whatever your costs.

Common mistakes

• Unlimited plans. "Unlimited" invites the one customer who'll run your product in a loop all night. If you must offer it, define fair use clearly.
• Pricing on tokens. Tokens are your cost, not your customer's value. Almost nobody outside engineering wants to budget in them.
• Hiding the AI cost entirely. Some founders bury model costs in a high flat fee, then struggle when a competitor offers a cheaper entry plan.
• Changing prices too often. Every change costs trust. Get the model right, then adjust the numbers, and do that rarely.

The one-sentence test

Before you ship your pricing page, try to finish this sentence: "Customers pay us more when ___, which is also when they get more value." If you can fill in the blank honestly, you've found a pricing model that will survive contact with real customers. If you can't, keep working on it, because the pricing page is part of the product.