A tiny group of AI users can drive most of a company's costs

Pricing expert Ulrik Lehrskov-Schmidt says a tiny group of heavy users can drive most of what AI costs a software company. At one of his clients, the costliest customer used 380 times more in AI processing than it paid in revenue. He has worked on pricing full time for about ten years and runs the firm Willingness to Pay.

What he found

Across a couple dozen high-growth AI pricing projects, his team sees AI costs bunch up at the top. In a product with 1 million users, he says, 1,000 people can generate three quarters of the bill for running the AI. He links this to gross margins, the share of revenue left after direct costs, falling from 85-90% to about 55% at some AI companies.

Cost and value are separate

Many firms try to pass AI costs on to customers. Lehrskov-Schmidt says cost and value are independent. The 380x customer sat in a category where the value was not there to charge for, so a price rise would have caused immediate churn, meaning the customer would have left.

The reverse can also hold. Another client with roughly $500K contracts worried about AI spend. Its largest accounts used about $20K a year in AI costs. He called that a good business.

The fix: brakes, not just prices

When costs are spiky, he recommends fair usage limits that throttle heavy use, whether or not the company charges more. His method:

• Split users into groups, such as new users, fraudulent users and everyone else.
• Set a kill point for each group, the usage level where the product shuts off.
• Add brakes that slow the product at a threshold, like halving speed or adding a short wait.
• Tell users by email, in-product message or a notice through the MCP.
• Tighten slowly, starting with about 0.01% of users, then 0.1%, and track who keeps going, uses less, leaves or upgrades.

At around 4% of users affected, he says a company might have fixed 60-80% of its cost problem. His team has seen companies recover up to 22 points of gross margin within weeks, with throttles that touched a few percent of users and no new pricing.

Higgsfield and Figma

A few days before the session, Higgsfield's pricing page showed plans at $19, $49 and $99, with several features listed as unlimited. Lehrskov-Schmidt says that caps revenue per user while leaving cost uncapped. He did not know whether Higgsfield has fair usage limits and said it should if it does not.

Figma gave AI to everyone but built fair usage limits before launch. After heavy use, it turned on credit counting, then enforcement, then monetization. By his account, gross margins fell from 92% to 82%, but revenue grew faster than the costs.

What to check

Anyone selling an AI product on a flat monthly price should look at how usage is spread across customers. Buyers of AI tools can expect limits on heavy use, even on plans marked unlimited.