Harvey, the legal AI startup, spent part of this year losing money every time customers got more value from its product. According to Bloomberg, the company’s gross margin fell from about 50% at the start of the year to -50% by June (!). It was spending about $1.50 to deliver every dollar it sold.
The problem, as State of AI points out, was pricing. Harvey sells law firms a flat annual seat with unlimited use, but its AI providers charge by the token. Token usage rose twentyfold this year, while revenue only doubled. A partner running one query a week paid the same as an associate running an agent all night.
Harvey had a keel and no sails.
Think of a sailboat. The keel sits under the waterline keeps the boat steady. The sails drive it forward. A boat with only a keel goes nowhere. A boat with only sails tips over in the first strong gust.
Pricing works the same way. The subscription is your keel: a predictable base customers can budget for. Usage is your sails: revenue that grows as customers use more. For any software company paying for every token its new AI service consumes, each one protects against a different risk.
In a new analysis of pricing at about 200 B2B software companies, Bain found that most companies moving beyond seats are adding new meters on top of seats, not replacing them. About four in five vendors introducing AI pricing choose capacity models, where customers commit to a fixed amount up front. Bain says this gives vendors the stability of seat pricing and gives CFOs a commitment they can approve.
Now consider the opposite risk. The average price businesses pay per million tokens has fallen significantly this year. Cheaper tokens sound like good news for software companies who are selling AI services, and at first they are. But customers notice, and competitors cut their prices. If your revenue is priced purely on usage, it may fall along with token prices. That is when you need the keel.
Harvey eventually found a way out. In August, it launched its own model, built on a cheaper open-weight model. The company says the new model cut the cost per query on one feature by about 90%, and Bloomberg reports its margin turned positive again.
But look at what happened. In six months, heavy usage tripled Harvey’s cost to serve each dollar of revenue, then a model swap cut costs on some work by 90%. Talk about a financial storm! AI costs can swing hard in both directions. Harvey got lucky that a cheaper model came along in time. The next company may not.
So start with a subscription that reflects the ongoing relationship. Then add a usage layer that captures the value customers create as they grow. The keel protects you when token prices fall. The sails protect you when usage spikes. Together, they absorb the swings, whichever way they go.
There are some cases, of course, where committing to a single model makes sense. For example, I’ve seen lots of services that start off as pay-as-you-go and eventually formalize into a committed subscription. There will always be a push and pull between consumption and predictability.
But if you build both, you can sail in any weather.
