A token paywall is not SaaS

A founder ships an AI product and prices it the way they price everything: flat, $20 a seat a month, unlimited use, just like SaaS. It works great — until customers actually use it. The power users, the ones who love it most, quietly become the ones you lose money on. The instinct is to fiddle with the price. The real problem is one level deeper: you priced an AI product as if it were SaaS, and it isn't.

What made SaaS magic — and why AI doesn't have it

The reason software was such a beautiful business is one number: near-zero marginal cost. You build it once, and the ten-thousandth customer costs you almost nothing more than the first. That's the whole engine behind 80–90% gross margins and "grow now, the unit economics take care of themselves."

AI doesn't have that engine. Every single interaction burns tokens, so your cost to serve scales with usage and never flattens — sometimes it scales super-linearly. The numbers are stark and consistent now. ICONIQ's 2026 State of AI report finds AI-native products averaging ~52% gross margins versus 75–90% for mature SaaS, with inference alone eating 23% of revenue at scaling-stage companies. And as one analysis put it, this gap "isn't a temporary inefficiency — it's an architectural consequence of inference costs." You will not optimize your way back to SaaS margins, because the cost is in the physics, not the waste.

The flat-fee trap, with receipts

A flat monthly price on a metered cost is a bet that your customer stays light. Your best customers exist to break that bet. In its early days, GitHub Copilot reportedly lost up to $80 per user per month on heavy users while charging a flat $10. Replit watched its gross margins go negative during usage surges before it changed pricing. For a typical $100-a-month AI product, it's now normal for $30–40 of that to disappear straight into tokens and compute — a 30-point hole that simply didn't exist three years ago. Flat pricing doesn't make that cost go away. It just hides it until your most enthusiastic users make it impossible to ignore.

The per-seat trap is even deeper

Per-seat pricing has a second, sneakier problem with AI, and it's almost funny. If your agent is good, it reduces the number of seats your customer needs. The whole pitch of an AI support agent is "you don't need 50 support reps, you need 5." So if you charge per seat, the better your product works, the less you make. You've tied your revenue to the exact headcount you're bragging about eliminating. Succeed fully and you price yourself to zero. No amount of discounting fixes a model that punishes you for being good at your job.

AI products are closer to a utility than to software

Here's the reframe that makes the pricing obvious. A traditional SaaS product has almost no cost of goods sold. An AI product has a real one — tokens — under every single transaction, the way an electric utility has fuel under every kilowatt or a manufacturer has materials under every unit. That's not a worse business; it's a different one, and it has to be priced like what it is: pass the variable cost through. That's why 92% of AI software companies now use mixed pricing — a base fee plus usage — and why GitHub moved Copilot itself to usage-based billing on June 1, 2026. Outcome-based pricing (charge for the result, not the seat) is emerging for the same reason: align what you charge with what each customer actually costs and delivers.

And this is where the business meets the engineering, which is the part I care about most. When tokens are your cost of goods, efficiency is margin. Every token you don't burn is money you keep — so routing the boring work to cheap models, grounding so the agent doesn't flail, and keeping agents narrow aren't engineering niceties anymore. They're the difference between 52% and 32% gross margin. The discipline that controls your token bill is the discipline that keeps your company alive.

The one fatal mistake

There's no single correct pricing model — mixed models are winning precisely because the right answer depends on your product. But there is one fatal mistake, and it's the SaaS reflex: pricing as if serving the next user is free. It isn't, it never will be, and the magical economics of software simply don't apply to the part of your product that thinks. AI isn't SaaS. Put a real cost of goods in your model, or your most enthusiastic customers will be the ones that sink you.