There is a specific, avoidable mistake that AI founders keep making, and it's baked into the pricing page before the product ships. You charge per seat — because that's what every SaaS company you admire does — and then you build an agent whose entire job is to make seats unnecessary. You have just wired your revenue to move in the opposite direction of your value.
That's not a vibe. It's arithmetic, and buyers now do it before they'll take your demo.
Per-seat: the model that fights your product#
Seat pricing worked for a generation of software because software made people more productive without replacing them. One more analyst, one more Salesforce seat. The headcount and the license grew together.
An agent breaks that link. If your product lets a ten-person support team handle the same volume with three people, a per-seat contract shrinks as the product succeeds. You are, quite literally, paid to under-deliver — the more work you automate away, the less you're owed. Procurement teams have figured this out. Across the market, seat-only pricing fell from roughly 21% to 15% of SaaS companies in a single year, and seat-only AI vendors increasingly get disqualified from deals before a demo is even booked.
Per-seat isn't dead. If humans sit inside your product all day — a coding copilot, a writing tool, a design assistant — a seat still maps to real, recurring value. The trap is using it reflexively for an agent that works instead of a person.
Usage-based: honest, and jumpy#
The obvious correction is to charge for what the thing actually does: tokens consumed, tasks run, actions taken, agent-hours burned. Usage pricing has one enormous virtue — it tracks your own cost. When a customer hammers the product, your infra bill and your invoice move together, so a runaway user can't bankrupt you. It also scales with adoption instead of against it, which is exactly the property per-seat lacks. (If you don't yet know how your own costs scale, that's the first thing to fix — because AI agent costs scale quadratically, not linearly, and a usage price set against a linear mental model quietly loses money on your heaviest accounts.)
Usage pricing aligns your invoice with your cost. It does nothing to align either one with the customer's sense of worth.
The cost is predictability — on both sides. The customer gets an invoice that swings with a number they can't forecast, and "bill shock" is the fastest route to churn in usage-priced products. You, meanwhile, get revenue that follows their usage curve, which means it dips when they're quiet and spikes when they're busy, none of it on your schedule. That two-sided volatility is why almost nobody ships pure usage pricing anymore — and why, once you do meter, getting the metering plumbing right matters as much as the number you charge.
Outcome-based: charge for the win#
The sharpest alignment is to charge only when the product works. Intercom's Fin bills $0.99 per resolved ticket and nothing on the ones it escalates. Sierra runs pure outcome pricing — it gets paid only when its agent resolves the issue without a human — and rode that model past $150M in ARR by early 2026. Salesforce, hedging, now runs three pricing models for Agentforce at once, including a $2-per-conversation tier and a $0.10-per-action credit system.
Outcome pricing is the most honest promise a vendor can make: no result, no charge. It's also the hardest to operate, for two reasons. First, you have to define and measure the outcome cleanly — "resolved" has to mean resolved, not "the user gave up," or you'll be charging for failures and losing the trust the model was supposed to buy. Second, you absorb the full cost of everything that doesn't work: every failed attempt, every fuzzy edge case, every dispute about whether the outcome really happened. You're now underwriting your own quality. That's a fantastic incentive and a real balance-sheet risk, and it only pencils out where a "win" is unambiguous — a resolved ticket, a booked meeting, a completed transaction.
The 2026 default is hybrid — and that's fine#
Here's the un-dramatic answer most successful AI companies land on: a platform fee plus a meter. A fixed monthly base covers access, onboarding, and support and gives you a revenue floor you can forecast. On top of it sits a usage or outcome meter that captures upside from your heaviest customers. An included allowance in the base fee means most customers never touch the meter, so their bill feels flat — the predictability of a subscription — while your power users pay for the value they extract.
Hybrid is now the industry standard, adopted by something like 41% of AI vendors in 2026 versus roughly 27% the year before. Even OpenAI moved workspace-agent runs to token-metered pricing in July 2026 while leaving the base ChatGPT subscription intact — a platform fee with a meter bolted on, by another name.
Pick your model by asking one question: does my price go up when my product gets better, or down? Per-seat answers "down" for anything that automates work. Usage answers "sideways." Outcome answers "up," if you can define the outcome. Choose accordingly — and if you can't yet, ship a platform fee with a meter and earn the right to get more precise later.
