Pricing the Invisible
Strategy · July 4, 2026
The pricing page is the most honest document a company writes. Everything else says what you wish you were. The price says what you actually think you're selling, and to whom, and what you believe it's worth on a Tuesday when nobody is watching. Founders treat it as the last thing, a cosmetic decision after the product is built. It is the first thing. The number is the strategy, and the strategy is a bet.
It is a harder bet for an AI product than for anything that came before it, and the reason is simple. You are pricing something the buyer cannot see.
The invisible product
Sell a seat in a spreadsheet and both sides know exactly what changed hands: one login, one predictable thing, the same today as tomorrow. Sell the output of a model and you are selling a distribution, not an object. Sometimes it nails the task in one clean pass. Sometimes it wanders, retries, and gets there the long way. Sometimes it is confidently wrong. The customer cannot hold it, cannot count it, and cannot fully predict it before they have paid for it.
Old software was a thing you bought. An AI product is a promise you keep, unevenly, one request at a time.
So the pricing question is not "what is this worth." It is "how do I put a number on a promise whose cost I don't control and whose value the customer can't see until after the invoice." There are three doors out of that room. Each one aligns the price to a different thing, and each hands the risk that the model is wrong to a different party.
Three ways to charge, three different bets
Inherited from SaaS. You bill for who can log in.
The trap. The agent that replaces fifty people still bills like five.
You bill for what the model actually did.
The trap. You're reselling inference, and the floor is your supplier's price.
You bill for the result: the resolved ticket, the Tuesday done.
The trap. Your reliability is now your revenue, and you have to prove the result.
Per seat: the number that lies
Seat pricing is inherited furniture. It was built for a world of human users who log in, and it worked because in that world a seat was a fair proxy for value: more people using it, more value delivered, more you charge. Clean.
An AI product breaks the proxy from both ends. On the value side, the whole point is that one agent does the work of many people, so billing per human who can log in caps your upside at exactly the moment you deliver the most. You built something that replaces fifty people and you are charging for five seats. On the cost side, seat pricing is blind to the thing that actually costs you money. The user who fires ten thousand requests a day and the one who logs in monthly pay the same, while your inference bill for them differs by four orders of magnitude. The number on the invoice has come loose from both the value and the cost. It is measuring headcount, which is neither.
The market has noticed. Seat-based pricing fell from 21% of SaaS companies to 15% in a single year, while hybrid models jumped from 27% to 41%. That is not a fashion cycle. That is the proxy breaking in public.
Per token: the number that scares
So you align the price to your cost. You charge for what the model actually did: tokens in, tokens out, metered and marked up. Now your revenue finally moves with your COGS, and the power user who was quietly bankrupting you on a flat plan pays their freight.
The trouble is what you have handed the customer. You have handed them a meter they cannot read. Every retry, every extra step of reasoning, every time the model thinks a little harder shows up on their bill, and they have no way to forecast it before the month closes. Buyers do not fear a high price nearly as much as they fear an unpredictable one, because an unpredictable bill is something they have to defend to their own boss. A meter turns every heavy month into an argument.
And it exposes you. Priced per token, you are visibly reselling inference at a markup, which invites the one question you cannot win: why is yours more expensive than the model underneath it. You have entered a price war whose floor is your supplier's price, and you do not own the supplier. This is the same rent I've written about on the engineering side, seen now from the sales side of the desk: when you meter the customer, you have made your cost structure their problem, and your margin their negotiation.
Per outcome: the number that bets the company
The fashionable door, and the honest one, is to charge for the result. Not the seat, not the token, but the resolved ticket, the closed deal, the Tuesday actually done. Gartner expects 40% of enterprise software spend to move to usage, agent, or outcome-based models by 2030, and it is easy to see why everyone wants it. It aligns the price to the customer's value perfectly. You win when they win. Nobody argues with a bill for outcomes they can feel.
But read what you just signed. Outcome pricing bets the company on two things most startups do not yet have. First, your reliability is now your revenue, directly and mercilessly. If the model produces the outcome eighty percent of the time, you have priced in a twenty percent hole, and every reliability bug is a line item bleeding money, not just a support ticket. The scaffolding I keep insisting matters stops being an engineering virtue and becomes the P&L. Second, you have to measure the outcome, cleanly enough to put it on an invoice both sides trust. What counts as a resolved ticket? Who decides the deal closed because of you? You have re-created, in dollars and in front of the customer, the exact problem of defining what "good" means that makes reward functions so hard. A vague outcome is a billing dispute waiting to happen.
The margin underneath all three
There is a floor under this whole conversation, and the SaaS instinct walks right off it. Traditional software had near-zero marginal cost: once the seat was provisioned, the next request was free, which is why those businesses ran at 80 to 90% gross margins and could afford to be careless about which pricing model they picked. An AI product does not get that gift. Inference is a real, variable cost that rises with every request and every feature you ship, and AI-first companies are running at 40 to 70% margins, often less. The whole SaaS circle has become a thin slice.
Which flips a rule founders learned in their bones. In SaaS, growth fixed everything: more customers, more margin, the model got healthier as it got bigger. In AI, growth can make the P&L worse, because your best, most engaged customer is often your worst margin. Replit lived this out loud: revenue rocketed from about $2M to $144M ARR in a year, and they still had to move to usage-based pricing just to drag gross margin out of the single digits into the twenties. They did not have a growth problem. They had a pricing problem wearing a growth problem's clothes.
In SaaS, scale was the cure. In AI, scale is the test, and the wrong price fails it faster the more you sell.
The price is what you're saying
Here is the part that outlives any pricing table. The number does not just capture value. It tells the customer what the thing is. Price per seat and you have said "this is a tool, judge it like a tool, one of many your team logs into." Price per outcome and you have said "this is a worker, hold me to a result." You cannot charge like a tool and be trusted like a worker, or charge like a worker and deliver like a tool. The price sets the expectation you will then be measured against, so picking it is not the end of positioning. It is positioning, stated in the one language the customer cannot pretend not to understand.
Where this lands
So the honest counsel, the kind I would give from a small company's chair rather than a keynote stage, is this. Do not cargo-cult the outcome-pricing headlines before you can deliver the outcome. Outcome pricing is a bet on your own reliability, and if that reliability is not real yet, you are not pricing, you are gambling with a rubric the customer will contest. Most young companies belong at the hybrid door that the market has quietly settled on: a predictable floor so the buyer can sleep, a variable component so your best customer does not sink you, and a slow migration toward outcomes as your reliability earns the right to charge for them.
The one real edge is the boring one. If you control your cost structure, if you run smaller models on your own hardware instead of renting every token off someone else's meter, you have room to price in ways the company across the street cannot, because their floor is a bill they do not control and yours is a machine you own. That is not a pricing trick. It is why the reliability work and the cost work turned out to be the same job, and the pricing page is just where that fact finally shows up in public.
Write the price last and it will tell the truth about a business you built by accident. Write it first and it becomes the plan: what you're selling, who you're selling it to, and which risks you were brave enough to keep on your own side of the invoice.
Kha PhanCo-founder & CTO, Easy AI