Pricing AI products by token is a trap.

Bhai, ekta kotha shune ja. Thirty years on Block 3 of the Karwan Bazar fish market, calling auctions before the city wakes up, and I have watched the unit on this chalkboard get smudged out and rewritten more times than the Ramzan moon. Per fish, per bucket, per kilo, per maund, per tray of eight, per quality grade. Each one held for a season, and each one broke for the same reason. The thing we were measuring stopped being the thing the buyer thought he was paying for, and the moment the trader at the back noticed, the price collapsed in a morning. Saquif bhai sat on this stool one September asking me about it, because his SaaS friends in Gulshan are losing their shirts over the same problem with their AI products.

The first proper unit I worked under was per fish. Each hilsa at its own called price, the buyer pointing, the runner pulling it out with both hands like a baby. Beautiful for the market, murderous for the caller. By eight my throat was sandpaper, and the buyers had figured out that whichever hilsa placed second always went for thirty taka less than the first.

So we moved to per maund, which is forty kilos at one shouted price, and it held for almost a decade. A maund is a maund, the colonial scale on the back wall measured it in plain brass. What killed it was the inventory. By 2008 we were getting half trays from Chandpur and quarter trays from Barisal in the same morning, with size and quality so mixed that the wholesalers stopped trusting the average.

Now the SaaS bhais sitting in Mohakhali office towers with their AI agents are about to rediscover this, on faster timescales because their customers can churn in an afternoon instead of a season. The standard playbook for software, going back to early Salesforce and the Tomasz Tunguz reading list on pricing based on usage, was per seat. One named user, one monthly fee, predictable for both procurement and revenue. It worked because the marginal cost of one more user was roughly zero.

AI broke per seat the way the half trays from Barisal broke per maund. The thing inside the unit stopped being uniform. One designer at a global hospitality brand puts the image agent on a rendering loop for three days, generates a thousand variants of a banner, and his single seat costs the vendor four hundred and twelve dollars in inference for the month. The seat next to him belongs to a finance analyst who logs in twice and forgets the tool exists. Both seats bill at thirty dollars.

The Bessemer team's annual State of the Cloud report says it in more diplomatic language. Vendors built natively for AI who held per seat pricing into 2025 watched gross margins fall from the standard low eighties down into the low fifties, while legacy per seat SaaS held intact. The unit had stopped being uniform inside the seat the way the fish stopped being uniform inside the maund, and nobody had built a way to mark up the heavy ones and discount the light ones inside one contract. Saquif bhai chased that per seat thread all the way down on his own a few weeks later, once the agents stopped being the heavy user on one seat and started being the reason the seats emptied out.

So the second wave moved to per token. The deck pitch is tidy. The customer pays for what the agent computes, heavy users pay heavy, light users pay light, vendor margin per request is a known constant. The wholesalers in our market have offered me variants of this for years. Per kilo, per gram of edible meat, per fillet. The trouble is always the same. The unit you priced last month is not the unit you're selling this month, and the buyer notices on his second invoice.

This is the precise moment I lived through in 2008. The wholesaler asks why the maund weighs different at six in the morning than at six in the evening, and the honest answer is that the ice is melting. The honest answer in the AI case is that the model is reasoning longer, which is the vendor's choice and not the customer's. Per token pricing gives the vendor a unilateral knob on the bill that the customer cannot see and did not consent to, and the relationship lasts only until the CFO notices the line item.

The deeper problem with per token pricing, the one that takes the buyer six months to articulate, is that he was never trying to buy tokens. He was trying to buy the resolved ticket, the booked room, the recovered chargeback. The token is a unit of how the vendor produces the thing, not what the customer is consuming. It is the equivalent of me selling fish by the gram of saliva the buyer expends shouting his bid.

The third model, the one actually working in pockets right now, is per outcome. The customer pays when the agent closes the ticket, books the room, completes the chargeback recovery. Vendor revenue is mechanically aligned with customer value, because there is no revenue without customer value. Margin pressure turns into a real engineering incentive to make the agent better, because every failed attempt is unpaid labour. The CRO at the vendor side absolutely hates it.

The reason per outcome is hard, and the reason most teams retreat to per token pricing after the second failed pilot, is the definition of done. When does a support ticket count as resolved? When the customer clicks "thank you," or when no ticket comes back as a follow up within thirty days, or when the human agent never had to be paged? Each definition gives you a different bill. Aaron Ross and Jason Lemkin's From Impossible to Inevitable made the point about contracts based on outcomes in its chapter on enterprise sales fifteen years ago, so this is not a new problem. What is new is that the agent is doing the work the human used to do, so the vendor and the buyer are now arguing about a definition of done that used to be silently negotiated inside the human's head.

The teams I am watching in 2026 are doing one of three things. Some are holding per seat pricing with a surcharge for heavy users, a fig leaf that will not survive a second renewal. Some are doing per token pricing and absorbing the drift in their margin, sustainable until the model provider raises rates by twenty percent in a quiet email. The handful quietly outperforming have moved to per outcome pricing with a narrow definition, a base fee, and a clause that lets them redefine the outcome each quarter. That last clause is the one the buyer's legal team fights hardest, because it is the clause that lets them rewrite the chalk, and the buyer remembers in his procurement scar tissue what happens to a unit when the seller can change it without consent.

The chalk on this board has been smudged out and rewritten six times in my thirty years. The current price is on it now in white chalk under the bare bulb at Block 3, and by eleven o'clock when the wholesalers from Chandpur arrive with the next tide it will already be smudged out and rewritten once more. The unit you bill in is not a contract you sign in March and forget about. It is a public chalk mark the whole market reads each morning, while some caller down the row is figuring out what to write on his own board to take your buyer away from you. Saquif bhai pays his cha bill on bKash before he walks back to Banani, and we both know the price he paid today will be a different price next Tuesday.