You can't sell SaaS seats anymore.

Right now, on some Mohakhali office floor, a finance team is being sold an AI agent at thirty dollars a month for its seat. The bot goes on the invoice next to the humans, as if it badges in at nine and takes its tea break at four. It's a fig leaf, and it won't survive the second renewal. One agent seat runs at a hundred times the hours a human seat does, and the morning the buyer works out that he can route the whole department's volume through one licensed agent, the per seat line stops describing anything that's really happening on the floor.

I know how that ends, because I watched a different unit dissolve the same way. There's a printing press on the second floor of a building off Bangla Bazar Road, above a paper wholesaler who keeps the reams stacked until the whole staircase smells of damp newsprint. Mintu mama has run it since before I was editing anything worth printing, and I was there last week to reprint a slim book of my chacha's poems. The composing room at the back used to hold eight men picking lead type with a speed that looked like a card trick. There is one boy in it now, Shohag, at a screen, and he does in an afternoon what the eight used to do in a week.

Mintu mama billed the publishers by the forma, which is sixteen pages, the unit every Bangla book is still quoted in. So many taka per forma composed, the number sitting on the job sheet before the publisher walked in. I asked him what he charges per forma now and he laughed at me. "Forma diye ki hishab korbo, bhai, lok-i to nai," he said. The composing is the boy and the screen, and the screen doesn't care whether it's one forma or twenty, so the unit he spent thirty years pricing has nothing left to measure. This is the cousin of a problem I wrote about last month, the Karwan Bazar fish caller whose chalkboard price kept getting smudged out and rewritten. Mintu mama's case is worse, because his unit didn't change into a different unit, it just dissolved under him.

Per seat software held for fifteen years for the same reason the forma held: one more user cost the vendor almost nothing, and every seat was a person doing the actual work. The agent breaks it the way the screen broke the composing room. When Klarna said in early 2024 that the assistant it built on OpenAI's models was doing the work of seven hundred service agents in its first month, the part nobody had priced was that a buyer running that doesn't need seven hundred seats. He needs one person to watch it and a contract for the volume. Six hundred and ninety nine seats walk off the invoice, and the better the agent gets, the faster the rest follow.

The honest objection to all this has James Bessen's name on it. He studies this kind of thing at Boston University, and he likes to point at the automated teller machine. The ATM was supposed to wipe out the bank teller. Instead, across the decades the machines spread, the number of tellers in the United States went up. A branch got cheaper to run, so the banks opened more of them, and the tellers moved off counting cash and onto the relationship work the machine couldn't do. Maybe agents go the same way. Maybe making knowledge work this cheap just means there's far more of it to do, and the seats don't vanish so much as change shape into something that supervises the agents instead of competing with them. Jason Lemkin of SaaStr replaced most of his sales team with around twenty agents, and his Salesforce bill went up roughly eighty percent rather than down, because the agents he kept running were hammering everything else the contract metered. The unit moved, but the bill did not fall the way the seat count did. I half believe the optimistic case. Even in that version, the seat stopped being the thing the buyer was paying for a while ago. It's just taking the invoice a little longer to admit it.

The louder objection comes from the companies where per seat is working beautifully. Cursor sells its coding assistant mostly by the seat, its users say they prefer it, and per seat is quietly underwriting most of the vibe coding boom, a boom worth billions of dollars on its own. Every Cursor seat is still a human the model makes faster, though, not an agent working alone, which is the one case where per seat holds. When Cursor tried in mid-2025 to meter its heaviest users by tokens, the developers revolted. Nobody thought token pricing was unfair. They just couldn't tell what it would cost them next month. Plenty of vendors meter by tokens anyway, since the agent's compute is the cost that actually moves, and a few large platforms now bill even for their own internal agents that way. Cursor's developers already showed where that goes. You cannot forecast a token bill, and a nervous CFO cuts what he cannot forecast first.

Tunguz's answer to that is the three part tariff, the shape your phone bill already uses: a base platform fee, an included allotment of usage, and a metered rate on whatever runs past it. It hands the buyer a floor he can budget against and the vendor a meter that tracks the agent's real cost, and it is the likeliest place token pricing settles once the novelty wears off.

So what do you charge for, once you've accepted that you're renting out a worker and not licensing a chair? The three part tariff is the safe answer, but it still meters the tokens the agent burns, which is the vendor's cost and not the buyer's value. The shape a smaller and stubborner group is chasing meters the outcome instead. Salesforce put Agentforce out at roughly two dollars a resolved conversation. Sierra, the company Bret Taylor started after he left Salesforce, charges a rate negotiated in advance when its agent resolves a case and nothing when it has to escalate to a person. As Taylor puts it, the atomic unit of productivity in AI is a process, not a person.

None of this is settled, and the people closest to it say so first. Everyone agrees the seat is wobbling and that outcomes is roughly the direction. But an outcome is a word somebody has to write into a contract, and a resolved ticket means one thing to the vendor's revenue team and another to the buyer's procurement team. It's the same argument the Karwan Bazar fish caller has every morning with his wholesalers about what counts as a finished maund, except this version has lawyers in the room and an agent clearing tickets in the background while they fight. The seat was easy because you could point at it. Nobody has worked out how to point at a finished ticket, and until they do, the contract gets reopened every renewal.