Seat pricing is ending. Every vendor selling machine work now has to say what it counts instead, and in my corner of the market the buyers have already answered.
The plant manager and the clock
In September I sat on a call with a plant manager who did not know he was being interviewed about pricing. He was telling me about his quality problems, and he told me about them in hours. Hours of rework. Days a line stood still waiting for a signature. Weeks an engineering change sat open because the person who had to approve it was on another site. He never once said "compliance risk." He said time.
A few days later another buyer did the same thing. Then a third. A research company told us it handles roughly a thousand quality problems a year, some of which cost nearly 60 days of downtime. A contract manufacturer said scanning and data entry eat two to three hours every day. A vehicle maker said its main measure of success was cycle time, full stop.
Nobody asked them how they would like to be billed. They told us anyway.
Why seats are over
Two things happened to seat pricing in 2026 and neither is reversible.
The first is that the user stopped being a person. On one large model marketplace, agents began consuming more tokens than humans in February and have not given the lead back. A seat counts people. When the work is done by something that is not a person, the seat is counting the wrong thing.
The second is that the cost underneath the work is in free fall. On a single day in September, one frontier model maker cut its price by a fifth and its rival released two models at half what their predecessors cost. Any price built on inference inherits that collapse. Veeva's chief financial officer said the quiet part out loud: the company will price its agents per business transaction precisely to avoid being marked down to what the underlying model costs.
So every vendor now has to finish the sentence "priced per..." and the interesting question is what goes after it.
Three ways to finish the sentence
Per call. Jason Lemkin found one large incumbent proposing to charge agents $5,000 to $100,000 per million calls, against roughly $83 per million for the same calls from an ordinary integration. Same endpoint, same record, same write, 60 to 1,200 times the price. This unit measures how often the agent knocked on the door, and the price depends on who is knocking. It is not value pricing. It is a toll booth.
Per transaction. A document reviewed, an inquiry resolved and filed. This is Veeva's answer and it is better, because the quality manager recognizes the thing being counted. But it is still the vendor's output. The buyer is left to work out what a reviewed document was worth to them.
Per buyer outcome. A charge for a change in something the buyer already measures. This is where everyone says they want to get to, and the largest agent business in software shows how hard it is. Salesforce's own outcome metric, the Agentic Work Unit, had still not made it into customer-facing pricing in September. They have the volume, the data and the sales force. What they do not have is a unit any particular buyer already uses to describe value.
That is the whole argument in one line. An outcome unit has to be countable from the record without anyone signing a statement about cause, and it has to be a number the buyer already budgets against. The first condition is technical and travels anywhere. The second does not. A support leader counts tickets. A pharma reviewer counts documents cleared. A plant manager counts time.
Which is why outcome pricing will not arrive as one shift across software. It arrives one industry at a time, and in each industry the first vendor to name a unit that is both countable and native gets to set the frame everyone else is priced against. Being first is not a marketing act. To count a unit from the record you have to sit where it happens and watch it happen thousands of times, and Tomasz Tunguz reported a Berkeley study showing that this kind of watching cuts the cost of the same result by 71 percent with no loss of accuracy. The cost advantage and the moat are the same asset.
For quality, the unit is time. Cycle time on a named process, counted from the timestamps the system already keeps, valued by the buyer in the downtime hours they already feel. We have not measured it yet. That is a hypothesis with three plant managers behind it, not a claim.
The case for seats
The best objection is that buyers like predictable bills. In the same month, four prospects told us their budgets sat between $80,000 and $100,000 a year, with anything near the top needing an executive signature. A unit that floats can cross that line without warning. The answer is not to abandon the unit but to post it with a ceiling: a price per unit and a cap that keeps the year inside the gate.
The second objection is that outcome units invite arguments. Downtime has many causes. If I am paid on downtime avoided, every invoice becomes a debate about attribution. Correct, and that is exactly why the unit is cycle time on a named process rather than downtime itself. Cycle time is counted. Downtime is what the buyer uses to decide what a day of it is worth. If a unit cannot be read from the record without a causation argument, it is not ready to carry a price.
The third objection is that the per-call meters are not pricing at all but rent on data the buyer already paid to create. I think that is right, and it is the opening. Six mid-market prospects in September were replacing systems they had held for years after incumbent price rises or forced migrations, comparing six to fifteen vendors each. Rent makes people ask what they are paying for. A vendor with a unit native to the buyer's work has an answer ready.
How I lose this bet
Two dates, and I will keep score here.
If by 30 June 2027 no quality or life sciences incumbent has published an agent-access meter at more than ten times its integration rate, I was wrong about which unit they reach for first, and the toll-booth story was unfair to them.
If the first customer we measure before and after on one process tells us their finance team will not put a value on cycle time and would rather keep buying seats, I was wrong about the unit. That is the one I would hate to lose, so it is the one I am going to test first.
And if by the end of 2028 outcome pricing in two or more regulated industries has settled on a single horizontal unit from a cross-industry vendor, an Agentic Work Unit or a token charge, then the whole "one industry at a time" thesis was wrong and this post becomes a cautionary tale. I will leave it up.
