When the agent does the work, what do you charge for?

Per-seat pricing has a hidden assumption baked so deep that most of us never say it out loud: value scales with the number of people using the thing. Ten seats, ten people getting work done, ten units of value — charge accordingly. It was never really about seats. Seats were a proxy, and for a long time an honest one, because the work was done by humans and each human needed a login to do it.

Agentic features quietly snap that link. When an agent drafts the reply, resolves the ticket, opens the pull request, the work happens whether or not a person is sitting in a seat watching it. You can now have fewer seats and more value flowing through the product — or the same seats and ten times the output. The proxy has come loose from the thing it was standing in for. That’s the whole disruption, and it’s less about AI than about a measurement suddenly measuring the wrong quantity.

Here’s the useful way to hold it: your pricing model is a theory of where your value comes from. Seat pricing is the theory “value comes from humans using the tool.” The moment that stops being true, your price is pinned to the wrong variable, and you’ll feel it as a slow ache in one of two directions — either margins bleed (heavy agent usage, flat per-seat price, your costs climbing under a price that can’t move) or adoption stalls (you meter so tightly that people ration the very thing that creates the value).

That leaves three honest options, and none of them is free.

Keep charging for seats. Still right when the human is genuinely the unit of value — when the agent makes each person dramatically better but a person is still the one deciding, judging, steering. You’re not selling output; you’re selling a more capable operator. The risk is you leave money on the table exactly as the agent gets good enough to work unattended.

Charge for usage — runs, actions, tokens, tasks. Easy to meter, and it moves with load so your margins stop bleeding. But usage is a proxy for your costs, not the customer’s value, and metering has a nasty side effect: it makes your best, highest-value power user watch a meter and hesitate. You end up taxing the behavior you most want.

Charge for outcomes — tickets resolved, PRs merged, dollars collected, meetings booked. This is the one that actually tracks value, which is why it’s tempting and why it’s hard. You have to define the outcome so tightly that both sides agree it happened, attribute it fairly when the human and the agent share the credit, and absorb the cases where the outcome is real but messy. Get it right and price finally means what the customer feels they’re buying. Get it wrong and you’re arguing about definitions every invoice.

The mistake I’d watch for is the non-decision: bolting an agent onto last year’s per-seat plan and hoping the numbers still work. They quietly won’t. Not because per-seat is dead — it isn’t — but because the question underneath it changed from how many people to how much got done, and a price that can’t answer the new question slowly stops describing your product.

You don’t have to pick the perfect model. You have to pick one whose hidden assumption is still true for your product this year — and be willing to notice the day it stops being true.

Worth reading around this: value-based pricing vs. cost-plus; the “jobs to be done” framing (what outcome is the customer actually hiring you for). The seat → usage/outcome shift is a widely-observed 2026 trend in software pricing, not a settled rule — treat it as a lens, not a law.