Launch log · Pricing
Pricing an AI product when the unit of value moved.
Seats measured how many people could log in. Nothing about that number describes what an agent does. What replaces it, and what breaks when you switch.
Per-seat pricing worked because it measured something real. Software made a person more productive, so charging per person was a rough proxy for value delivered, and everybody understood the arithmetic.
An agent breaks the proxy. If one system handles work that used to occupy six people, seat count now moves in the opposite direction to value. Charge per seat and you get paid less precisely as you succeed, which is an unusual position to design a business around.
Most of the market has worked this out. What follows is less settled, and the go-to-market consequences are larger than the pricing debate usually admits.
What the money people are actually saying
Bessemer Venture Partners published its AI pricing and monetisation playbook in February 2026, and two findings in it deserve more attention than the pricing-model taxonomy that gets quoted from it.
The first is about margins. AI economics differ from software economics because every query has a real compute cost behind it, and Bessemer puts AI-native gross margins in the range of 50 to 60% against 80 to 90% for classic software. Their advice follows from the arithmetic: maths that fails at ten customers will fail at a thousand.
The second is sharper. Bessemer’s phrase is that soft return-on-investment positioning kills willingness to pay, and the example they give is the copilot that offers advice without closing the loop, leaving the customer asking whether they are really getting value. Their warning about timing is the part I would put on a wall: as 2025 pilots reach 2026 renewals, pricing has to reflect actual value rather than a promise.
That renewal wave is happening now, which is why this is a 2026 problem rather than a 2027 one.
The operative principle underneath all of it: price against the labour category you are replacing rather than the software category you are entering. A product automating three quarters of a workflow that used to be billed by headcount should benchmark against the combined cost of software plus human review, rather than against the other platforms in its analyst quadrant.
The models, briefly
Four shapes are in play, and hybrids of them cover most of the market.
Usage pricing charges per token, call or inference. It maps cleanly onto cost and it creates billing anxiety, because the customer cannot forecast the invoice.
Outcome pricing charges when a defined result occurs. Customer support led here, with per-resolution pricing now public at several large vendors, and Zendesk’s own explanation of the model is a fair summary of the logic: the seller takes on part of the performance risk and revenue tracks customer success.
Hybrid combines a base subscription for predictability with a usage or outcome meter on top. Bessemer’s own guidance is that hybrid wins when you are uncertain, which describes almost everyone.
Per-seat survives, mostly where the AI augments a human whose job is unchanged in shape.
The interesting evidence that this is now a structural shift rather than a fashion is that the accountants have arrived. Deloitte published a technology spotlight on accounting for outcome-based pricing in agentic AI products in June 2026. When revenue recognition guidance appears for a pricing model, the model has stopped being an experiment.
The part nobody plans for
Changing the pricing model changes six other things, and teams routinely budget for one of them.
Your value story has to become measurable overnight. Outcome pricing requires both sides to agree what a resolved case is, what counts as a completed task, and who adjudicates a disputed one. That definition is a positioning decision disguised as a contract clause. Get it wrong and you spend every quarterly review arguing about the meter.
Sales compensation stops working. A rep paid on annual contract value has no incentive to sell a model where revenue accrues with consumption. Fix the comp plan in the same quarter as the pricing, or the pricing will quietly not get sold.
Procurement wants a ceiling. Enterprise buyers run on approved budgets, and a bill that varies with usage is a governance problem for them regardless of how fair it is. Commitments with tiered overages are the usual answer, and the honest version is that most enterprise outcome pricing has a subscription hiding inside it.
The business case changes shape. Under seats, the buyer’s case was a cost comparison against the incumbent tool. Under outcomes, it becomes a comparison against labour, which pulls a different executive into the room and raises questions about headcount that your champion may not want to raise yet.
Your own forecasting gets harder. Consumption revenue is less predictable than subscription revenue, which matters if you are raising money or reporting to a board that likes straight lines.
Failure becomes free for the customer. That is the selling point and the exposure. If your system underperforms you earn nothing while still paying for compute, and the margin range above is where that stops being survivable.
How to decide
Three questions, answered honestly.
Can you count the unit without arguing about it? If the outcome needs interpretation, outcome pricing will produce disputes rather than alignment. Support resolutions count cleanly. Improved decision quality does not.
Does your cost per unit fall as volume rises, or stay flat? Flat unit costs plus outcome pricing means you are running a services business with a software valuation, and the market will eventually notice.
What is the buyer replacing? If the answer is a software licence, price near it. If the answer is hours of human work, price against that and accept the longer, more political sale that comes with it.
The launch consequence
Pricing changes are launches, and they fail the same way launches fail. The most common version is a company that announces a new model publicly before finance has signed the floor price, before the comp plan changed, and before anyone wrote down what counts as a successful outcome.
Six weeks later, the first customer disputes a meter reading and there is no agreed definition to point at.
Settle the definition before the announcement. It is a paragraph of work and it prevents a year of arguments.