Samyak Jain
Pricing

Four cents a copy

When software does the work itself, charging per seat stops making sense. What replaces it depends on a question from contract theory, namely who should carry the risk when the work goes wrong.

Samyak JainNovember 9, 20268 min read

In 1959 a small company from Rochester, New York, showed the world the first office copier that used plain paper. The Xerox 914 weighed about 650 pounds and made seven copies a minute. It was also very expensive to build. The purchase price was $27,500, more than most offices would consider for a machine whose usefulness they could not yet judge.

So the company did something unusual. It leased the machine for $95 a month, with the first 2,000 copies included and each copy after that priced at four cents. A meter on the machine kept count.

Customers turned out to make far more copies than anyone had forecast, because nobody had known how many copies they wanted until copying was easy. By the mid-1960s the 914 is estimated to have produced about two-thirds of the company’s revenue. The copying technology was the invention. Charging by the copy is what made it a business.

Software is going through the same change now, for a similar reason.

Why seats break

For twenty-five years most business software has been sold by the seat. A seat is a person’s right to use a tool. It is a sensible unit when the person does the work and the software helps.

It stops being sensible when the software does the work. A support agent that resolves half of all tickets means the customer needs fewer support staff, and so fewer seats. The better the product performs, the less its maker earns. A pricing model that punishes the seller for succeeding will not last.

There is a second problem on the cost side. Traditional software costs almost nothing to run for one more customer, which is why software companies have enjoyed gross margins of 70 to 80%. An AI product pays for model usage every time it does anything.

AI products earn thinner margins than software used to

Average gross margin of AI products at surveyed companies (ICONIQ, State of AI reports)

0%25%50%75%100%Typical softwarebusiness, 70 to 80%41%202445%202553%2026projected59%2027projected
Figure 1. Survey averages as reported by ICONIQ; 2026 and 2027 are the respondents’ projections. In the same survey, model inference took about 23% of revenue at scaling-stage companies.

A common reply is that model prices are collapsing, so the problem will fix itself. Prices are falling, by about ten times a year for a fixed level of capability. But products rarely stay at a fixed level. They move to the newest model, and they move from single answers to agents that plan, call tools, check their work and try again, using many times more tokens for each task. The price of a token falls while the number of tokens in a task rises.

The economist William Stanley Jevons described this pattern in 1865. More efficient steam engines did not reduce Britain’s use of coal. They made coal useful for more things, and consumption grew. Cheaper tokens have so far had the same effect.

Four units

If a seat is the wrong unit, there are three others to choose from. Together they form a ladder.

Four things a customer can pay for

Seat

You pay forA person's access to the toolExampleMost software sold since 2000Risk sits withThe buyer, who pays even if nobody logs in

Usage

You pay forTokens, credits or actions consumedExampleModel APIs; Agentforce credits at about 10 cents an actionRisk sits withThe buyer, who pays for failed attempts too

Task

You pay forEach job the agent takes onExampleAgentforce at a list price of $2 a conversationRisk sits withBoth: the seller absorbs cost overruns, the buyer absorbs failures

Outcome

You pay forEach job doneExampleFin at 99 cents a resolutionRisk sits withThe seller, who is unpaid for every failure
Risk moves from buyer to seller
Figure 2. Examples are list prices as publicly reported in 2026 and will change. The direction of the ladder is the point.

Each step up the ladder moves the price closer to what the customer actually wants, which is finished work. Each step also moves risk from the buyer to the seller. Under seats, the buyer pays whether or not anyone uses the product. Under outcome pricing, the seller pays for every attempt and is paid only for the ones that succeed.

Buyers like the top of the ladder, for obvious reasons. Whether a seller should climb it is a harder question, and economics has a good answer.

Who should carry the risk

In 1979 the economist Bengt Holmström published a paper called Moral Hazard and Observability, part of the work for which he later shared a Nobel prize. It asks when one party should be paid according to results. His answer, known as the informativeness principle, is roughly this: tie pay to a measure only to the extent that the measure tells you about the performance of the party being paid. Paying someone for results they do not control just hands them risk, and they will charge you for carrying it.

Turned into questions for a pricing decision, that gives three tests.

Can both sides agree that the work was done? A copy is easy to count. A “resolution” is not. If a customer asks a question, gets an answer and leaves without replying, was the issue resolved? Vendors answer this differently, and the definition can move a bill by a large amount. If you cannot write a definition that a sceptical buyer would accept, you are not ready to price on outcomes.

Is the result mostly your product’s doing? A sales agent that drafts emails contributes to a closed deal. So do the salesperson, the product and the economy. Pricing on closed deals would make your revenue depend mostly on things you do not control. A resolved support ticket is much more clearly the agent’s own work.

Do you control the variation? Under outcome pricing your margin depends on how often the agent succeeds and what each attempt costs. If you know both numbers and they are stable, you can carry the risk cheaply. If your success rate swings from one customer to the next, you are making a bet on each contract.

When the answer to any of the three is no, the right move is one step down the ladder.

The margin on a finished task

The third test can be made precise. Suppose you charge 99 cents for each resolved conversation and nothing for the rest. Your revenue per attempt is 99 cents times your resolution rate. Your cost per attempt is whatever the model usage comes to, paid whether or not the attempt works.

Under outcome pricing, margin depends on reliability

Gross margin at 99 cents per resolved conversation, for two levels of model cost. Illustrative figures.

-25%0%25%50%75%100%Break-even20%30%40%50%60%70%80%90%Share of attempts that end in a paid resolution89%10 cents an attempt72%25 cents an attempt
Figure 3. Gross margin here is one minus cost per attempt divided by revenue per attempt. Human fallback, support and hosting would lower every point on both lines.

At ten cents an attempt and a 60% resolution rate, the margin is 83%, as good as traditional software. At 25 cents an attempt and a 30% rate, it is 16%. Below a 25% rate, that second product loses money on every conversation it handles.

The product, the price and the customer are the same in both cases. The two things that differ are the cost of an attempt and how reliably the agent finishes the job. That is why I think of pricing and reliability as one subject. A company that prices on outcomes is selling its own reliability, and it should know that number before it signs.

Choosing a rung

A few working rules follow.

Start one rung below where you want to end up. Price by task or usage while you learn your own success rate across different customers. Measure outcomes from the first day, even though you are not yet charging for them.

Move up when the numbers are boring. When resolution rates are stable across customers and the cost of an attempt is predictable, outcome pricing becomes an advantage, because competitors with shakier products cannot afford to match it.

Use a hybrid to cover fixed costs. A platform fee plus a price per outcome protects the seller against low volume and still ties most of the bill to results. The 914’s lease, with its monthly fee and included copies, was the same arrangement.

Treat the definition of the outcome as product work. Whatever you get paid for is what the whole company will improve. If a resolution counts when the customer goes quiet, you will build a product that makes customers go quiet. Write the definition as carefully as you would write a specification, because it is one.

Xerox could charge by the copy because a copy is easy to recognise. Everyone agrees on what one is, and a meter can count them. Most of the difficulty in pricing AI work comes from agreeing on what counts as a copy.

Sources