Business Models for AI Products: Seats, Usage and Outcomes

The margin curve of each pricing shape is arithmetic and it is written out elsewhere. What is not is the behaviour each shape produces — in your product team, in your customers, and in which customers you attract at all. That is usually what decides whether a model survives its second year.

Two questions people merge into one

“What is the business model” is two separate questions wearing one name.

• The revenue model: what triggers an invoice. Seats, metered units, credits, outcomes, a licence.
• The value-capture position: what you own that makes anyone pay at all. Distribution, a workflow the work lives inside, data that accrues, a compliance posture, a relationship.

They are independent. A company with a strong position can charge in almost any shape and be fine; a company with no position cannot fix it with a clever meter. If you are choosing a revenue model to solve a value-capture problem, you are solving the wrong one, and the tests a claimed position has to pass are the place to start instead.

What is new with inference in the stack is that the revenue model now has to absorb a real variable cost, and the shapes differ enormously in how they do it. The margin curves — how each shape behaves as usage grows — are derived in the pricing-shapes comparison, and this page does not repeat them.

The four shapes and what each rewards

Per seat

Revenue is flat, cost rises with use. The incentive this creates internally is the problem: your finance function now benefits from customers using the product less, at exactly the moment your product function is trying to make them use it more. Nobody states that out loud, and it surfaces as an unwillingness to ship the feature that would drive engagement, or as quietly lowered limits. Seat pricing is right where usage per seat is bounded by something physical — a person’s working day, a fixed number of cases — and wrong the moment a seat can be automated.

Per unit of usage

Revenue tracks cost, so margin is constant by construction. The incentive is subtler and worse than it first appears: if you charge per token, you are paid more for being verbose and paid less for the efficiency work that would improve the product. Charging per token also exports an implementation detail, so every model change becomes a price change the customer notices. Charge for a unit the customer recognises — a document, a query, a resolved ticket — and absorb the token variance yourself, which puts the incentive back where it belongs.

Credits

A currency you have issued. It buys legibility and prepayment, and it fixes an exchange rate between your cost and your price at the moment you publish it. If the work behind one credit gets more expensive — a verification pass, a longer context, a better but pricier model — the margin moves and the price cannot without a customer-visible change. Credits also create an accounting question about unused balances that belongs to a finance person before it belongs to an engineer.

Per outcome

The most attractive and the most demanding. You are paid for a resolved ticket, a completed reconciliation, a booked appointment. The alignment is genuine: the buyer is buying the thing they actually want, and your incentive is to produce it as cheaply as possible, which is also their interest. Three conditions have to hold, though, and they are strict:

• The outcome must be observable and attributable without argument. If a human touched it, both parties can claim it, and you have a dispute per invoice.
• You must be able to refuse work. Otherwise you are obliged to attempt cases whose expected cost exceeds their price.
• Cost per attempt must be bounded. An agent that retries indefinitely on a hard case turns a fixed price into an unbounded cost, so per-outcome pricing needs a hard budget per task in the runtime, not a policy in a document.

Adverse selection, especially in outcomes

Every pricing shape sorts customers, and the sort is not random. This is the part that is invisible in a margin curve, because the curve assumes the customer mix is exogenous. It is not.

Flat price P for a task whose true cost to serve
varies across buyers: c_i

A buyer buys if P < their alternative.
Buyers with high c_i are exactly the ones whose
alternative is most expensive — so they buy first.

Result: the buyers who arrive under a flat price
are drawn disproportionately from the top of the
cost distribution, and the average cost of your
customer base is worse than the average cost of
the market.

That is the classic insurance result and it applies without modification. A flat price for a task with variable difficulty attracts the difficult cases. A per-outcome price for “resolve this ticket” attracts the tickets nobody else could resolve. A generous flat plan attracts the heaviest users of the category, who knew they were heavy users before they signed up.

The defences are not exotic: price on an observable that correlates with difficulty, cap the work per unit, keep the right to decline, and measure the cost distribution of arrivals separately from the cost distribution of the whole base, because the first is a leading indicator of the second.

The working capital nobody models

Classic software has almost no cash gap: the cost of serving a customer this month is trivial, so revenue collected in arrears is fine. Once cost of goods is real and incurred before the invoice, a growing company finances its own growth.

cash gap = COGS_monthly * (days from cost incurred
                            to cash collected) / 30

Assume COGS of $60,000 a month and net-45 terms
with a further 15 days of billing lag:

  cash gap = 60,000 * 60 / 30 = $120,000 tied up

And it scales with growth: double the volume and
the gap doubles, permanently, even at a healthy
margin.

Two shapes make this materially worse. Per-outcome pricing means you pay for every attempt including the failures and invoice only for the successes. Annual invoicing in arrears means you fund a year. Two shapes make it better: prepaid credits and any plan billed in advance are negative working capital, which is one of the genuinely underrated arguments for them and has nothing to do with margin.

Margin enters every acquisition decision

The last reason the model matters is that gross margin is a multiplier on everything downstream of it. Acquisition is paid for out of gross profit, not out of revenue:

months to recover acquisition cost = CAC / (ARPU * m)

Assume CAC = $600, ARPU = $50 per month:

  m = 0.85  ->  600 / 42.50 = 14.1 months
  m = 0.60  ->  600 / 30.00 = 20.0 months
  m = 0.45  ->  600 / 22.50 = 26.7 months

A twenty-five point difference in margin is nearly six extra months of payback on identical revenue — which is six extra months of capital, at whatever capital costs. This is the mechanism by which a cost-of-goods problem becomes a fundraising problem, and it is worked through in what a variable COGS does to a software business.