In short: pick the pricing model that matches how value shows up, know your cost per task, and report the eight numbers that show whether the business works.
Key points
- Seat pricing often undercharges because AI products can do the work of several people.
- Buyers accept usage pricing only with caps, estimates and alerts.
- Cost per task is model cost plus compute plus human review, divided by tasks completed, and must fall.
- Treat a free tier as acquisition cost and never count free users as users in board reports.
- Stop reporting sign-ups, prompts sent and anything counted as AI-generated.
Three models and the fit test
There are three ways to charge for an AI product. Each fits a specific situation, and forcing the wrong one creates friction with buyers and finance teams.
| Model | Fits when | Watch for |
|---|---|---|
| Seats | A named human uses it daily and the value is their time | Usage spread across few heavy users |
| Usage | Value scales with volume and the buyer accepts variable bills | Bill shock, procurement blocking variable spend |
| Outcomes | You can measure the result and the buyer trusts the measurement | Disputes over attribution |
Seats work when the product sits in a person's daily workflow. The buyer understands the model and can budget for it. The weakness is that AI products often do the work of several people. Seat pricing then undercharges.
Usage pricing matches value to volume. It fits products that process documents, tickets, calls or records. Buyers accept it only with caps, estimates and alerts. Without them, finance will not sign.
Outcome pricing is the most attractive and the hardest to run. You need a result you can measure, and a buyer who accepts your measurement. Resolved tickets, qualified leads, completed filings. If either side can argue about the number, the model breaks.
Hybrids are normal
Most AI companies land on a hybrid. A platform fee plus usage. Seats with a usage allowance and overage. A base fee with an outcome bonus.
Hybrids are fine as long as they are simple to explain in one sentence. If your pricing page needs a calculator, buyers will assume the worst case.
The platform fee covers your fixed cost of serving a customer. The variable part tracks inference. That split keeps gross margin predictable even when usage swings.
The gross-margin trap
Every AI pricing model fails in the same way: the company does not know what a task costs to run.
Cost per task is model cost plus compute plus any human review, divided by tasks completed. If you cannot produce that number for last month, you do not know your gross margin. Pricing decisions made without it are guesses.
The number must be falling. Caching repeated context, routing simple tasks to smaller models, shortening prompts, batching requests. Each one is an operating project with an owner and a target. Part 14 covered why investors ask for this. Here it is the difference between a business and a hobby.
Set a margin floor per task and price above it. If a customer's usage pattern drops below the floor, you need to know within the month, not at renewal.
Free tiers are acquisition cost
A free tier is a marketing expense that happens to look like a product. Treat it that way.
Every free task has an inference cost. Add those up by month and you have a number to put next to paid channels. If free-tier conversion to paid is weak, the free tier is an expensive campaign with no stop date.
Do not count free users as users in any report that goes to the board or to investors. Count them as a channel with a cost per acquired team, and compare that cost to the others.
What to report, and what to stop
Part 9 made the case for activation, retention, revenue, efficiency and source over reach. For an AI product those become eight monthly numbers.
- Activation: the share of new teams that complete a first real task within a week. Not a sign-up, not a prompt, a task.
- Weekly active teams: teams, not individuals, because AI products are bought by teams.
- Tasks completed: the unit of value delivered, defined once and not changed.
- Retention at 30 and 90 days: measured on teams that activated.
- Net revenue retention: what last year's customers are paying now.
- Gross margin: net of inference, with the trend.
- Cost per task: the operating number behind the margin.
- Cost per acquired team by source: with the free tier shown as a source.
Stop reporting sign-ups, which measure a landing page. Stop reporting prompts sent, which measure curiosity. Stop reporting anything counted as "AI-generated". None of them tells you whether the product works or whether the business does.
Pricing changes as experiments
Pricing is not set once. It is changed in small steps, each run as an experiment.
Write down the hypothesis, the segment it applies to, the number you expect to move and the stop rule. A stop rule is the condition under which you revert. Activation falls by a set amount, or churn rises, or support tickets about billing double.
Run it on new customers first. Existing customers get grandfathered until the result is in. Then decide, record the decision, and move on to the next one.
The checklist
- [ ] Pricing model chosen against the fit test, explained in one sentence
- [ ] Cost per task known for last month, with a reduction target and an owner
- [ ] Margin floor per task set, with alerts when a customer drops below it
- [ ] Free tier costed and reported as an acquisition channel
- [ ] The eight monthly numbers in one page, sign-ups and prompt counts removed
- [ ] Next pricing experiment written down with a stop rule
Common questions
Should I price my AI product per seat or by usage?
Seats fit when a named human uses it daily and the value is their time. Usage fits when value scales with volume and the buyer accepts variable bills, provided you offer caps, estimates and alerts.
How do you calculate cost per task for an AI product?
Cost per task is model cost plus compute plus any human review, divided by tasks completed. If you cannot produce it for last month, you do not know your gross margin.
What metrics should an AI startup report to its board?
Activation, weekly active teams, tasks completed, retention at 30 and 90 days, net revenue retention, gross margin net of inference, cost per task, and cost per acquired team by source.
Is a free tier worth it for an AI product?
Treat it as a marketing expense. Every free task has an inference cost, so report the free tier as an acquisition channel with a cost per acquired team and compare it to other channels.