The unit economics of AI products

Why this is different from software

Conventional software has a high fixed cost to build and a near-zero cost to serve one more customer. That pattern shaped a generation of pricing and valuation habits. AI products break it. Every request consumes compute, many require retrieval from a database, and some need a person to check the result. The cost of one more unit of work is small but real, and it varies with how the customer uses the product.

Start with the unit of work

Define the thing the customer is actually buying: a processed invoice, a resolved ticket, a drafted contract, a completed review. Then measure the full cost of producing one. A useful breakdown has four parts.

• Inference: the cost of model calls, including retries, long inputs and any second model that checks the first.
• Retrieval and tools: search, databases and external services called along the way.
• Human time: review, exception handling and customer support, measured in minutes per unit.
• Overhead: evaluation, monitoring and security work that scales with usage.

A worked example

Now change one assumption. If the customer's tickets are longer than expected and model cost doubles to $0.40, margin falls to $0.20 per ticket, or 20 percent. If better instructions cut review from one in five to one in ten, review cost falls to $0.15 and margin recovers. The point is not the numbers. It is that two operating levers, inference and review, each move margin by more than a typical price change would.

Where margin quietly leaks

• Retries and long contexts. A product that resends the whole conversation on every turn can cost several times what a lean design costs.
• Heavy users on flat pricing. Per-seat plans reward the customer who uses the system most and penalize the company serving them.
• Premium models used everywhere. Routing easy cases to cheaper models and hard cases to stronger ones often protects quality and cost together.
• Unmeasured review. If nobody tracks review minutes, the cost is hidden in payroll.

Pricing choices

Three common structures each have a trade-off.

Model	Strength	Watch out for
Per seat	Simple, predictable	Cost can outrun revenue for heavy users
Per use	Cost and revenue move together	Customers may resist unpredictable bills
Per outcome	Aligned with value	Requires a clear, auditable definition of success

Many companies blend them: a base fee for access, with usage or outcome pricing above an included amount.

Does scale help?

Often, but not automatically. Fixed costs such as evaluation and security spread over more units. Negotiated model pricing may improve. Better instructions and routing reduce waste. But if growth comes from customers with harder cases, average cost per unit can rise even as volume grows. The honest view is a cohort view: cost per unit by customer type, tracked over time.

Measuring it in practice

Start small. Pick one workflow, instrument it for a month, and compute cost per unit by hand before building a dashboard. Include a sample of review time by timing real cases rather than estimating. Then repeat the calculation for the customer segments that differ most. Many teams discover that a few segments, often the ones with the longest or messiest inputs, explain most of the cost. That finding alone often shapes both pricing and product priorities.

Questions for founders

Do you know your cost per unit today? How does it differ between your easiest and hardest customer? What would you change in the product to lower it without harming quality? Which customers would you decline at the current price, and why? These are ordinary operating questions, and they are the ones that decide whether growth improves the business or stretches it.

A note on forecasts

Be cautious with forecasts that assume costs fall steadily. Usage tends to grow as products improve, and new features often call models more than once. Build the forecast from units of work and cost per unit, and test it with a higher-usage case.

What to ask

When we review an AI business, we ask for cost per unit of work, split into the four parts above, by customer segment and by quarter. We ask how price was set and when it was last tested. We ask what happens to margin if the main model's price doubles, and if usage doubles. And we ask which lever the team would pull first. A team that knows the answer has usually looked at the numbers before. Our guidance for growth-stage companies describes the other questions we ask.

This article is for general information only and is not investment, legal, tax or accounting advice.