The cost of running an AI model is easy to measure. It shows up in cloud bills, API invoices, and unit economics spreadsheets, and that visibility makes it a natural anchor for pricing decisions. It shouldn’t be.
The Floor and the Ceiling
For any AI product, you’re working between two numbers.
The floor is your full cost to deliver: inference, infrastructure, development, integration, and the ongoing operational work of monitoring, evaluation, and human review that most AI systems require. Price below it and you’re losing money on every transaction. Build this number carefully. Unlike the ceiling, which is determined by your customer’s economics, the floor comes entirely from decisions within your organization.
The ceiling is the maximum your customer would pay before the economics stop making sense for them. It depends on what your product does for them, and the calculation differs depending on which situation you’re in.
Three Situations, Three Ceiling Calculations
When you’re replacing a process
If your AI product replaces something a person or team currently does, the ceiling starts with the fully loaded cost of that process. Take customer support. If your product replaces a support team, the ceiling starts with what that team costs: agent wages, benefits, software, management overhead, and facilities. Your customer knows that number.
If your product is also meaningfully better than what it replaces, that quality improvement extends the ceiling beyond the cost baseline. An AI system that resolves issues faster, more consistently, or with fewer escalations creates value beyond cost displacement alone.
But if your cost per support resolution exceeds what a human agent costs and your product doesn’t deliver meaningfully better outcomes, you don’t have a pricing problem. You have a product problem.
When you’re augmenting a process
Augmenting is different from replacing. The person stays in the role, but AI reduces the time or effort the work requires. The ceiling here is the calculable value of that improvement: the hourly cost of the person, the time saved per interaction, and the number of interactions affected. For a support operation handling 50,000 contacts a month, shaving two minutes off average handle time has a dollar value.
When you’re creating something new
The harder situation is when your AI product enables something the customer has chosen not to do because it was not possible, or because the cost was prohibitive. A company without off-hours support or multilingual capability didn’t necessarily lack the option. They made a cost decision. Your product changes that calculation, and the ceiling isn’t set by what they currently spend. It’s set by the business value of the capability they’ve been going without.
Here, the ceiling is bounded by what that capability is actually worth to them: what is 24/7 multilingual coverage worth in reduced churn, in newly accessible markets, in improved customer retention?
Most AI products don’t fit cleanly into one category, so treat each component separately, calculate the ceiling for each, and combine them.
The Floor Is Moving
All three ceiling calculations assume a relatively stable floor. For AI products, that assumption needs qualification.
Epoch AI’s analysis of inference price trends found that the cost to reach a given level of model performance has fallen dramatically, with the fastest declines concentrated in the most recent period.
What this means for the investment case: if your unit economics are marginal today, they may not be in a year. A cost projection built on current inference prices understates your likely margins. Build a cost curve, not a point estimate, and use it explicitly in your business case.
That cost curve also opens a strategic option worth naming. Some teams make a deliberate choice to price below their current cost, betting that falling inference costs will close the gap before the runway runs out. This bet is more defensible for AI products than it would be in most other categories.
The starting point for any AI pricing decision isn’t your cloud bill. It’s your customer’s economics and which situation you’re in. Everything else follows from that.
