A pricing architecture built from scratch for a new AI product.

How do you price a product that has never been sold? Our client was preparing to launch a new AI product. There were no sales figures, no price history and no customers yet whose behavior could be analyzed. All we had was the product, its costs and a market still taking shape. The first price would set expectations that are hard to correct later.

The starting point

Most pricing work starts with data: past deals, discounts, usage, churn. Here there was none. At the same time, a lot was riding on the first price:

• Too low, and the price anchors the market’s expectations; raising it later costs trust and creates friction.
• Too high, and adoption stalls before the product can prove its value.
• The wrong metric, and revenue moves neither with the value delivered nor with the computing costs caused by every use of an AI model.

There were opinions on all three questions, but no evidence yet.

Our approach

Without historical data, structure has to replace history. Following our pricing method, we broke the pricing decision down into explicit hypotheses and tested them before the launch.

From assumptions to testable hypotheses

The starting point was the value the product creates: which customer segments benefit, for which tasks, and how can that value be measured in time saved, costs avoided or additional revenue? We simulated different approaches to segmentation and examined how each user segment perceived the product’s value. From this we derived possible value metrics, packages and price ranges. Each option was recorded as a hypothesis together with the evidence that would confirm or refute it.

Costs as a guardrail

AI products have costs that rise with usage. For every candidate model, we calculated what it costs to serve different user segments. This brought perceived customer value into line with the cost structure and ensured that the model remains commercially viable as usage grows.

A structured test framework

We tested the hypotheses with prospective customers: interviews on value and the buying process, questions on willingness to pay and comparisons of alternative packages and metrics. Every test had a defined question, a target group and a decision rule, so the results could be compared rather than interpreted.

The pricing architecture

The validated hypotheses became the architecture: the value metric, the packages and the fences between them, an entry offer and the path along which customers expand as their usage grows. Deliberately designed levers leave room to evolve prices, package boundaries and usage logic after launch. The system provides direction from day one without anticipating what the market will teach us next.

The results

• A validated pricing roadmap, ready for launch: introductory prices, the next test stages and the decision points at which the model will be reviewed.
• Built 100% without historical data: every decision rests on tested hypotheses instead of past transactions.
• Risk of mispricing minimized: the most consequential assumptions were tested before the market saw a price.

What others can learn from it

A new product does not need a perfect first price. It needs a pricing architecture that connects perceived value, user segments and the cost structure. Deliberate flexibility is equally important, because it allows new market insights to flow back into the system. A greenfield launch is the best moment for this, because there are no existing contracts to migrate yet.