The cost of an AI feature is set while you build it

The product works and adoption climbs. Gross margin slides, and the room assumes AI is making things efficient. Then the team finds out from the bill that the product has become too expensive to sell.

Most teams I talk to can explain what their AI product does. When I ask what one use of it costs to run, fewer of them have the number.

The cost of a single use comes down to 4 choices. They're which model you pick, how much text you send it per call, how many calls each task makes, and whether a person reviews the output. All of them get made at the keyboard. So cost is a build decision, and the people building are the ones making it.

Model spend is a cost of goods sold, the line that sits against revenue. Park it in operating expenses and the margin looks better than it is. Booked against revenue, the slide shows up in gross margin.

The version that works puts a price on each unit of work the system does: one answer, one document processed, one agent run. A monthly total gives you nothing to act on. Once each action has a price, you can check it while you build, the same way you check speed or accuracy.

There are 2 levers that shift the price more than model choice does. The first is how much text you send with each call. Most of the waste sits there, and so does most of the quality, which makes trimming it a quality decision as well as a cost one.

The second lever is who does the checking. A rule costs close to nothing, a model costs cents, and a person costs dollars and minutes. Route the cheapest reliable check first, and save the person for the calls that set a standard.

None of this is glamorous, which is why it's an edge. Far fewer teams build an AI feature that survives its own economics than build an impressive one. Make each tradeoff on purpose, and check the price of each action during the build, next to speed and accuracy.