You cannot ship AI you cannot afford

Most teams I talk to can tell you what their AI product does. Fewer can tell me what one use of it costs to run. That gap is where AI products become too expensive to sell, and the team finds out from the bill.

Cost is a build decision, not a finance one. Which model you pick. How much text you send it per call. How many calls per task. Whether a person reviews the output. Those four choices set what one use costs, and all four get made at the keyboard.

The failure looks like this: the product works, adoption climbs, and gross margin slides while the room assumes AI is making things efficient. 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.

The claim

Park it in operating expenses and the margin looks better than it is.

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 tells you nothing you can act on. A per-action price you can check while you build, the same way you check speed or accuracy.

Two levers move it more than model choice. The first is how much text you send with each call, where most of the waste and most of the quality both sit. The second is who does the checking: a rule costs close to nothing, a model costs cents, a person costs dollars and minutes. Route the cheapest reliable check first and save the person for calls that set a standard.

None of this is glamorous, which is why it is an edge. Plenty of people can build an impressive AI feature. Far fewer build one that survives its own economics. Measure while you build, price per action, make the tradeoffs on purpose.