You can't ship AI you can't 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 instead of from the design.

Cost is a build decision, not a finance one. Which model you pick. How much text you send it on each call. How many times you call it per task. Whether a person reviews the output before it goes out. Those four choices set what one use costs, and all four get made at the keyboard by the person building. Hand the question to the CFO after launch and the product has already been built around numbers nobody checked.

Most teams I talk to can tell you what their AI product does.

Nobody measures it because it used to not matter. In ordinary software, one more user costs you close to nothing, so nobody learned to watch the cost of each transaction while designing. AI ended that. Every call to a model has a price, and the price rises with how much you send and how much comes back. More usage means more cost, which turns a good demo into a business that loses money on its heaviest customers.

So the failure looks like this. The product works, adoption climbs, everyone celebrates. Gross margin flattens or slides while the room assumes AI is making things more efficient. It isn't, because the automation was never measured against its own cost. Growth stacks loss on loss, and it stays invisible until someone moves model spend to where it belongs. It's a cost of goods sold, the line that sits directly against revenue. Park it in operating expenses instead and the margin looks better than it is.

The version that works meters cost per action and reads it like a phone bill. A monthly total tells you nothing you can act on. Put a price on each unit of work the system does: one answer, one document processed, one agent run. You can check that number while you build, the same way you check speed or accuracy while you build. It belongs on the same live surface as everything else you steer with, because cost moves in the same hour as usage, and a monthly finance review can't see that.

Two levers move it more than model choice does. The first is how much text you send with each call. That's where most of the waste sits, and it's also where most of the quality comes from, so it's a tradeoff and not a cleanup. The second is who does the checking. A rule that checks the output costs close to nothing. A model checking the output costs cents. A person costs dollars and minutes. Route the cheapest reliable check first and save the person for calls that set a standard. Teams that skip that routing pay human prices for machine work. It never shows up as an AI cost line, because it's payroll.

None of this is glamorous, which is why it's an edge. Plenty of people can build an impressive AI feature now. Far fewer build one that survives contact with its own economics. Measure cost while you build. Price it per action. Make the tradeoffs on purpose. Then you ship things that are still standing a year later, at a margin that pays for the next one.