We Lost Fifty Dollars A Month On Every Customer We Signed

Here is a number I do not enjoy saying out loud: more than fifty dollars a month, lost on every client we signed.

Not a rounding error. Not a bad quarter. A structural loss built into the price, which meant every sale made it worse. We were growing and bleeding simultaneously, and growth was the problem.

What actually happened

I priced the product against what a light user consumed. That was the whole mistake, stated plainly. I ran the numbers on realistic-looking usage, added margin, and shipped.

Real customers are not light users. People who find a tool genuinely useful use it a lot, which is the entire point, and consumption-based costs scale with exactly that enthusiasm. Software pricing trained all of us to think in seats: one customer, one price, marginal cost near zero. AI pricing does not work that way. Every successful customer costs more to serve than the last one you modeled.

No vendor lied to me. Nobody buried a term. I did the arithmetic against optimistic inputs and the market corrected me.

How we fixed it

The instinct is to raise the price. We tried that first and it was the wrong first move, because it treated an engineering problem as a commercial one.

What actually worked was re-architecting.

Cheaper models for the easy work. A large fraction of what we were sending to an expensive model did not need it. Classification, routing, extraction from clean documents: smaller models handle those well.

The expensive model only where it changed the answer. We went function by function and asked whether the premium model produced a materially better result. Sometimes yes. Often no.

Aggressive caching. A surprising proportion of requests were near-duplicates.

Cutting what never needed to go. We were sending context on every call that mattered on one call in fifty.

That work took real engineering skill and, honestly, some craft. Getting consumption down without the product getting worse is a genuine discipline, and it is invisible in every pitch deck you will ever see.

What this means if you are buying

Ask for the pricing curve, not the entry price. Model the volume you expect in year two. Ask what the vendor does to manage consumption, and listen for whether they have an answer or an assurance.

And if a vendor tells you their margins are fine at any volume, ask them how. The honest ones will tell you about caching and model routing. That answer is a good sign, not a red flag.