AI Cost Per Customer: The Question Our Board Asked (And We Couldn't Answer)
Our board meeting last year: "What's our AI cost per customer?" We had total spend. We had usage by model. We did not have cost per customer. We mumbled something about averages and moved on. It was embarrassing. We're a data company.
Why it matters more than total spend
Total spend tells you if you have a problem. Cost per customer tells you what to do about it. A $30k AI bill could mean 500 customers at $60 each (fine) or 100 customers at $300 each (not fine). Or 2,000 free users eating 60% of it (definitely not fine). We had no idea which world we were in.
The board wasn't asking for vanity metrics. They wanted to know if our pricing was sustainable. If we were losing money on certain segments. If we had a free-tier problem. We couldn't answer any of that. Our finance team had revenue by customer from Stripe. Our eng team had usage by model from OpenAI. Nobody had connected the two.
What we discovered
Once we could actually attribute spend to Stripe customers, the picture was ugly. Our free tier was our biggest cost centre. A handful of power users on Pro were profitable; the rest were marginal. And our enterprise tier? Some accounts were net negative — we were losing money on AI before we even counted support and infra.
One enterprise customer was spending $4,200/month on our API. They paid us $3,000. We were subsidising them. Another was at $800 spend on a $2,500 contract — healthy. The averages hid both extremes. You can't fix what you can't see.
What we changed
We gated our heaviest AI feature (document analysis) behind Enterprise. We added a 100-request cap on the free tier. We raised prices for high-volume Pro users and added usage-based overages for enterprise. None of that would have been possible without the numbers.
The board asked the same question at the next meeting. This time we had a dashboard. Cost per customer by tier. Margin by segment. Which features were profitable. It changed the whole conversation. You can't have a pricing discussion with your team — or your board — without knowing who costs what.
If you're in the same boat, we built PerUnit to connect OpenAI and Anthropic usage to your customer list and Stripe data. Cost per customer, margin by tier, and which features are profitable. Early access is $99/mo — we're onboarding a small group now. To understand your margins first, try our free AI margin calculator — it shows cost per customer and AI as % of revenue.