Is your AI product losing money on its best customer?

Most founders look at blended margin and move on. Here's what happens when you slice it by customer — and why the largest contract on your books might be the one that's quietly draining you.

The scenario

Suppose you run an AI document-processing product. You charge a flat monthly fee per account. You have a mix of smaller customers and one significantly larger one — call them Account A — that closed at a number that made the whole team happy.

Your P&L looks healthy. Stripe revenue is up. OpenAI bill is growing but proportionally, you think. Nobody's looked at the per-customer numbers because you don't have a tool that shows them.

The blended view (what most teams see)

Suppose your monthly picture looks roughly like this:

That looks fine. Tight, but fine. A lot of early-stage AI companies are operating in this range. You ship the number to your board and move on.

The customer-level view (what you're missing)

Account A is your largest customer by contract value. They're also heavy users — let's say they generate 45% of your total API call volume. Here's the same data sliced by customer:

Still positive. But notice the gap: they represent 35% of your revenue and 45% of your LLM costs. That spread is the risk. Now account for scope creep.

Three months later, Account A adds two new internal workflows to their setup. Their usage climbs from 45% of API calls to 68%. The contract price hasn't changed.

They haven't churned. They haven't complained. They just expanded organically — exactly what you wanted. And now you're operating 7.7% gross margin on your biggest customer, subsidized by everyone else.

Why this is structurally hard to see

Your payment processor knows who pays and how much. Your LLM provider knows how much compute you consumed and what it cost. Neither of them knows anything about the other. The overlap — who's paying what, and what it costs to serve them — lives in a spreadsheet nobody is building.

Flat-rate pricing accelerates the problem. The more a customer uses your product, the more you pay in inference costs. If your pricing doesn't move with usage, you're exposed every time a customer grows.

High-volume customers are also usually your most vocal advocates. They give you case studies (hypothetically speaking). They expand seats. They refer others. Losing them is painful. So the math stays unexamined until a crunch forces it.

What to do when you find one

If you find a margin-negative customer, you're not automatically in trouble — but you need a deliberate decision, not an oversight. The options:

• RepriceMove to usage-based billing so revenue tracks cost. Existing customers often accept this at renewal if the value is clear.
• OptimizeIdentify which workflows are burning the most tokens. Often one use case drives 80% of costs and is a candidate for caching or smaller models.
• CapAdd hard rate limits on high-consumption tiers. Blunt, but sometimes the right call while you build the pricing model.
• Accept it deliberatelySometimes a large customer is a strategic anchor worth subsidizing short-term. That's a valid choice — but it should be a choice, not an accident.

None of these decisions are obvious without the numbers. What's not acceptable is not knowing.