While AI is an experiment, the spend is read as an infrastructure line. The moment it sits inside what you sell, it stops being infrastructure and becomes variable cost: every new customer brings revenue and brings an invoice. From then on, the number that matters is not the monthly spend, it is what it costs you to serve one customer.
At a glance
| What you measure | AI spend divided by active customers |
| What it is for | Margin per customer, pricing, plan limits |
| The trap | The average hides a very uneven distribution |
| Rule | Look at the median and the high percentile, not just the mean |
How to calculate it
You need two things: attributed spend and the right denominator. Attributed spend comes from tagging each request with the customer that caused it, which is why separating by key and by project is not an administrative luxury.
The denominator depends on your business model: active accounts if you charge per account, seats if you charge per user, cases processed if you charge per usage. Pick the one that appears on the invoice you send the customer, so you can put both figures side by side.
The average lies almost always
In AI products the distribution of consumption is very uneven: a handful of customers concentrates a good share of the spend. The average of that distribution describes almost nobody.
Look at three numbers. The median tells you what a normal customer costs. The high percentile tells you what the one getting away from you costs. And the maximum tells you whether you have a customer eating everyone else’s margin.
If the most expensive one costs twenty times the median, you do not have a token pricing problem: you have a badly designed plan.
What to do with expensive customers
Before raising their price, check whether their cost is intrinsic or accidental. Often the expensive customer is simply the one using the flow nobody optimized: the one who triggers the agent that takes twenty turns, or the one uploading three-hundred-page documents that travel whole on every request.
There the answer is technical and helps everyone: route per task, cache the fixed part and trim the context. Only when the cost is intrinsic to the usage does it make sense to touch the plan, with fair-use limits or a higher tier.
From cost per customer to price
With cost per customer in hand, three decisions stop being opinion. How much margin each plan actually leaves. Where to put the cheap plan’s limit so it does not eat the expensive one. And how much you can spend on acquisition knowing your cost to serve. All of it rests on cost per task, which is the layer underneath: cost per task times tasks per customer.
Frequently asked questions
How often should I recalculate it? Monthly is enough if the product is stable. After every model or prompt change, it is worth a look even off cycle.
What if I cannot attribute spend per customer? Start by attributing it per business function, which is easier, and split with the best proxy you have while you instrument the rest.
Should cost per customer fall over time? Yes, that is the sign optimization is working. If it rises while volume grows, something is getting away from you.
At e-ficient we measure the cost per task of companies that already run AI in production and tell you how much you would save with each lever at your real volume. The first audit is free if you start a plan with a three-month commitment and returns a diagnosis within 72 hours.