There has never been a better time to send an invoice for AI, and there has never been a worse time to receive one.
Orb, Metronome, Stripe, and a wave of usage-billing platforms sit alongside your product, meter every call, attach a price to each event, and turn raw usage into a clean invoice in real time. What used to be a hard engineering project is now a checkbox, while the bill coming in from your model providers is still a black box.
Every one of these tools is built to charge for usage. What none of them do is tell you whether that usage makes money.
You can meter every call and bill it flawlessly and still watch your margin disappear. While the invoice looks right, the P&L doesn’t. In AI, those two numbers pull apart faster than most teams expect.
The bill is a black box before it is ever an invoice
When a customer uses an AI product, a single request rarely maps to a single cost. It fans out into model calls, tool calls, retrieval steps, retries, and in agentic products a chain of downstream actions the user never sees. Your billing system meters the event it was told to meter. It doesn’t see the full cost of producing that event.
The number on the invoice is real, and the number underneath it is a mystery. You are charging with precision on top of a cost you can’t break down.
Finance teams know the feeling already. The provider bill arrives, it is bigger than last month, and no one can say which customers, features, or agents drove the jump. Agents can spend money whether or not anyone is watching.
Same price, ten times the cost
Here is where it turns from an accounting annoyance into a margin problem.
Set usage pricing at a fixed rate per call. Reasonable, clean, easy to explain to a customer. Now look at what a call actually costs you. Context length varies. Model routing varies, and the spread inside a single vendor runs 5x or more between a cheap model and a flagship. Retries and tool use vary. Put it together and your real cost per call can swing 10x from one customer to the next.
You charge everyone the same, even if you don’t spend the same to serve them. Some customers pay far more than they cost. Some cost more to serve than they pay, and on the invoice they look identical.
Traditional SaaS ran at 80-90% gross margin because the marginal cost of another user was close to zero. AI does not work that way. Inference is a real variable cost on every request, and it does not shrink with scale. Bessemer puts AI gross margins at 50-60% against the 80-90% SaaS was built on, and ICONIQ's 2026 State of AI report puts the average AI product margin at 52%. The margin you assumed is not the margin you have.
The timing makes it worse. You find out at quarter close, in aggregate, with no way to trace which accounts or features drained the P&L.
Roughly 80-85% of enterprises miss their AI forecasts by 25%or more. The cause sits upstream of the spreadsheet. You cannot forecast a cost you cannot attribute, and you cannot defend the number to a board that wants to know what the AI spend returned.
Why better billing cannot fix this
The instinct is to reach for a better billing platform. Move from invoice-based to wallet billing. Adopt usage-based pricing, or the outcome-based pricing that Salesforce and others are pushing into the market, a model that jumped from 2-18% adoption among AI companies in six months. These are real improvements, and every one of them improves how you charge.
Changing how you charge tells you nothing about what it costs. Usage-based and outcome-based models make that worse, because they move cost variance onto your side of the table. When you charge per resolution and the cost per resolution swings 10x, you have taken on the variance the customer used to carry. The pricing model raises the stakes on cost attribution. It does not supply it.
Every billing engine assumes an input it does not produce. Cost to serve, per customer. Most teams run the entire stack on a number they have never actually measured.
Attribute first, then price and control
The fix is to build the layer that billing takes for granted.
Attribute cost at the transaction level, by customer, by feature, by workflow, and join it to the revenue you already bill. Once cost to serve sits next to price, the questions that used to wait for quarter close get answered in real time. Which customers are unprofitable. Which features bleed margin. Which agents run up spend with nothing to show for it.
That visibility is also what makes aggressive pricing safe. You can set usage rates or outcome prices against your real cost curve instead of a guess. You can put guardrails on the customers and workflows that run hot. You can walk into a board meeting with cost, revenue, and margin per customer, not one aggregate figure and an apology.
This is the layer Revenium provides. It sits under the billing stack, connects cost per customer to what you charge, and turns "was it worth it" from a quarterly surprise into something you watch and act on daily.
Billing infrastructure will send a flawless invoice. It will not tell you whether that invoice describes a profitable business. In AI those two things diverge by design, and the gap is where the money leaks.
See how Revenium connects cost per customer to your billing stack.
