English · Nederlands
Every AI pricing discussion I join sooner or later arrives at the same question: how do we bill this per token? It is a fair question for your finance team and the wrong one for your customer. In 2011 I wrote about a buddy finder demo we gave at T-Mobile in 2003. Reading it again in 2026, the manager’s reaction sounds like something I heard last week.
What I saw then
Early 2003, at T-Mobile Netherlands. We had a working buddy finder on a Windows smartphone, an app before anyone called it one. The GPS came from a separate receiver that we had to pair over Bluetooth. Our showpiece: you get a ping when a friend enters a geofenced zone you defined yourself. Back then, a real feat.
The manager across the table had one question: “So how can I bill this service out on a per minute basis?”
His bonus depended on call minutes, so call minutes were the only lens he had. In the post I blamed us as much as him. We had thought only from the consumer’s side and not from the telco’s business model. My conclusion was that innovation needs people who can see both perspectives, and that the business model has to be fitted to the customer, not the other way around.
What is the same now
Replace “minute” with “token” and the meeting replays itself. Tokens are what the model provider charges you, so tokens are what finance can measure, so tokens end up on the price list. The minute was the telco’s cost and control unit in exactly the same way. Neither was the unit the customer cared about. Nobody wanted minutes; they wanted to know where their friends were. Nobody wants tokens; they want the contract reviewed, the ticket closed, the tender answered.
The telco story also shows what happens next. In May 2011, two months after my post, KPN told investors that SMS and voice traffic in the Netherlands were falling because customers were moving to messaging apps such as WhatsApp. CEO Eelco Blok said SMS traffic was down 10% year on year in March (Telecoms.com). The unit the operators had defended for a decade became something an app gave away on top of a data connection. The mobile plans I see today lead with data, with calls and texts thrown in.
That is the risk for anyone pricing AI per token. The unit you meter is the unit a competitor, a model provider or the customer’s own IT department can make free. Token prices keep falling, agents make token use unpredictable, and customers who host their own model stop buying tokens from you at all. I made the full case in Price AI on outcomes, not tokens.
There is a second parallel that gets less attention: the meter changes behaviour. Per-minute billing taught people to keep calls short. Per-token billing teaches customers to ration the very AI you want them to adopt. If your growth plan depends on an agent working in the background all day, a visible meter works against it.
What is different
The analogy breaks in three places, and all of them matter.
First, a minute on an existing network cost the operator very little. A token has a real variable cost, and with agents that cost can swing widely from one task to the next. You cannot bundle the meter away the way telcos eventually did with minutes. You need it, just not on the customer’s invoice.
Second, for some buyers the token is the right unit. A developer buying raw model access through an API understands tokens and can forecast them from their own traffic. The mistake is not metering tokens. It is metering tokens for a business buyer who pays for an outcome and has no way to turn tokens into a budget line.
Third, the clock runs faster. Operators had years to watch messaging apps eat into SMS before they rebuilt their plans around data. Inference prices move in quarters, so a token-based price list can be out of date before the first renewal.
What to do about it
- Start with what the customer counts. Before you design a meter, write down the unit your buyer already uses in their own budget: documents, cases, tickets, tenders, users. If you cannot name it, you are not ready to price.
- Keep tokens as an internal cost spine. Track consumption, margin and heavy accounts per customer. Finance needs this. The price list does not.
- Check the scorecards. The T-Mobile manager was not stupid; he was paid on minutes. If your sales team is paid on credits consumed, expect them to sell consumption. Pay on retained revenue and on adoption of the outcome instead.
- Ask who can make your unit free. For every billing unit, name the party that could give it away: a model provider bundling it, an open-weight model, the customer’s own platform. Then price on what remains when that happens.
- Run the budget test. Show your draft model to three customers and ask whether they could predict next year’s bill from their own plans. If the answer is no, procurement will stall it or discount it.
The lesson, 23 years later
We lost that meeting in 2003 because we did not speak the manager’s language. In AI the risk runs the other way: vendors speak only the language of their own cost base and expect customers to learn it. The companies that win this round will be the ones that translate.