From practice3 min read
The first meter hung on our own AI
Tokenometrics did not come out of a workshop. It exists because we could not explain our own invoice, and then found the same gap at every client we looked at.
by Jonas Leeb
Tokenometrics had a customer before it had a waitlist. The customer was us.
Nine products, one invoice
The lab holds nine built products, from wind farm permitting to genomics. They share infrastructure, tooling and, as it turned out, one line in the books. While AI sits inside a prototype, that line is a rounding difference. It does not stay one.
haki agent was the tipping point
haki agent is our own development process as a product: you commission a task instead of writing a prompt. A team of roles works through it, and a human approves every step.
That construction is exactly what makes cost unpredictable. A commissioned task is one operation to us and dozens of calls to the provider: planning, research, drafting, review, correction, and after a rejected approval possibly from the top. We could say what a month had cost. We could not say what a task had cost.
Why the invoice did not help
The provider's export knows the model, the tokens, the key and the timestamp. It knows no tasks. Attribution by operation was missing precisely the field that existed only inside our systems, and there only for as long as the workflow was running.
What we built instead
A meter is conceptually simple and unpleasant to implement. Every call carries an identifier that ties it to its operation. The operation gets its outcome attached at the end: accepted, discarded, repeated. From there you can work backwards to what an outcome cost, instead of forwards to what tokens cost.
Capture runs through an SDK and a proxy, and evaluation happens where capture happened. Keeping the data inside the environment started as convenience rather than positioning: we did not want our own prompts sitting in somebody else's service. The first customer conversations showed that this is exactly where vendor reviews get stuck.
Why there is no savings percentage here
This is the point where a post like this normally quotes a percentage. We have numbers, and they come from a single building, namely ours.
One building is one data point. Turning one data point into a rate and putting it on a website is precisely what we hold against everyone else. The number appears here once three measurements carry it, with the derivation beside it.
What became an offer
The part that carried revenue first is not the platform. It is the parser: the piece that reads billing exports mechanically instead of sorting them by hand every time. It is the stage before measurement, it needs no access to anyone else's systems, and we call it the Kostenbild.
The platform is the scaling step, not the entry point. What it does, we deliver today as a service, and every engagement sharpens the attribution logic that later sits inside the product. A product built out of a service sold twelve times has proven its demand before it goes live.
The electricity meter did not make electricity cheaper. It made accountability possible.
That is where AI stands right now. We walked the length of it before we started selling it.
what does one outcome cost you?
Not one token, not one month. One finished outcome. I'll get back to you within 24 hours.
