What This Post Cost Us in AI

This post cost us 61.0 cents in model fees by the time the German version was finished. Measured across its own four runs, not estimated:

• Topic suggestion: 15.4 cents
• Draft: 12.7 cents
• Editing run: 12.4 cents
• Cover image: 20.5 cents

The English version added 10.1 cents. Those are the figures for this one text, not an average. A post with more research, or with a second image run because the first motif did not fit, comes out higher. For us these numbers do not sit in a report at the end of the month, they sit on every single run. We see them before we press the next button.

That sounds like small change, and that is exactly the point. Small change gets dangerous when nobody is counting.

Why the bill explodes elsewhere

Reports of catastrophically expensive internal AI use have leaked out of Amazon, with costs well above what anyone had planned for. We have no insight into anyone else's books and we are not going to recalculate something we cannot see. In any case, the interesting part is not the total but the mechanism, and anyone who has ever built an automation knows it.

Depending on the model, a single call costs a fraction of a cent up to a few cents per thousand tokens. An agent that checks its own work calls itself twice. One that uses a tool and feeds the result back into the model gets to five calls. One that hangs in a loop because an interface returns an error it does not understand gets to a thousand. And because an automation runs without complaining, nobody notices. A person who does the same task wrong ten times in a row stops at some point. A loop does not stop.

Then there is the second effect: what is cheap gets used generously. The whole contract into the context window instead of the three relevant paragraphs. The complete product database on every run instead of just the changed rows. Harmless one by one, and together they are the actual bill.

Three build decisions that prevent it

This is not a question of discipline, it is a question of construction. Three things we build in as a matter of course.

One: every model call hangs on a button that a person presses. For this blog that means, concretely: the topic suggestion happens because someone asks for it. The draft happens because someone picked a suggestion. The editing run happens because someone wants the draft read over. The cover image happens because a person read the post and decided it was good. Four runs, four decisions. There is no path on which the system writes itself twelve posts because some condition was phrased wrong.

That is not just a brake on costs, it is our stance on AI in general: the machine proposes, a human approves. The pleasant side effect is that a system where a human grants approval does not order its own next run. It does not help against a loop inside a single run, that is what the time limit is for. But if you answer the quality question properly, you have largely settled the cost question along with it.

Two: every run has a time limit and a cost line. A run that is not finished within the agreed time is aborted. Not extended, aborted. And every run records what it cost. Not in a dashboard you could go and open, but in the same view where the result sits. If a system puts out a number and nobody can trace where it came from, the number is worthless. That holds for results, and it holds just as much for costs.

From that the third question almost follows on its own, the one to ask before any automation: what may a run cost for it to be worth it? For a blog post, 61 cents against the time a person needs for a rough draft and an image is not worth discussing. For an automation that runs ten thousand times a day, half a cent per run is a very different sum. The number itself is harmless, what matters is whether you know it before you switch the thing on for real.

Three: the model is replaceable. Model prices change, in both directions. Providers retire models. A model that is the right choice for a task today may not be in six months, because a smaller one is enough or another one has got cheaper. If the model call sits behind an interface of your own and the prompts belong to you, a switch is a configuration change and a test run. If not, it is a project.

That is exactly what we mean by sovereignty: not that nothing leaves the building, but that you keep the choice. Cloud AI is a legitimate decision, a local model on your own hardware just as much. What counts is that the decision stays reversible.

Where we see this in operation

Our second system of this kind is the social automation for Kunst gegen Bares, a live art stage in Düsseldorf. Campaign planning, a visual template editor, AI texts per channel, brand personas, publishing and analytics. The same rule applies there: the texts come out as a proposal, a person looks at them and approves them, and only then do they go out. And the last mile is outside the building anyway: if you want to publish on Instagram, you send data to Instagram. That changes nothing about the cost logic, quite the opposite. The more parts of a workflow sit with third parties, the more it matters that you know, for every run, what it cost and who triggered it.

The yardstick you can apply yourself

If someone offers you an AI automation, ask three questions: what does one run cost? Where is that number shown, without anyone having to work it out? And what happens if the model doubles in price tomorrow?

Anyone with an answer to all three has probably run something in production and not just demoed it. Anyone who starts estimating at the first question will still be estimating later.