The True Cost of Agentic AI: Why We Need a Value Logic

Most AI business cases pay off (as long as you use the cheap models).

I burned 200 euros on AI tokens last week. And in doing so, I finally understood why many AI transformation plans will fail.

As long as AI is thought of in the model of a monthly €20–40 subscription (as is STILL common today), the math is simple: A small license item meets high personnel costs. This makes almost any productivity increase seem immediately attractive.

This logic reaches its limits as soon as truly powerful agentic setups incur costs of >€100 per day.

I experienced it myself: The AGI feeling only arises with frontier models like Opus 4.6. There, millions of tokens are burned daily at high token costs when automating real workflows.

Anthropic is already preparing a new model called Mythos. Even more powerful than Opus, but also much more resource-hungry and therefore even more expensive. If this trend is confirmed, we will soon no longer be talking about subscription costs for agentic workflows, but about infrastructure costs equivalent to the salaries being replaced.

A new AI multi-class society is then emerging. A small part works with the best models and experiences this almost magical productivity. The vast majority works with significantly weaker models and wonders why the results are not nearly the same.

There are significantly cheaper models, especially from China, but without the same “it just works” feeling. For a text summary, a cheap model is sufficient. But anyone who wants to use it to automate complex controlling workflows across SAP system breaks needs frontier models—and those cost a pretty penny.

If frontier models generate costs in complex workflows that are not far off from salary costs, the classic efficiency argument is no longer sufficient. Then the question is not just whether employees work more cheaply with AI, but what additional value the company creates with these capabilities.

That is exactly where I see a gap in many AI transformation plans.

A lot is still thought of from a cost logic perspective:

• less effort
• fewer personnel requirements
• more output per capita

Too little from a value logic perspective:

• What new services become possible?
• Which processes can be mapped economically for the first time?
• What additional quality, speed, or customer proximity is created?
• Where does this lead to additional revenue or a strategic advantage?

If the strongest models remain expensive, the best AI setups will not automatically pay for themselves through savings. They only pay for themselves if new value is created from them.

This is exactly what companies should be thinking about much more seriously now.