Habr breaks down why token prices don't explain the economics of AI agents in business

Habr has published a piece examining the economics of AI agents in enterprise customer service, a topic that, in the author's view, usually gets reduced to a simplistic discussion of token prices.

What questions are usually asked about agent economics

The author lists the typical questions that come up when discussing the cost of AI agents: how much a million input and output tokens cost, what the budget burn rate will be, which model to pick to avoid overpaying, and how many GPU accelerators will be needed when hosting inside the customer's own environment, along with their utilization.

• Discussion of agent economics usually focuses on token prices
• Budget burn rate and model choice are taken into account
• For deployment inside the customer's environment, teams count GPUs and their utilization
• The author argues that for enterprise customer service this is not enough

Why token prices alone are not enough

The author stresses that all of these questions are the right ones, but when it comes specifically to enterprise customer service, the answers fall critically short of a full picture of the economics. A proper assessment has to account for additional factors beyond the direct cost of compute, including integrations, people and the learning curve of the system itself, which the title of the piece points to.

What this means

The piece is a reminder that when bringing AI agents into business processes, focusing only on token prices is a mistake: the real economics are built from integration costs, the changing role of people in the process, and the time it takes for the system to reach working efficiency.