Nobody Can Tell You What An AI Task Costs: Not Even The AI

Konstantin Klyagin is the Founder of Redwerk and QAwerk, driving innovation in custom software development and quality assurance since 2005.

​Over the last year, we’ve seen top companies shift from leaderboards that track AI use to capping AI tokens per employee. Both extremes miss the mark. We now have solid proof that the true cost of an AI task is impossible to predict reliably. So, how do you protect yourself from massive, unexpected AI bills?

​The answer isn’t just tracking the dollars you spend on tokens. In this article, I’ll give you an honest breakdown of how to calculate the real value of your AI use, so you can make the best decisions for your business.

​Tokenmaxxing In Real Life

​First, let’s recap the tokenmaxxing hype that has just died down. We have some very good examples here:

​Amazon ran an internal leaderboard to rank employees on token consumption. This led people to use AI for every trivial thing, and computing costs shot right into outer space. Uber burned through the 2026 AI budget within four months. The company’s course correction was to cap tokens at $1,500 a month per employee (per tool), adjustable for some engineers. Microsoft canceled most internal licenses for external coding agents and moved thousands of engineers to its own tooling.

​The message is clear. Companies went from using AI for the sake of using AI to frantically budgeting its use. However, this solution might be just another trap.​

Here’s why. Measuring your AI value in dollars is just as faulty as measuring it in consumed tokens. Let’s look at OpenAI to see this in practice. It recently published a report on research acceleration within its labs. According to the data, they see an eight-hour workday of a researcher’s labor as equal to about 3.1 agent-workdays of research effort. However, the same report also states that over half of successful tasks that run four to eight hours require human intervention.​