AI agent token consumption leads to unexpected corporate costs

Companies utilizing AI agents are facing unexpected financial strain due to the shift from traditional per-seat licensing to token-based consumption models. Because AI agents operate autonomously in the background, technical errors can trigger infinite loops that consume massive amounts of tokens before being detected.

Agorum® Software GmbH reported an instance where an image tagging process ran in a loop for a full day, costing 1,400 euros instead of the usual 80 euros. Oliver Schulze, CEO of agorum®, identified four primary drivers for these costs: multiple agents working simultaneously, long conversation histories being repeatedly sent, varying model prices, and technical errors causing undetected loops.

Large-scale examples highlight the severity of this issue. Amazon reportedly faced a project that went 860 percent over budget, costing approximately 1.8 million US dollars after a coding error caused an agent to restart tasks repeatedly. Similarly, Uber exhausted its 2026 AI budget within four months, with a single two-hour programming session costing 1,200 US dollars.