Million-Dollar Budget Overruns: AI Usage by Employees Inflates Amazon's Costs

Errors in how Amazon implemented and operated AI models and the lack of control mechanisms in its organizational systems have led to "catastrophic overruns" of millions of dollars on these models, the Financial Times reports.

Over the past two years, many organizations have provided employees with access to AI models to assist with routine tasks such as writing code, creating presentations, summarizing meetings, and identifying insights in organizational data. Initially, companies, especially in high-tech, encouraged employees to expand their use of AI, sometimes measuring performance by the volume of model usage via the number of tokens consumed. This led to a trend dubbed "Tokenmaxxing," in which employees maximized their use of AI, sometimes artificially.

However, companies have recently realized that the cost of running these models often outweighs the value they provide, particularly when employees use them indiscriminately for relatively simple tasks. Amazon learned the hard way how much unchecked AI usage can cost. According to the Financial Times, at a team meeting held at the tech giant last week, senior engineers stated that attempts to offload programming tasks to models led to "unplanned" expenses.

In one case, Amazon spent $1.8 million trying to use Anthropic's Claude to match author details to records on the company's site, without success. This expenditure represented an 860% overrun compared to the project's planned budget and was only discovered after five months. This is not an isolated incident. In another case, Amazon incurred an unplanned expense of $541,000 on a project to build a financial audit tool. In a third incident, where AI was used to improve delivery speeds in Amazon's logistics network, the company recorded unplanned expenses of $134,000 that were only discovered after two weeks.

According to senior engineers, coding errors that can be "cheap and trivial" in traditional systems become "catastrophically expensive" once AI is involved. Engineers are now working on creating automatic protection mechanisms to prevent cost overruns in future projects.

"It's hard to understand how much everything related to AI costs," a senior Amazon official told the Financial Times.

These incidents are negligible relative to the scale of Amazon's operations, which recorded revenues of $200.61 billion and a net profit of $62.65 billion in the second quarter. However, they illustrate the challenges organizations face in integrating AI and the high, unexpected costs these systems can entail. In May of this year, Amazon shut down an internal leaderboard that ranked employees by their volume of AI usage after staff inflated their token consumption to climb the rankings.

Amazon responded: "As with any new technology, we are experimenting, learning, and improving how we use it, including how to optimize costs. Selectively choosing small, isolated examples where teams are learning from each other and presenting them as business as usual does not reflect how teams at Amazon use AI."