Amazon has identified several AI projects with drastic cost overruns, according to its own statements. One initiative, which aimed to match author data with product data using Anthropic’s language model Claude Sonnet, cost $1.8 million and exceeded its budget by 860% – the shortfall was only noticed after five months. The project ultimately delivered no usable results.
Three AI Projects Significantly Exceed Their Budgets
At an internal meeting on July 28, 2026, first reported by the Financial Times, Amazon’s senior engineers revealed several cases to their colleagues and described the results as “catastrophically expensive.” In the largest case, Claude Sonnet was supposed to identify duplicate or incorrectly assigned author entries in the product catalog and match them with the entries on the e-commerce website; costs totaled $1.8 million with a budget overrun of 860%. The responsible team had gradually replaced handwritten code with AI-generated queries without continuously monitoring the usage-based token costs or setting a spending limit.
In a financial auditing tool, an additional $541,000 in unforeseen costs was incurred. A third project, which aimed to reduce delivery times in the logistics network, caused $134,000 in unplanned expenses – the error was noticed after about two weeks here, while in the Sonnet project it was only noticed after five months. Only a few teams within the approximately 300,000 employees of the Amazon corporate division were affected.
Amazon Establishes Automated Cost Controls for AI Projects
The cause of the overruns is the usage-based billing of modern AI models: programming errors that would have been practically free to fix in classic code can quickly consume several hundred thousand dollars under AI systems. An Amazon official admitted to colleagues: “It’s difficult to figure out what AI usage actually costs in individual cases.”
Previously, Amazon had operated an internal ranking that classified employees based on their AI tool consumption, which was shut down in May 2026 because it inadvertently encouraged pure token consumption instead of saving costs. Now, the responsible engineers are working on automated control mechanisms with fixed spending caps and mandatory reviews by colleagues, which are intended to stop ongoing costs early in the future. Planned measures include automated alerts for unusual cost increases and a mandatory internal approval before a team switches from classic code to AI-supported methods. Other companies like Tesla, Uber, Meta, and Walmart have also recently set fixed caps on AI spending for their employees after similar issues were noticed.
Amazon Disputes the Representation as a Rule
An Amazon spokesperson denied to the Financial Times that the cases described reflect the usual handling of AI spending in the company, calling them isolated incidents. The report comes just days after Amazon’s quarterly figures from July 30, 2026, in which the company estimated its investment budget for 2026 at around $200 billion – mostly for AI and data center infrastructure. At the same time, Amazon recently cut tens of thousands of jobs in the company; CEO Andy Jassy repeatedly referred to future efficiency gains from AI, while the company emphasized that AI was not the main reason for most of the cuts. The cost problem is known industry-wide: Just recently, OpenAI CFO Sarah Friar publicly admitted that many companies find it difficult to quantify the economic benefits of their AI spending.
It remains to be seen whether the announced spending caps will be sufficient to prevent similar miscalculations or whether Amazon is merely addressing symptoms of an industry-wide problem with usage-based AI billing. The crucial factor will be whether the billion-dollar AI investments in the coming quarters translate into measurable efficiency gains.
