Amazon’s internal teams are grappling with massive cost overruns tied to AI initiatives, with some projects wildly overshooting their budgets. An internal session on July 28 exposed a worrying trend of escalating expenses, described by one group as “catastrophically expensive.”
The standout example was a product listing AI project that blew past its budget by 860%, with the overspending unnoticed for nearly half a year. That alone highlights a serious failure in oversight.
Another costly misstep involved an AI deployment utilizing Anthropic’s Claude Sonnet model. It racked up $1.8 million in expenses before anyone caught on and the project never even made it to market. Additional initiatives didn’t fare much better: financial audit AI tools overshot by $541,000, while logistics enhancement efforts exceeded projections by $134,000.
Beyond the numbers, staff behavior seems to have fueled the problem. Amazon’s internal leaderboard for AI token usage, KiroRank, was shut down in May after employees gamed the system, boosting their consumption to climb the rankings. Meanwhile, Kiro, the company’s AI coding assistant, has faced reliability issues tied to notable outages.
Amazon is now rolling out automated guardrails for AI work including peer reviews and strict spending limits aiming to prevent runaway costs. This comes as Amazon continues corporate layoffs, making the $1.8 million burn on an unshipped project particularly sensitive.
