Amazon AI Projects Face 860% Cost Overrun, $1.8M Wasted on Metadata Matching

Internal audits reveal severe AI cost overruns at Amazon, with a failed metadata project costing $1.8M, 860% over budget. Engineers cite lack of controls; automation measures planned.

Woofun AI reports that Amazon internally identified multiple AI initiatives suffering from significant budget deviations. One project utilizing Claude Sonnet for author data matching failed and incurred $1.8 million in expenses, exceeding its budget by 860% before detection after five months. Additional overspending included $541,000 in a financial audit tool and $134,000 in logistics optimization, with the latter discovered only two weeks post-expenditure.

Amazon engineers attributed these discrepancies to deployment errors and insufficient cost controls, noting that minor mistakes in AI systems can result in disproportionately high costs compared to traditional infrastructure. The company is implementing automation to prevent future spiraling expenses. Amazon stated these incidents are isolated and do not reflect overall AI usage, following a previous shutdown of an internal leaderboard where employees generated meaningless tasks to inflate metrics.