Amazon staff report costly AI deployment errors as internal meeting reveals ‘catastrophically expensive’ overruns

Amazon’s engineers have a spending problem, and it starts with two letters: AI. During an internal meeting on July 28, staff laid out a pattern of budget overruns across multiple AI projects that one team described as “catastrophically expensive.”

The most jarring example: a product listing project that exceeded its budget by 860%. The overspend went undetected for five months.

The damage, project by project

The listing project wasn’t an isolated incident. A separate deployment built on Anthropic’s Claude Sonnet racked up $1.8 million in unplanned costs. That project never even shipped.

Other AI initiatives still blew through their budgets. Financial auditing tools ran $541,000 over budget. A logistics improvement project came in $134,000 above plan.

A culture problem, not just a cost problem

Consider KiroRank, an internal AI token leaderboard that Amazon suspended on May 29 after employees gamed the system. Staff were essentially competing to use the most AI compute, and token consumption ballooned as employees inflated their usage to climb the rankings.

Then there’s Kiro itself, Amazon’s AI coding assistant, which has been linked to notable outages.

Amazon’s response: automated guardrails

Amazon is now implementing mandatory safeguards for AI projects, including peer reviews and spending caps designed to catch runaway costs before they spiral.

These revelations come amid sustained corporate layoffs across Amazon, which makes the optics of burning $1.8 million on an unshipped AI project particularly uncomfortable.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.