AI code wreaked havoc with Amazon outage, and now the company is making tight rules

Amazon has been aggressively pushing its engineers to adopt AI tools. At least 80% of its developers are expected to use AI for coding tasks at least once a week. However, recent events suggest that this fast-tracked rollout may have come at a cost.

As reported by the Financial Times, Amazon Web Services suffered a 13-hour outage in December after engineers let its Kiro AI coding tool update code without requiring any oversight. Kiro decided the best solution was to “delete and recreate the environment.” That’s one way to fix a problem, I suppose.

That wasn’t a one-off. A follow-up FT report revealed that Amazon’s e-commerce business has been dealing with a “trend of incidents” since Q3 2025, prompting a company-wide deep dive meeting led by SVP Dave Treadwell. 

Some employees were already skeptical about how useful these AI tools actually are for day-to-day work, and these incidents haven’t exactly helped build confidence.

Just how bad did it get?

Business Insider obtained internal documents that paint a clearer picture of what actually happened. On March 2, 2026, Amazon’s AI coding tools contributed to an incident that caused 120,000 lost orders and 1.6 million website errors. 

Three days later, on March 5, 2026, a separate outage caused a 99% drop in orders across North American marketplaces, resulting in 6.3 million lost orders. That’s a number that will surely show on the bottom line of a financial sheet, even for a company as big as Amazon. 

What is Amazon doing to ensure it never happens again?

Amazon is now rolling out a 90-day safety reset targeting around 335 critical systems. Engineers must get two people to review changes before deployment, use a formal documentation and approval process, and follow stricter automated checks.

The company maintains that these were user errors, not AI errors, and that the same mistakes could happen with any developer tool. That’s a fair point, but it doesn’t change the outcome. 

When artificial intelligence tools are handed broad permissions without adequate oversight, things break, and the scale of AI-generated code only amplifies the damage.