This week, reports emerged that Amazon had held a mandatory engineering meeting to address a string of recent outages on its retail website and app. The Financial Times broke the story on Monday, reporting that internal materials cited “GenAI-assisted changes” as a contributing factor, and that the incidents had a “high blast radius”. CNBC confirmed that the meeting was part of Amazon’s regular operational review, but that the AI angle had given it unusual urgency.
Amazon pushed back hard. In a public statement, the company said only one of the recent incidents involved AI tools “in any way”, and even then the root cause was an engineer following inaccurate advice that an AI tool had inferred from an outdated internal wiki. None of the outages, Amazon insisted, involved AI-written code. The company also denied that AWS services were affected, or that it had introduced new approval requirements for engineers using AI tools.
Since late 2025, Amazon has been steering its engineers towards Kiro, its in-house AI coding assistant, over external alternatives like Claude Code and Codex. Internal guidance actively discourages third-party tools for production work, requiring formal approval before engineers can use them. About 1,500 Amazon employees signed an internal thread calling for Claude Code to be formally adopted. One engineer wrote that Kiro’s “only survival mechanism becomes forced adoption rather than genuine value”. This is what happens when dogfooding (using your own products) becomes doctrine.
Kiro itself runs on Anthropic’s Claude models, and Amazon is one of Anthropic’s largest investors. Some Amazon engineers responsible for selling Claude Code through the company’s Bedrock platform questioned how they could credibly promote a product they were not permitted to use themselves. Meanwhile, Amazon has cut roughly 30,000 corporate roles since October, creating an environment where thinner teams face pressure to move faster with less oversight.
Whether AI directly caused these outages is debatable. What is harder to dispute is that Amazon pushed a young, internally controlled tool into production workflows before the surrounding safety processes were ready, while simultaneously restricting access to more mature alternatives and reducing headcount.
Takeaways: The lesson from Amazon’s rough week is not that AI coding tools are inherently dangerous. It is that mandating adoption of immature tools without re-engineering the surrounding processes is a recipe for exactly the kind of incident Amazon is now scrambling to explain. Software may be “solved” but the deployment, security, maintenance and governance of it is not.