When Amazon Senior Vice President Dave Treadwell told engineers that site availability "has not been good recently," nobody needed a dashboard to confirm it. The usually optional weekly operations meeting, known internally as TWiST (This Week in Stores Tech), was made mandatory. The briefing note cited incidents with "high blast radius" and flagged "Gen-AI assisted changes" as a contributing factor.

The company is now requiring senior engineers to sign off on all AI-assisted code changes made by junior and mid-level staff before deployment. It is a significant admission from a company that has spent the past year aggressively pushing AI adoption across its engineering organisation.

The incidents have been piling up. Earlier this month, Amazon's main retail website and shopping app went down for nearly six hours. Customers could not view product details, access account information, or complete transactions. The company attributed the disruption to an erroneous software code deployment.

But the retail outage is only one piece of a broader pattern. In December 2025, AWS experienced a 13-hour service interruption after engineers allowed the company's Kiro AI coding tool to make changes to a live system. According to the Financial Times, Kiro determined that the best course of action was to delete and recreate the entire environment, triggering an extended outage of AWS Cost Explorer in one of the company's China regions.

A second incident, involving Amazon Q Developer, also surfaced in recent months. A senior AWS employee told the Financial Times that both outages were "small but entirely foreseeable."

Amazon has pushed back on the narrative. The company published a blog post calling the December incident "user error, specifically misconfigured access controls, not AI." But the fact remains: the AI tool was given operator-level permissions equivalent to a human developer, and it made a decision that most human developers never would have made.

Here is where the story gets harder for Amazon to manage. The company has cut approximately 30,000 corporate roles since October 2025, including 16,000 announced in January 2026. CEO Andy Jassy has been explicit that AI-driven efficiency would allow the company to operate with a leaner workforce. According to WARN filings analysed by FinalRound AI, nearly 40 percent of the October cuts targeted engineering roles specifically.

At the same time, Amazon had reportedly set an internal target of 80 percent of developers using AI coding tools at least once a week, tracking adoption as a corporate OKR. In November 2025, weeks before the Kiro incident, an internal memo established Kiro as the standardised AI coding assistant across the company.

The tension is obvious. You cannot simultaneously reduce engineering headcount, mandate AI tool adoption, and then act surprised when AI-generated code introduces production failures that fewer humans are available to catch. Multiple Amazon engineers have told the Financial Times that their teams are dealing with more "Sev2" incidents, urgent issues requiring rapid response to prevent outages, and they attribute the increase to job cuts.

Amazon disputes that workforce reductions are responsible for the rise in outages. But the pattern is hard to ignore.

Amazon is not operating in a vacuum. The collision between aggressive AI adoption and inadequate oversight is playing out across the industry.

Microsoft CEO Satya Nadella said in mid-2025 that AI writes up to 30 percent of the company's code, with some projects entirely AI-generated. By late January 2026, Microsoft acknowledged it was working to fix Windows 11 flaws and restore its reputation, nine months after that admission.

Google's 2025 DORA report on AI-assisted software development found that 90 percent of developers use AI for coding, but only 24 percent say they trust it "a lot." That trust gap is significant. Developers are using tools they do not fully trust because they are being told to, not because the tools have earned confidence through reliability.

Gartner has predicted that over 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. Forrester has gone further, predicting at least two major multi-day hyperscaler outages in 2026, driven partly by AI infrastructure upgrades taking priority over legacy system maintenance.

Treadwell's memo outlines Amazon's immediate response. The company is implementing what it calls "controlled friction," temporary safety practices that add deliberate checkpoints to changes in the most critical parts of the retail experience. In parallel, it plans to invest in "more durable solutions including both deterministic and agentic safeguards," according to CNBC.

The senior engineer sign-off requirement is the most concrete change. Junior and mid-level engineers can still use AI tools, but their AI-assisted changes to production systems now require approval from someone with the experience to spot the kind of errors these tools introduce.

It is a sensible policy. It is also one that probably should have existed before an AI agent was given the permissions to delete a production environment at a company that generates 57 percent of its operating profit from cloud services.

The broader question for the industry is structural. Companies are racing to make AI agents more autonomous because the productivity gains are real. But the guardrails are not keeping pace with the capabilities. The pressure to adopt is coming from boardrooms and earnings calls, not from engineering teams who understand the risks.

Amazon's response, slowing down, adding friction, requiring human review, is the right instinct. Whether it represents a genuine shift in how the company balances speed against reliability, or a temporary patch before the next push for AI-driven efficiency, will depend on what happens after the headlines fade.

Is "controlled friction" the answer, or are we watching companies apply plasters to a structural problem they created by cutting the humans who were the original guardrails?