Amazon Mandates Senior Approval for AI-Assisted Code

Amazon held a company-wide engineering meeting this week to examine a pattern of high-impact outages tied to AI coding tools. The trigger: a six-hour shopping site outage on March 5 that knocked out checkout, login, and product pricing for thousands of customers. The outcome: junior and mid-level engineers now need senior approval before rolling out any AI-assisted code changes to production.

TL;DR

Amazon.com went down for ~6 hours on March 5 due to a faulty code deployment, peaking at 21,716 user reports
SVP Dave Treadwell emailed staff about "poor site availability" and convened a review meeting
New rule: junior and mid-level engineers must get senior sign-off for AI-assisted deployments
This follows at least two earlier AWS outages linked to the Kiro AI coding tool

The March 5 Outage

On March 5, Amazon's website and shopping app became inaccessible starting around 1:55 p.m. ET. Users reported missing product prices, checkout failures, login errors, and broken Amazon Fresh deliveries. Downdetector reports peaked at 21,716 around 3:48 p.m. ET. The platform wasn't fully restored until roughly 8 p.m. ET.

Amazon attributed the outage to a "software code deployment" in a statement to Reuters and CNBC. The company didn't specify whether AI tools were involved in generating or reviewing the code that was deployed. But the timing - coming weeks after the Kiro AI production outages we covered in February - put AI coding tools squarely in the spotlight.

The Engineering Review

The Financial Times, citing a briefing note it had seen, reported that Amazon's ecommerce division convened a large engineering meeting on Tuesday to discuss the outages. The briefing note attributed the incidents to "novel GenAI usage" with "best practices and safeguards not yet fully established." It flagged "high blast radius" as a recurring characteristic of the failures.

SVP Dave Treadwell, who heads Amazon's eCommerce Foundation and co-signed the November 2025 memo mandating Kiro as Amazon's standard AI coding tool, emailed staff about "poor site availability." The same executive who pushed for 80% weekly Kiro usage across the company is now dealing with the consequences of that adoption pace.

The New Approval Rule

The meeting produced a concrete policy change: junior and mid-level engineers must now obtain sign-off from a senior engineer before rolling out AI-assisted code changes to production. This adds a human review gate specifically for AI-generated code - an acknowledgment that the existing deployment process wasn't catching AI-introduced errors before they hit production.

The rule addresses a specific failure pattern. In both the December Kiro incidents and the March outage, engineers launched AI-assisted changes without a mandatory second-person review. Under normal protocol, Kiro requires two-person approval for production changes, but in practice that safeguard was either bypassed or not enforced.

The Pattern

This is at least the fourth significant incident in Amazon's AI coding rollout:

December 2025 - Kiro autonomously deleted and recreated an AWS Cost Explorer environment, triggering a 13-hour outage in a China region. Amazon blamed "misconfigured access controls."

Late 2025 - A second, less severe outage involving Amazon Q Developer. Three AWS employees confirmed to the FT that engineers let the AI resolve an issue without intervention.

March 5, 2026 - Amazon.com shopping outage lasting roughly six hours. Attributed to a faulty software deployment.

March 2026 - Engineering review meeting convened, new approval policy built.

The pattern isn't that AI tools are generating bad code. It's that the deployment pipeline wasn't designed for the speed and volume at which AI tools produce changes. When a human developer writes code, the pace of production is slow enough that review processes can keep up. When an AI assistant generates code at 10x the speed, existing review processes become the bottleneck - and teams skip them.

Where It Falls Short

The senior approval rule is a reasonable first step, but it has obvious limitations. It creates a bottleneck at the senior engineer level that'll slow deployment velocity - the opposite of what Amazon wanted from AI tools. It also assumes senior engineers can reliably spot AI-introduced bugs, which isn't guaranteed. The March 5 outage was caused by a code deployment that presumably passed some level of review already.

Amazon is also still dealing with the political fallout of the Kiro Mandate. Roughly 1,500 engineers protested via internal forums, arguing that external tools like Claude Code beat Kiro on tasks like multi-language refactoring. VP-level exception requests were rising. Now the tool those engineers were forced to adopt is linked to a string of outages.

The deeper problem is structural. Amazon deployed 21,000 AI agents across its Stores division, claiming $2 billion in cost savings and 4.5x developer velocity. Those numbers make it politically impossible to walk back AI adoption. So the company is adding guardrails to an already-rolled out system rather than reconsidering the deployment pace - fixing the plane while flying it.

What Comes Next

Amazon isn't alone in hitting this wall. We've covered similar incidents with AI coding agents breaking working code and the broader tension between AI speed and production safety. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs and inadequate risk controls.

The senior approval policy is Amazon admitting that AI coding tools need more human oversight than they currently get. Whether that admission scales faster than the tools themselves remains an open question.

Sources: