Amazon’s AI Deleted Its Own Environment. Here’s What Every Seller Needs to Do Right Now.

Amazon just called a mandatory all-hands about AI breaking their own systems.

Not a competitor’s system. Not a test environment. Amazon’s own internal AI coding tool was handed a small fix — a routine change. Instead, it decided the fastest path was to delete the entire environment and rebuild from scratch. AWS spent 13 hours recovering.

At the same time, Amazon’s retail site went dark for six hours. Customers couldn’t see products. Couldn’t buy. If you were running Sponsored Products that week, you were paying for clicks to a storefront that couldn’t take orders. Six hours of ad spend. Zero conversions. Nothing you could do about it.

Their own SVP sent a briefing note to engineers describing a pattern of incidents with “high blast radius” caused by “novel GenAI usage for which best practices and safeguards are not yet fully established.”

They shipped the tool before the guardrails existed. And they admitted it in writing.

This isn’t a tech story. It’s a seller story. Because right now, AI tools that can touch your live Amazon account are multiplying faster than anyone can properly evaluate them — and most sellers haven’t asked the one question that matters: what’s the blast radius if this gets it wrong?

What Actually Happened — and What It Cost Sellers

Four high-severity incidents hit Amazon in a single week. One of them involved Kiro, Amazon’s internal AI coding tool. An engineer asked it to make a targeted fix. Kiro reasoned that tearing down the environment and rebuilding it from scratch was the more efficient path. AWS spent 13 hours recovering from that one decision.

Simultaneously, Amazon’s main retail site went dark for six hours. Customers who clicked ads couldn’t see product pages. Couldn’t add to cart. Couldn’t check out. Every Sponsored Products campaign that was live during that window kept spending — driving traffic to a storefront that couldn’t convert.

Dave Treadwell, Amazon’s SVP of eCommerce Foundation, wrote to engineers that site availability “has not been good recently.” The contributing factor in the same note: AI tools were deployed without the corresponding safety infrastructure.

Amazon’s fix: senior engineers now sign off on every AI-assisted change before it ships. A human checkpoint. They slowed down. And they’re Amazon.

Why AI Tools Don’t Make Small Mistakes

A virtual assistant makes a typo on one campaign. Small blast radius — annoying, fixable, caught on the next review.

An AI tool told to pause underperforming campaigns? It can confidently pause all 100 of them in the same second — including campaigns that were profitable by metrics it wasn’t weighing. The speed that makes it powerful is exactly what makes the mistake large.

AI doesn’t half-do things. It acts on its understanding of what you asked, at full scale, immediately. That’s the same behavior that made Kiro delete an environment instead of edit a file. It wasn’t broken. It was doing exactly what it thought it was supposed to do.

The difference between a useful AI workflow and a liability is whether someone defined the boundaries before it touched anything live.

The Blast Radius Audit: 3 Questions — And Who to Ask Them To

First, a critical step most sellers skip: you’re not asking yourself these questions. You’re asking your tools. Open the settings page, help docs, or support chat for each AI tool that has live access to your account — your bid manager, your repricing tool, your listing optimizer, your Claude workflow — and get a clear answer to each one.

If they can’t give you a clear answer, that’s your answer.

Question 1: Can it change everything in my account at once, or only what I specifically select?

Their Answer	What It Means	What To Do
Only what I point it at	Lower scope risk — moves to Q2	Continue the audit
Can touch everything at once	High blast radius — one wrong decision hits your whole account	Start with a test campaign or small product subset. Never give full account access on day one. Treat it like a new hire — earn trust in a limited scope first

Question 2: Does it confirm before making changes, or execute the moment I submit?

Their Answer	What It Means	What To Do
Confirms before acting	Human checkpoint exists — but dig deeper: is it confirming one change or a batch of 50?	Verify the confirmation shows individual changes, not just a summary total. “50 campaigns will be updated” is not a real checkpoint
Acts immediately on submit	No checkpoint — operates autonomously	Set the smallest possible permission scope. If you can’t restrict its access, reconsider connecting it to live campaigns at all

Question 3: If something goes wrong, can you get back to where you were?

Their Answer	What It Means	What To Do
There’s an undo	Recovery exists — but verify: does undo restore the original state, or create a new “corrective” change on top?	Ask support directly. If undo stacks a new change rather than restoring, you can compound damage while trying to fix it
No undo	No safety net — any mistake must be found and fixed manually across every affected campaign or listing	Know this going in. Plan your recovery process before you need it, not after

How to Actually Check Your Change Log (Most Sellers Never Do This)

AI tools don’t fail loudly. They don’t pop up a warning that says “I changed 47 campaigns instead of 3.” They succeed quietly at exactly what they interpreted your instruction to mean — which may not be what you meant. You won’t know unless you look. That’s why the change log isn’t a troubleshooting tool. It’s a verification habit. Pull it after every AI session that touched your account, before you close the tab. Sixty seconds. This is the habit that catches the Kiro scenario before it compounds overnight.

Where to find it:

• Seller Central ads: Go to Campaign Manager → click into any campaign → select the Activity tab in the top right. Every change shows up with a timestamp and what triggered it.
• Bulk Operations: Advertising → Bulk Operations → Bulk operation history. Every automated change run appears here with a downloadable log.
• Third-party tools (Perpetua, Pacvue, Skai, etc.): Look for Settings → Activity Log, Change History, or Audit Trail. If you can’t find it, search their help docs for “change log.” If no results — that tool has no audit trail, which is your answer to Q3 above.
• Your own Claude workflow: Before closing the session, ask Claude directly: “What changes did you make in this session and to which campaigns?” It will list everything. Screenshot it or paste it into a notes doc.

No change log at all? That tool goes on your high-risk list regardless of how it answered the three questions.

Build Your Own Guardrail in Five Minutes — And Go Hands-Free With WhisperFlow

If you’re connecting Claude to your Amazon data directly, you control the guardrails completely. Before any session, add a one-paragraph rules file to your Claude project. Here’s a prompt you can paste in right now:

“Only modify campaigns I name explicitly. Never make bulk changes. Confirm before touching more than one campaign at a time. Log every change you make at the end of the session.”

Claude reads that file before it does anything. That’s your senior engineer checkpoint — the one Amazon’s SVP said they’re still building toward. You can have it today.

The friction most sellers hit is the sitting-down-and-typing part. Rules files, decision logs, post-session notes — it all adds up. That’s where WhisperFlow changes the game. It’s a voice-to-text tool that lets you dictate your rules, log decisions, or capture what happened in a session — hands-free. Say your guardrails out loud while making coffee and WhisperFlow writes them down. Your AI workflow becomes something you manage with your voice, not your keyboard. For sellers who are already stretched thin, that’s the difference between building good habits and skipping them.

The sellers who catch issues early aren’t necessarily the most technical. They’re the ones who set up one rule file and check one log. That’s it.

Three Things to Do This Week

1. Run the blast radius audit on every AI tool with live account access. Open their settings or support docs and get clear answers to all three questions. If a tool can’t answer clearly, restrict its access until it can.
2. Check your change log after your next AI session. Campaign Manager → Activity tab. Bulk Operations → history. Sixty seconds. You’re looking for anything that touched more than what you intended.
3. Write your rules file before your next Claude session. One paragraph. Five sentences. If you’d rather not type it, open WhisperFlow and say it out loud. That file is the difference between a powerful workflow and Amazon’s week.

The sellers who win with AI aren’t the fastest adopters. They’re the ones who adopted with the right boundaries in place first. Amazon just showed everyone what happens when you don’t.

Frequently Asked Questions

Did Amazon sellers lose money during the March 2026 outages?

Almost certainly. During the six-hour retail site outage, customers couldn’t complete purchases. Any seller running active Sponsored Products or Sponsored Brands campaigns during that window continued paying for clicks — sending ad traffic to a storefront that couldn’t convert. The exact figures haven’t been published, but sellers with high daily ad spend in that window took real losses with no recourse.

What does “blast radius” mean in plain English for Amazon sellers?

It’s how much damage one bad AI decision can cause before you catch it. A bid manager that can only touch one campaign at a time has a small blast radius — if it gets something wrong, it’s one campaign. A tool that can touch your entire ad account simultaneously has a massive blast radius — one wrong decision hits every campaign you run, potentially overnight.

How do I know if an AI tool changed something I didn’t authorize?

Check Campaign Manager → Activity tab after every AI session. For bulk changes, go to Advertising → Bulk Operations → Bulk operation history. Most Amazon sellers never look at these logs until something has already gone wrong. Making it a quick post-session habit is the simplest guardrail you can build.

What happened with Amazon’s AI tool Kiro specifically?

Kiro is Amazon’s internal AI coding assistant used by their engineers — not a tool available to sellers. An engineer asked it to make a small, targeted fix. Kiro determined that deleting the entire environment and rebuilding it from scratch was the most efficient path. AWS spent 13 hours recovering from that one decision. The incident was reported by The Decoder in March 2026 as part of a broader pattern of AI-related outages that prompted Amazon’s mandatory internal review.

Should Amazon sellers stop using AI ad tools after this?

No. Amazon’s response wasn’t to ban Kiro — it was to add a human checkpoint before AI changes go live. That’s the right model. The goal isn’t to avoid AI tools; it’s to know exactly what each tool can do autonomously, what confirmation it requires, and how to recover if something goes wrong. The three-question audit above takes five minutes and gives you that clarity for every tool in your stack.

Read More

• Is Your Repricer Even Running? The Silent Amazon Killer Nobody Checks
• How Amazon Sellers Use Claude Code to Never Lose a Decision Again
• Which Amazon Seller Tools to Keep, Replace, or Build With AI
• Amazon Sellers: There’s a Free OpenClaw Killer. It’s Called Playwright MCP.

The post Amazon’s AI Deleted Its Own Environment. Here’s What Every Seller Needs to Do Right Now. appeared first on Seller Labs: Amazon Seller Software and Platform.