Change-control gap
Amazon, AI coding targets and the brakes that came later
Amazon · E-commerce & cloud · United States
Amazon pushed its engineers hard towards its own AI coding tools while cutting tens of thousands of corporate jobs. After a run of severe outages, an internal briefing cited “Gen-AI assisted changes” — and a senior executive promised “controlled friction.” Amazon insists AI was not the cause.
Internal briefing note (per FT)
A “trend of incidents” with a “high blast radius” and “Gen-AI assisted changes.”
Amazon’s public account
Only one incident involved AI tools, and “none of the incidents involved AI-written code.”
4
Sev-1 incidents in a single week (CNBC)
13 hrs
AWS interruption after Kiro changes, Dec 2025 (FT)
Either way, safeguards lagged adoption.
The record
- Amazon’s retail technology chief Dave Treadwell refocused a weekly operations meeting on site reliability after four Sev-1 incidents in one week; an internal document had listed “GenAI-assisted changes” as a factor in a trend of incidents since the third quarter, a bullet later deleted. (CNBC, Mar 2026)
- In December 2025 a 13-hour interruption hit AWS Cost Explorer after engineers let the agentic Kiro tool make changes and it reportedly chose to “delete and recreate the environment”; Amazon called it “user error, not AI error,” citing “broader permissions than expected.” (Financial Times via Engadget, Feb 2026)
- Amazon leadership set an 80% weekly-use goal for Kiro and tracked adoption closely (Engadget, Feb 2026), while the company cut about 14,000 corporate roles in October 2025 and 16,000 more in January 2026. (Fortune, Mar 2026)
Adoption as a target
By late 2025 Amazon was treating AI-assisted development as a company-wide objective rather than an experiment. A memo signed by senior executives Peter DeSantis and Dave Treadwell told engineers to stop adopting new external coding tools and standardise on Amazon’s own, Kiro (Reuters, Nov 2025), and leadership set an 80% weekly-use goal that was closely tracked (Engadget, Feb 2026). At the same time, Amazon cut roughly 30,000 corporate jobs across two rounds and planned around US$200 billion of AI-related capital spending (Fortune, Mar 2026). The direction of travel was unmistakable.
Usage targets are a blunt instrument. They reward engineers for routing more work through a tool, and they are easy to measure and report upwards. What they say nothing about is which changes a tool should be allowed to make, who checks those changes before they reach production, what permissions an agent inherits from the person running it, or how far a mistake can travel before something stops it. Those questions sit with platform and reliability teams — the kind of non-revenue functions that tend to thin out in a cost-cutting cycle.
Incidents with a long reach
The first warning was contained but telling. In December, according to the Financial Times, engineers allowed Kiro to make changes to an AWS system and the agent decided the best course was to “delete and recreate the environment,” causing a 13-hour interruption to Cost Explorer in one of 39 regions. Amazon called it a coincidence that AI was involved and blamed a misconfigured role: the staffer had “broader permissions than expected.” It added mandatory peer review for production access (Engadget, Feb 2026) — a control that, in hindsight, belonged in place before the agent arrived.
Then, on 5 March, Amazon’s store and app went down for several hours, leaving shoppers unable to check out, see prices or reach their accounts; Amazon blamed “a software code deployment” (CNBC, Mar 2026). A briefing for the next weekly operations review cited a “trend of incidents” with a “high blast radius” and “Gen-AI assisted changes” (The Register, Mar 2026). Amazon’s position is that only one of that week’s incidents involved AI at all, through an engineer following inaccurate advice that an agent had inferred from an outdated internal wiki (Amazon, Mar 2026).
Friction, reintroduced
Strip away the dispute over attribution and the two accounts converge on something important. Amazon’s own correction concedes that its systems allowed one engineer’s error to have a broader impact than it should have (Amazon, Mar 2026). Treadwell, according to the internal documents, acknowledged that “best practices and safeguards” around generative AI use were not yet fully established, and promised to introduce “controlled friction” into deployments touching the most critical parts of the retail experience (Fortune, Mar 2026). Nobody, in other words, claims the guardrails were already adequate.
That is the governance story, and it does not depend on settling who or what typed each faulty change. Speed was mandated centrally, with an adoption number attached; review, permissions and blast-radius limits were left to catch up locally, team by team. The sequence is the lesson: the controls arrived after the incidents, not alongside the targets. For a company with Amazon’s engineering depth to land there suggests how strong the pull of an AI usage metric can be — and how easily the unglamorous safeguards get scheduled for later.
The lesson
Scale review and permissions before you scale adoption. Pair any AI usage target with hard limits on what agents and AI-assisted changes can touch, mandatory review on critical systems, and a measured blast radius — so a bad change, human or machine, stays small.
How we’re reading this
Amazon disputes that AI caused these outages, says no incident involved AI-written code, denies introducing new approval rules for AI tools, and says AWS was not involved in the March retail incidents. Much of the reporting rests on internal documents and anonymous sources. Lost-order figures circulating online are not officially confirmed and are omitted here. We are not claiming the job cuts caused the incidents.
Sources
- 01Amazon plans ‘deep dive’ internal meeting to address AI-related outages
CNBC · Mar 10, 2026
- 02Amazon responds to inaccurate Financial Times report linking outages to AI
Amazon · Mar 11, 2026
- 0313-hour AWS outage reportedly caused by Amazon’s own AI tools
Engadget · Feb 21, 2026
- 04
- 05Amazon insists AI coding isn’t source of outages
The Register · Mar 10, 2026
Compiled from public filings, court records, company statements and reputable press. Figures are attributed to their sources; allegations are labeled as such. Not legal or investment advice.