A flight to Seattle, and 100,000 lines after a 25-year gap
Beth Galetti, Amazon's senior vice president for People Experience and Technology, spent 25 years running the teams that manage software engineers without writing a line of code herself, and this week Amazon's own workplace blog said she has now written more than 100,000 of them using an internal AI tool called Kiro. Galetti described the return in personal terms: "Do you know how empowering it is to realize that after 25 years away from coding, you can actually build again?" Her first project was a family calendaring app, built for her own household rather than for Amazon's business.
The bigger story is the one Amazon chose to publish around her. Galetti wrote a submission for an internal competition called Everyone Can Build during a flight to Seattle, and the contest drew more than 1,500 employees within weeks of launching. Amazon's post highlighted other entries in the same spirit: one employee automated a four-hour manual process down to one second, another built a desktop pet called Momo for fun rather than for any business need.
Amazon is direct about what this is for. The post ties Galetti's coding return to Future Ready 2030, the company's newly announced 2.5 billion dollar commitment to prepare more than 50 million people for the future of work, including 1 billion dollars earmarked for the existing Career Choice education benefit and free AI training and AWS certification for every Amazon employee. A senior executive's personal app has become the human face of a very large training budget.
Five months earlier, Amazon's own engineers got the opposite instruction
Read next to the internal memo Amazon's engineering organization circulated in March 2026, the celebration looks like half the story. Five months before Galetti's post, Amazon required its trained software engineers, not first-time hobbyists, to get a senior colleague's sign-off before shipping any code an AI tool had helped write.
The trigger was a run of outages tied to what Amazon's internal communications called Gen-AI assisted changes with a high blast radius, a pattern the memo traced back to the third quarter of 2025. In December 2025, an AWS outage ran 13 hours after the Kiro coding tool was allowed to modify infrastructure without a human check and chose on its own to delete and recreate an environment. In early March 2026, two more incidents followed within three days: one cost roughly 120,000 orders and produced 1.6 million website errors, and the other cut order volume across North America by 99 percent, a reported 6.3 million lost orders. Amazon SVP Dave Treadwell told staff plainly that the availability of the site and related infrastructure had not been good recently.
TechRadar reported the resulting policy, citing an internal memo obtained by the Financial Times: junior and mid-level engineers must now get a senior engineer's sign-off before any AI-assisted code change ships, inside a 90-day code safety reset covering roughly 335 critical retail systems, with two-person review and stricter automated checks. Amazon disputed part of the framing, calling at least one incident a case of human access-control error, but it left the sign-off requirement itself in place.
A family calendar and a checkout page carry different risk
The difference between Galetti's app and the code that triggered Amazon's outages was never who typed the prompt, it was what the resulting code could touch. A family calendaring app and an internal contest submission read and write nothing that belongs to a customer, and nothing they do reaches Amazon's paying public.
The March policy targeted a narrower category: AI-written code running inside production systems that process real orders, real prices and real customer accounts, at a scale where one bad deployment removes millions of transactions in hours. That category needs a second trained person checking it before it ships, whoever, or whatever, wrote the first draft.
The threshold Amazon drew for itself is the transferable part of this story. It kept Kiro in use and kept encouraging non-engineers to build. It added a human check only once code reached production, touched customer data, or carried a blast radius large enough to matter.
The decision every operator now has in front of them
For an owner watching a marketing manager or a finance lead build their own tools with an AI coding assistant, the decision is not whether to allow it, Amazon's own year already answered that question. The decision is where to put the line Amazon put at 335 production systems and a mandatory senior sign-off.
Below that line, let people build. A spreadsheet automation, an internal dashboard fed by data the builder already has legitimate access to, a personal scheduling tool like Galetti's calendar app, none of it needs a formal review gate, and gating it anyway just teaches staff to build quietly instead of asking.
Above it, apply Amazon's own rule directly. The moment a non-engineer's AI-coded tool is set to read or write a production database, touch real customer records, or become reachable from outside the company, require a qualified technical reviewer to sign off before it goes live, on the same 2026 standard Amazon just used internally. If no one on staff can actually perform that review, pause the launch until someone qualified can.
What to check before the next AI-built tool goes live
Start this week by listing every AI-coded tool built outside the engineering team in the last quarter and asking one question of each: does it touch production data, a paying customer, or a system reachable from outside the company. A yes on any part of that requires the same sign-off Amazon now demands of its own engineers, regardless of who inside the company built it.
Put a name against the sign-off, not a policy document. Amazon's fix works because one specific senior engineer has to attach their name to a specific change before it ships, and a rule with no named reviewer behind it is a memo, not a gate.
Read next: Ring's New Default Is Not the Encryption It Sounds Like | The EU Team Behind DuckDB Just Joined AWS



