GoKawiilSign up to see the future, today Can’t-miss innovations from the bleeding edge of science and tech Email address Sign Up Thank you!
Even the gazillion-dollar empire that is Amazon is reeling from the costs of letting its employees run loose with AI agents.
The Financial Times reports that staff at the company have identified cases of “catastrophically expensive” cost overruns as it switched over to using AI tools to perform what sound like pretty routine coding tasks.
In one case, Amazon blew $1.8 million by using Anthropic’s Claude Sonnet AI to match author details with product listings on its namesake ecommerce website. This menial job somehow went 860 percent over the allocated budget. And equally alarming was that it took five months to notice the piles of money being set on fire.
“It’s difficult to figure out how much anything [AI related] costs,” one senior Amazon employee complained to the FT.
Senior engineers who presumably got their start before “vibe-coding” became all the rage warned that these weren’t isolated incidents, lamenting that coding mistakes that were once “trivially cheap” were now “catastrophically expensive.” Amazon wasted $541,000 in unexpected costs related to building financial auditing tools, per the FT. And it also accidentally burned through $134,000 after giving AI the daunting task of improving delivery speeds across its logistics network. (Magical thinking much? These things aren’t genies, contrary to what Sam Altman promises.)
Amazon’s unintentional spending binges are part of a broader reckoning in the business world as leadership grapples with the costs of deploying AI tools across a company.
Not long ago, the prevailing ethos of “tokenmaxxing” encouraged companies and software engineers to use AI agents, and especially coding tools, as much as possible. Inevitably, this led to unfortunate incidents such as one company that reportedly blew half a billion dollars in Claude usage fees in a single month.
These costs concerns have been underscored by AI companies transitioning to token-based billing instead of flat subscription rates.
Amazon was certainly guilty of reveling in the debauchery “tokenmaxxing” era. It hosted an internal leaderboard, for example, that ranked employees based on how much they were using their AI tools. It was eventually shut down in May, presumably as the spending hangover started to set in.
... continue reading