The project sounded mundane enough: use Anthropic’s Claude Sonnet to match author records to product listings on Amazon’s retail site. Routine data work. Instead, it quietly generated a $1.8 million bill — 860% over its original budget — before anyone noticed, according to people familiar with an internal presentation reported by the Financial Times. Five months of unchecked spending. The project never shipped. Senior engineers reportedly described the outcome as “catastrophically expensive.”
Here’s the math that puts the number in context. Claude Sonnet’s public rates run roughly $3 per million input tokens and $15 per million output tokens. At those input rates alone, $1.8 million translates to approximately 600 billion tokens consumed. For perspective, a typical developer coding task on the same model costs well under $5. The gap between normal use and what happened here is staggering — roughly the difference between a parking ticket and a mortgage.
The author-matching project wasn’t alone:
- A financial auditing tool reportedly added around $541,000 in unexpected costs.
- A logistics optimization system burned roughly $134,000 extra, per sources familiar with the internal presentation.
Mistakes that cost pennies in traditional software — inefficient loops, poorly designed workflows, runaway batch jobs — become ruinously expensive when an LLM executes every step at per-token rates.
The Real Problem Isn’t Claude – It’s Governance
Cultural incentives to use more AI collided head-on with an absence of cost controls.
Amazon had previously maintained an internal leaderboard ranking employees by AI usage volume — essentially a step-count competition where the prize turned out to be a budget crisis. The company dropped the practice after costs spiraled. AI coding bot errors also reportedly contributed to AWS outages earlier in 2026, prompting Amazon to reduce agent permissions rather than expand them. The pattern is consistent: organization-wide pressure to deploy AI, without corresponding discipline on what that deployment actually costs.
Amazon’s official response frames these incidents as isolated learning experiences against a backdrop of $181 billion in quarterly revenue — excess AI spend representing well under 0.1% of monthly earnings. The proportionality point is fair. Yet published governance analysis argues that most AI cost overruns stem from “operational sloppiness”: duplicate AI tools across teams, unbounded usage without role-based controls, and agents running without spending thresholds or review cycles. Cherry-picked or not, the structural risks are real and they show up at organizations paying too much far smaller than Amazon.
For any enterprise deploying AI agents internally, the case offers an unambiguous blueprint of what to avoid:
- Assign spend ownership per workflow.
- Set hard kill-switches on agents.
- Treat autonomous AI systems as production workloads rather than side experiments.
That $1.8 million receipt — attached to a project that never launched — confirms the cost of skipping governance is no longer a hypothetical.
