Amazon Staff Flagged a $1.8 Million Claude Bill — And It Exposes the Enterprise AI Governance Gap

Amazon employees flagged a $1.8 million bill for using Anthropic's Claude AI to match author details with product listings, according to the Financial Times. The incident follows Amazon shutting down an internal AI usage leaderboard after employees gamed it, exposing a governance gap where enterprise AI costs compound silently without proper controls.

Amazon Staff Flagged a $1.8 Million Claude Bill — And It Exposes the Enterprise AI Governance Gap Financial Times reports Amazon staff flagged repeated 'catastrophically expensive' AI cost incidents including a $1.8M Claude spend on matching author details with product listings. Follows Amazon shutting down an internal AI usage leaderboard after employees gamed it. Illustrates the enterprise governance gap where API costs compound silently. The Financial Times reported this week that Amazon employees flagged repeated incidents of "catastrophically expensive" AI usage, including a single $1.8 million spend on Anthropic /glossary/anthropic 's Claude /glossary/claude for matching author details with product listings. The story follows Amazon's decision to shut down an internal AI usage leaderboard after employees started gaming it. The incidents matter because they're happening at Amazon — a company that builds and sells AI infrastructure for a living. If AWS can't keep internal AI costs under control, what chance does anyone else have? What the $1.8 Million Bought The specific project involved matching author details to product listings. It's the kind of task that sounds narrow until you consider Amazon's catalog: hundreds of millions of products, each with metadata that needs to be accurate, consistent, and cross-referenced against author databases, publisher records, and customer-facing listings. Running that matching at Amazon scale through Claude's API generated a $1.8 million bill. The number is eye-catching, but the real issue isn't the dollar amount — it's that nobody noticed until after the fact. The costs compounded silently while the work ran, and only surfaced when someone looked at the API billing dashboard. The Leaderboard Problem Separately, Amazon maintained an internal leaderboard tracking which teams used AI the most. The intent was to encourage adoption. The result was teams using AI for tasks that didn't need it, just to climb the rankings. Amazon shut the leaderboard down. Leadership issued guidance telling staff not to "use AI just for the sake of using it." The fact that a company at the center of the AI industry had to tell its own employees this is revealing. The Pattern The Amazon incidents fit a pattern emerging across large enterprises. Companies deploy AI APIs broadly — make them available to engineering teams, product teams, data science groups — without putting usage controls, cost caps, or approval workflows in place first. The assumption is that API costs are small enough to self-regulate. They aren't. Claude, GPT-5, Gemini /glossary/gemini , and comparable models charge per token /glossary/token , and at enterprise scale, tokens add up fast. A hundred developers running inference /glossary/inference on production-sized datasets can burn through six figures in hours without anyone raising an alarm. The governance gap is structural. Cloud infrastructure spending has decades of tooling around it — reserved instances, budget alerts, cost allocation tags, FinOps teams. AI API spending has almost none of that. The tools exist, but most organizations haven't integrated them into procurement and engineering workflows yet. What This Means Amazon isn't uniquely bad at managing AI costs. It's uniquely visible because of its size and because the Financial Times got access to internal discussions. Most enterprises running AI workloads have similar stories they're not sharing publicly. The lesson for organizations deploying AI at scale: cost controls need to go in before API access goes out. Usage caps per team, approval thresholds for large-scale inference jobs, automated alerts when spend crosses predefined limits — these aren't optional for responsible deployment. They're the difference between AI being a productivity multiplier and AI being an uncontrolled expense line. Amazon's leadership told staff to be more deliberate about AI usage. The better advice is to build systems that make responsible usage the default, not something that requires individual judgment calls after the bill arrives. Q: What exactly happened with the $1.8 million Claude bill at Amazon? Q: Why did Amazon shut down its internal AI leaderboard? Q: Is Amazon uniquely bad at managing AI costs? Q: What's the core governance problem? Q: What should enterprises do differently? Get AI news in your inbox Daily digest of what matters in AI. Key Terms Explained Anthropic /glossary/anthropic An AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Claude /glossary/claude Anthropic's family of AI assistants, including Claude Haiku, Sonnet, and Opus. Gemini /glossary/gemini Google's flagship multimodal AI model family, developed by Google DeepMind. GPT /glossary/gpt Generative Pre-trained Transformer.