The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI
Simon Willison’s Weblog Simon Willison ● Covered by 3 sources
Companies are freaking out over AI token bills, and Accenture's own staff blame a dumb habit: converting PDFs to markdown.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Simon Willison spotted a genuinely funny detail buried in a 404 Media piece from June 24th, sourced from leaked internal Accenture meeting audio. The consulting giant has apparently been digging into why its AI token usage keeps ballooning, expecting to find engineers running wild with expensive model calls. Instead, the culprits turned out to be non-engineers, and the specific behavior burning through tokens is almost absurdly mundane: converting PDFs into images, then into markdown files.
The exchange itself is the kind of thing that only happens when people forget the tape is rolling. Justice Kwak, Accenture's agentic AI strategy lead, starts explaining the data, and Stuart Henderson, the firm's client group lead, jumps in half-joking, half-horrified, asking if Kwak is talking about people turning PDFs into markdown. Kwak confirms it. That is exactly what the internal numbers show.
Willison's read on this is pointed and, frankly, correct: the real problem isn't the AI, it's the PDF. Companies have spent two decades treating PDFs as the universal format for sharing information, and now that same format is quietly torching compute budgets because it's a nightmare for machines to parse. Feed a document into a model as an image, ask it to extract text, reformat that into markdown so downstream tools can actually use it — every one of those steps costs tokens, and none of it would be necessary if the source document had been usable in the first place.
What makes the story land is the irony of it. A firm charging clients to build sophisticated agentic AI strategies is discovering that its biggest inefficiency isn't some exotic misuse of large language models. It's a file format problem that predates ChatGPT by twenty years. If Accenture manages to convince the broader business world that PDFs are a bad way to store information worth extracting, that might end up being a more useful contribution than any of its agentic AI consulting work.
My take — AI-written commentary, not fact-checked reporting
This is the most 2024 story imaginable: an AI company discovers its runaway costs aren't from some exotic misuse case, they're from feeding PDFs through three unnecessary conversion steps because nobody wanted to fix the actual data pipeline. Enterprises love bolting expensive AI onto broken document workflows instead of just... not using PDFs for machine-readable data. The token bill isn't an AI problem, it's a decades-old format problem finally sending someone an invoice.
Read more about this at: Simon Willison’s Weblog