The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI
Simon Willison’s Weblog Simon Willison ● Covered by 4 sources
Leaked meeting audio shows Accenture's own data blaming non-engineers for its AI token bill, not the engineers. Turns out a big chunk of that spending comes from turning PDFs into markdown files.
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
There's a small, telling moment buried in a 404 Media piece published June 24th, built off leaked audio from an internal Accenture meeting. Justice Kwak, who leads agentic AI strategy at the firm, tells colleagues that the company's internal data points somewhere unexpected: it's not the engineers driving up token consumption. It's everyone else.
Stuart Henderson, Accenture's client group lead, jumps in with a joke, asking Kwak if he'd just converted a PDF into images and then into markdown files. Kwak's response isn't a punchline. It's confirmation. That workflow, according to Accenture's own numbers, is one of the biggest token chewers in the building. Henderson repeats it back almost in disbelief: turning PDFs into markdown, that's really what's doing it.
And there's something almost absurd about a company deep in agentic AI strategy discovering that its biggest cost driver isn't sophisticated model reasoning or complex automation. It's people wrestling PDFs into a format a language model can actually read. PDFs were never built to be machine-friendly. They're built to look right when printed, which is a very different problem than being easy for software to parse.
The clip made its way around via 404 Media's TikTok account, which is its own small sign of how these stories travel now — internal corporate meeting audio surfacing not through a press release but through a leak turned into a short video. Kwak and Henderson weren't trying to make headlines. They were just talking shop, and in doing so accidentally revealed something a lot of companies adopting AI tools are probably not tracking closely enough: where the tokens are actually going.
My take — AI-written commentary, not fact-checked reporting
Blaming non-engineers for burning through the token budget is a convenient story, but the real culprit is the file format everyone's stuck using. Companies love to frame overspending as a discipline problem — use the tool better, train people harder — when the actual fix is boring and structural: stop making humans and machines fight over PDFs. That's a pattern worth watching well beyond Accenture, because plenty of firms are about to discover the same expensive habit hiding in their own logs.
Read more about this at: Simon Willison’s Weblog
