The “Tokenpocalypse”: Who’s Really Burning Companies’ AI Budgets

Spending on generative AI keeps climbing, and companies are starting to look in earnest for where exactly the money is leaking away. One telling episode is described in a piece by 404 Media (citing a leaked audio recording of an internal Accenture meeting) and retold on Simon Willison’s blog.

The engineers aren’t to blame

According to Justice Kwak, who heads up agentic-AI strategy at Accenture, the internal data reveals something unexpected: the bulk of the token consumption is not being driven by developers at all. The lion’s share of the spending comes from staff far removed from engineering — through their everyday ways of working with AI.

Stuart Henderson, who leads the client side of the business, jumps into the conversation. Half-joking, he says he hopes his colleague hasn’t been the one turning PDFs first into a set of images and then into Markdown files. “I’m learning here that this is one of the biggest token guzzlers,” he says. “Turning PDFs into Markdown — right?” Kwak confirms it: that’s exactly what the company’s statistics show.

Why it’s so expensive

The scenario looks harmless, but it comes at a steep price. To “feed” a PDF’s contents to a model, the document is often rendered into images first, and then the AI is asked to recognize the text and lay it out as tidy markup. Each such pass through images means a large volume of tokens — and therefore a noticeable line on the bill. When thousands of employees across an organization do this, it adds up to a hefty sum.

Willison rounds out the anecdote with an ironic thought of his own. Since Accenture has discovered from its own data that PDF is an extremely inconvenient format for conveying information, perhaps the company ought to share that revelation with the rest of the business world too. A format that has been considered the standard of business correspondence for decades turns, in the era of language models, into an expensive obstacle: what’s convenient to print and sign, a machine has to painstakingly take apart again piece by piece.

The story is amusing, but behind the joke lies a thoroughly practical takeaway. As AI tools become part of everyday work, the bill is increasingly determined not by complex engineering tasks but by the routine habits of ordinary users. And before optimizing their models, companies might do well to take a closer look at the form in which they store and share their documents in the first place.

Source: Simon Willison