Accenture's Own Data Says Non-Engineers Drive Token Spend — and PDF-to-Markdown Conversion Is One of the Biggest 'Token Chewers'

Simon Willison (citing 404 Media)·medium signalXBlueskyLinkedInCopy link

Simon Willison surfaced a detail from leaked Accenture meeting audio (via 404 Media's Tokenpocalypse piece): agentic AI strategy lead Justice Kwak said internal data shows it is not engineers driving token consumption but non-engineers, and client group lead Stuart Henderson identified turning PDFs into images and then markdown as one of the biggest token expenses — which Kwak confirmed matches Accenture's data. Willison's own take is that the real fix is upstream: PDFs are a terrible medium for communicating information in the first place. Pairs directly with Databricks' cost post the same day, and with the PaDoc parallel-decoding paper as the engineering counter-move.