PDF-to-Markdown Conversion Major Token Cost Driver - Accenture
Leaked internal meeting audio from Accenture reveals that converting PDFs into images and then into markdown files is one of the largest sources of unexpected AI token consumption inside the company.

👁 118 reads
Leaked internal meeting audio from Accenture, reported by 404 Media on June 24th, reveals that converting PDFs into images and then into markdown files is one of the single largest sources of unexpected AI token consumption inside the company - and it is non-engineers, not developers, who are responsible for most of it.
- Accenture's agentic AI strategy lead confirmed the company's own data shows PDF-to-markdown conversion is a major token cost driver
- The behaviour is being driven primarily by non-engineers, not technical staff, according to internal Accenture data
- The exchange surfaced in a 404 Media report published June 24th, based on leaked internal meeting audio
The exchange involves two senior Accenture figures. Justice Kwak, the company’s agentic AI strategy lead, notes that internal data shows token consumption is not being driven by engineering staff. Stuart Henderson, Accenture’s client group lead, then asks whether PDF-to-markdown conversion is one of the culprits.
“I’m learning that’s one of the big token chewers. Turning PDFs into markdown: is that right?”
Kwak confirms that is exactly what Accenture’s own data shows. The anecdote was highlighted on August 7th by developer Simon Willison in his link blog, where he added a pointed observation: if Accenture can identify PDFs as a poor medium for communicating information internally, it may be well placed to spread that message to its enterprise clients.
The story fits into a broader pattern the 404 Media piece describes as the “tokenpocalypse” - a scramble by companies to control AI inference costs that have grown faster than anticipated once employees outside technical roles began using agentic tools at scale.
Practical takeaway: For anyone building or operating AI workflows, PDF ingestion pipelines deserve a close look at the cost level before they go near a production agent. Converting a PDF to images and then parsing those images into markdown can multiply token usage dramatically compared to extracting plain text directly. Auditing which file types and conversion steps are flowing through an LLM - and who is triggering them - is now a straightforward cost-control measure, not just a best practice.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →More on this from QUANTSLANT
magazine·Jeff Dean and Senior Google AI Figures Move On in Shake-Up

