AI costs are turning from a budget line into an operational problem inside large organisations. A 404 Media piece published in June, based on leaked meeting recordings, shows that shift starkly through an internal discussion at Accenture.

An unexpected consumer

Justice Kwak, Accenture's agentic AI strategy lead, reports that internal data says something surprising:

"We're seeing from some of the data internally at least that it's actually not our engineers that are driving the token consumption. It's a lot of the non-engineers that are doing some of those behaviors."

That inverts most of the assumptions in cost planning. Organisations budget for AI based on what their software teams will use; but the bill is being driven up by the far larger group who use the tool in everyday work and have no intuition for what anything costs.

The PDF problem

What follows in the conversation identifies the concrete source. Stuart Henderson, Accenture's client group lead, interrupts to joke that he hopes Kwak did not just convert a PDF into images and then into markdown files:

"I'm learning that's one of the big token chewers. Turning PDFs into markdown: is that right?"

Kwak confirms that this is exactly what the company's own data shows.

Why it costs so much

The reason is technical and dull. PDF is a format designed not to carry text but to fix the appearance of a page. The text inside usually sits as positioned fragments; reading order, column layout and table structure are not written down anywhere in the file.

As a result the most practical way to convert a PDF reliably into text is to render the page as an image and have the model look at it. And images are far more expensive for language models than text: a page of plain text may cost a few hundred tokens, while an image of the same page can cost thousands.

Processed this way, a hundred-page report becomes a significant cost item on its own — and the person doing it generally believes they pressed a button.

The lesson

The real finding here is that the cost comes not from model pricing but from document formats and working habits inside the organisation. That explains the apparent contradiction of enterprise bills rising while model prices fall: the unit cost is dropping, but the number of units consumed is rising much faster.

The practical conclusion follows from the same place. The most effective way to cut AI spending is not choosing a cheaper model; it is keeping documents as text at source, caching conversion results, and preventing the same file from being reprocessed on every query.

Put another way: the problem is not the AI, but what we hand it.

Four steps that cut the cost

  • Keep documents as text at source; move to PDF only at the sharing stage.
  • Cache conversion results so the same file is not reprocessed on every query.
  • Treat rendering to images as a last resort; use the text layer directly when one exists.
  • Make spending visible per user — without knowing who consumes what, any limit is a guess.