Inefficient PDF Workflows Drive Up Enterprise AI Token Costs
Leaked internal audio from Accenture reveals that non-engineering employees are driving up AI costs by using highly inefficient workflows. Specifically, employees are converting PDF documents into images and then into markdown files, which consumes a massive number of tokens. This highlights a major challenge in enterprise AI adoption, where a lack of technical literacy among non-technical staff leads to skyrocketing operational costs. It underscores the need for better training and optimized document processing pipelines as companies scramble to control generative AI spending. The issue was highlighted by Accenture's agentic AI strategy lead, Justice Kwak, and client group lead, Stuart Henderson, who noted that converting PDFs to markdown via images is a major "token chewer." PDFs are notoriously difficult for LLMs to parse directly, leading users to adopt convoluted workarounds that multiply token usage.
## BACKGROUND
Large Language Models (LLMs) charge users based on "tokens," which are basic units of text or image data processed by the model. PDFs are a notoriously difficult format for AI to read because they lack structured layout information, often forcing users to convert them into images for vision-based LLMs or into markdown text. However, processing high-resolution images or performing multi-step conversions through LLMs requires significantly more computational power and tokens than processing raw text.