Structured data for AI explained: why LLMs waste tokens on messy formats, why Markdown wins, and a playbook to make your files AI ready.
Every company that rolled out AI in the last two years is hitting the same wall, and it is not the model. It is the files. Large language models reason over text. Give them clean, well-structured text and they answer with precision. Give them a scanned PDF, a spreadsheet exported to HTML or a slide deck made of floating text boxes, and they burn tokens reconstructing what your document should have said in the first place, then fill the gaps with guesses. That is what structured data for AI means in practice. Not a data science luxury: the difference between an assistant that quotes your price list correctly and one that invents a discount. And it is measurable. The same content costs about 16 percent more tokens serialized as JSON than as Markdown, and flat tables cost roughly 61 percent less as CSV than as formatted JSON. Multiply that by…