ROI Calculator Does It Cost Your Business to Read PDFs with AI?

PDFs were designed for humans, not machines. The format describes where ink goes on a page, not what the content means. A PDF table is a grid of visual cells held together by coordinates. A heading is just bigger text in a particular font. There is no semantic structure that says, "this is a heading", "this is a clause", or "this paragraph is the answer to that question".

To extract meaning, the AI has to do work the format was never built to support; parsing layout, inferring relationships, stitching together columns, recognising tables. That work is expensive in tokens and unreliable in outcome. The same document, read by two different models or on two different days, can produce different structure and different answers. This is the hidden cost behind a lot of "our AI is too expensive" and "our AI hallucinates" complaints. It often is not the model. It is the input.

Markdown is a lightweight plain-text format with simple, explicit cues for structure, # for a heading, - for a bullet, | for table cells. It was built to be readable by humans and parseable by machines without any of the visual overhead of PDF.

For AI, that matters in two ways. First, token efficiency: the same 20-page document that costs 60,000–100,000 tokens as a PDF often costs 4,000–8,000 tokens as Markdown, a 10× to 20× reduction before any other change. Second, semantic clarity: the model is not guessing where the table ends, or which paragraph belongs to which heading. It can focus its capacity on reasoning over the content, not on reconstructing the layout. Cheaper, more reliable, more auditable. That is what makes Markdown the natural input format for production-grade AI.

It helps, but only for a narrow slice of the workload. Prompt caching is effective when the same large block of text is sent into many calls inside a short window, a chatbot grounded in one handbook, for instance. It does not help when the document set is broad, when documents change often, when the cache window expires, or when each call references a different combination of sources. Batch discounts apply only to non-interactive workloads where latency is not a constraint.

Neither addresses the underlying problem: the input itself is bloated and unstructured. Even a perfectly cached PDF call still pays for the bloat on the first call, still produces less reliable output, and still requires more remediation downstream. Preparing the content once gives you the saving on every call, in every workload, with every model, whether caching is available or not.

This is the right question, and it is why the calculator is a directional tool, not the full picture. Input cost is the easiest line to model because it is a clean function of tokens × price. The other lines in the cost-per-defensible-answer equation, retrieval, remediation, and human review, are usually larger, but they are more workload-specific and harder to generalise.

Improving the input format compounds across all of them. Better-structured content makes retrieval sharper because the model has clean chunks to match against. It reduces remediation because outputs cite the right sources and need less factchecking. It cuts human review because reviewers are validating judgement, not janitorial errors. The input saving the calculator shows is the floor. The real return on AI-ready data is felt across the whole equation.

You keep your PDFs; they remain the source of record. What changes is what the AI actually reads. The Progress Data Platform sits between your sources and your AI consumers as a context layer. Semaphore enriches and classifies content semantically. MarkLogic stores it in a queryable, governed form. Orchestration Studio runs the pipelines that prepare each document once and route it wherever it is needed. Corticon enforces the policy rules that decide what is shown to whom.

The first AI workload that uses a document pays the preparation cost. Every workload after that, retrieval, summarisation, agents, audit, reads the prepared version for a fraction of the tokens, with sharper grounding and a clear governance trail. You are not replacing your PDFs. You are stopping every AI workload from re-parsing them.