Why Home Health AI struggles with PDFs?

In home health RCM, a missed line item in a clinical note directly causes claim denials. Reliable healthcare AI isn't about flashy demos — it’s about the high-precision engineering happening behind the scenes.

Foundational models are among the most capable AI systems in the world. Yet, it recently took six prompts to get one to find a single word in a PDF.

A quick experiment: I handed a state-of-the-art foundational model a scanned document and asked it to find the word "qualifying." Six tries, each more specific than the last. The prompt that finally worked: "Find this word on page 3 of the PDF, which is printed as page 10." Only then did it locate it.

This isn't a criticism of the model. It's a window into a problem most people badly underestimate.

A PDF isn't a document — it's a set of instructions for a renderer. When you open it, your viewer draws the page you see. The model never gets that page. It gets the raw text stream, stripped of anything visual.

"Page 3" doesn't exist in that stream — page breaks are rendering instructions, not markers. The "10" printed on a page is just the characters "1" and "0" floating somewhere in the text, indistinguishable from any other number. Scanned documents add a layer of OCR noise on top.

A human glances at the page and instantly perceives the frame, the header, the columns, where one report ends and the next begins — visual-spatial reasoning that happens before you read a single word. Text extraction throws all of these cues away and hands the model the residue.

The real lesson: the hard problems in applied AI are rarely the ones that demo well. They're the unglamorous ones — finding the right word, on the right page, at the right coordinates, reliably, on documents far messier than anything in a benchmark.

These are some of the problems we're quietly solving at Serent. Often, the most valuable engineering is the kind nobody sees.