The Scanned-PDF Problem Nobody Warned Your Diligence AI About

Why does your AI diligence tool keep missing clauses a junior associate would flag in five minutes? Nine times out of ten, the culprit is the document itself, not the model. A scanned PDF looks identical to a machine-readable one on your screen, but the software trying to read it sees two completely different files, and that difference is where a lot of AI review is going wrong.

Deal teams have started to notice. Reviewers who trusted a tool-generated summary can discover, late in the process, that indemnity sections or side letters were only partially ingested. The file loaded fine; the words inside it never made it through.

What Exactly Is a Scanned PDF, and Why Does AI Choke on It?

A PDF can carry two very different kinds of content. One is a real text layer, with characters the software can select, copy, and index. The other is a picture of a page, with no underlying text at all. As WebAIM explains, an image-only PDF has a visual layer but no machine-readable structure beneath it, which is why assistive technology (and every AI pipeline built on the same assumptions) can't parse it directly.

Scans, faxed contracts, signature pages photographed on a phone, older SEC filings, board minutes printed and re-scanned for the data room, all of these arrive as images wrapped in a PDF container. To make them queryable, the pipeline has to run optical character recognition first. If that step is skipped, degraded, or silently fails on a page, the AI acts like the page doesn't exist.

How Bad Does the Accuracy Actually Get?

Bad enough to change a diligence conclusion. OCR accuracy isn't a single number. It swings hard based on scan quality, layout, and whether the document was born digital or photographed. An NVIDIA engineering post compares specialized OCR pipelines against newer vision-language models that read the page more like a human would, and the trade-offs are real: vision models handle complex layouts better but cost more and run slower per page.

The failure modes are quiet, which is the real problem. A missed digit in a purchase price allocation, a merged cell in a cap table that gets read as one column, a signature block that swallows the counterparty name, none of it throws an error. It shows up in the summary as absence.

Where Does This Actually Bite in a Deal?

It bites in the places diligence teams already worry about, and a few they don't. The pattern is consistent: the older the document, the more likely it lives as an image, and the more consequential the clause tends to be.

• Legacy contracts. Signed agreements from a decade ago are almost always scans. Change-of-control provisions, exclusivity language, and assignment restrictions often sit in exactly these files.
• Real estate and lending files. Title work, recorded mortgages, and estoppels are frequently image-only. Reviewers rely on them to confirm liens and priority.
• HR and benefits. Executed offer letters, non-competes, and older plan documents get scanned once and never re-keyed. Retention analysis depends on reading them accurately.
• Regulatory correspondence. Letters from agencies still arrive on paper in many jurisdictions. The date stamp and the response deadline both matter, and both can be mangled by weak OCR.

What Deal Teams Are Doing About It

The teams handling this well have stopped treating ingestion as a background chore. They inspect it. They want evidence that every page was read, and they insist on a visible trail back to the source when the AI answers a question.

1. Audit the ingest. Before any review begins, confirm the extraction step produced a text layer for every page. A blank page should be blank because it's blank, not because OCR skipped it.
2. Re-scan the worst offenders. Documents captured below roughly 300 DPI, or with heavy skew and speckle, are worth requesting again from the seller. The cost of a re-scan is trivial next to the cost of a missed clause.
3. Require citation to source. Every AI-generated finding should link back to the specific page it came from, so a reviewer can eyeball the actual image and confirm the extraction is faithful.
4. Reconcile across documents. A contract term that appears in the model but not in the disclosure schedule, or the other way around, is often a scan that half-loaded. Cross-document checks catch what single-document review misses.
5. Keep a human on the exceptions. Low-confidence extractions belong in a queue that a person clears, not in a summary that reads as if everything was fine.

Platforms built for this workflow are starting to bake these habits into the product itself. The recent VDR.ai coverage on barchart.com describes an approach that combines OCR-aware ingestion with agentic review and cross-document reconciliation, so the tool is actively looking for the missing pages rather than pretending they aren't there. Whichever platform a team lands on, the principle is the same: an AI that can't tell you what it didn't read is an AI you can't trust with the deal.