OCR Errors: When Scanned Contracts Break AI Review

The Decimal Point That Cost $40,000

A claims manager I'll call Dana approved a settlement payout last spring on the strength of an AI summary. The summary said the outstanding medical lien was $1,250. The scan said $17,250. That "7" sat right on the fold line of the page, faded to a gray smudge, and the optical character recognition engine running underneath the summarizer dropped it entirely. The AI didn't invent the error. It inherited it. The audit that caught the gap landed four months later, after the check cleared.

Dana's story isn't unusual, and it isn't really about AI. It's about the layer underneath every document analysis tool: the moment paper becomes machine-readable text. Get that layer wrong and everything downstream, the summary, the obligation map, the risk score, sits on sand.

Most document review failures I've seen trace back to scan quality, not model quality. Here's how the pipeline actually breaks, and what to do before it breaks something of yours.

What Happens to a Page After You Hit Scan

OCR doesn't read. It converts pixel patterns into character codes using statistical models trained mostly on clean, printed, well-lit text. When a page is skewed, faded, or speckled with scanner noise, the engine guesses and moves on, silently, with no red pen in the margin.

Tesseract, the open-source engine embedded inside a surprising number of commercial products, exposes a confidence score for every word it produces. Almost nobody looks at those scores. Your summarization tool almost certainly doesn't surface them to you either.

So here's the mechanism: a document gets scanned at 150 DPI instead of 300. Compression smears thin strokes. The letterforms in "not," "no," and "unless" are among the thinnest in any typeface, which is why negations are the first casualties. A clause reading "Vendor shall not be liable for consequential damages" can emerge as "Vendor shall be liable for consequential damages", a total inversion of meaning, delivered in perfect grammar, with a confident summary on top.

Clean, high-resolution print can push OCR accuracy past 99%. Drop the resolution, photocopy a fax, or shoot a photo at an angle, and that number falls off a cliff. The tool won't tell you which page you're on.

The Five Ways Scans Break Document Analysis

The failures cluster into five predictable categories, and four of them involve numbers or negation, the exact things people most need to get right. You can hunt for all five in an afternoon.

1. Currency and decimal loss. Dropped digits, missing decimal points, and commas that vanish under scanner artifacts. "1,250.00" and "1.250.00" are very different numbers, and an OCR engine with a European locale model may flip them.
2. Negation destruction. The thin strokes in "not," "nor," "except," and "unless" erode first. The research brief behind this piece flags exactly this problem: a clause saying "no liability" may carry exceptions elsewhere, and "unless," "except," and "subject to" can reverse the practical result of the whole provision.
3. Digit confusion in dates. 3 and 8, 5 and 6, 1 and 7. A notice deadline of August 3rd becomes August 8th. One of those keeps you in compliance and one doesn't.
4. Table flattening. Multi-column schedules, fee grids, and payment tables scramble when OCR reads left-to-right across merged cells. Rows migrate. The total line lands in the wrong column.
5. Cross-reference corruption. "Section 4.2(b)" becomes "Section 4.2(6)." Now your analysis points at a clause that doesn't exist, and the summarizer will happily describe what it thinks lives there.

Any one of those can survive a dozen human skim-reads. Would you catch a missing decimal in a page of dense figures at 11 p.m.?

"Searchable PDF" Doesn't Mean Accurate PDF

The phrase "searchable PDF" describes one thing only: an invisible text layer has been placed behind the page image. It says nothing about whether that text is correct, a searchable PDF can carry a 40% character error rate and still pass the description.

This is the trap. Scanning software brags about searchability. Review tools accept searchable PDFs without complaint. And the failure mode is invisible, because the readable page looks perfect. You're looking at the image while the machine reads the guess underneath it.

Here's a habit worth building. When a tool extracts text from a scan, ask to see the extracted characters next to the page image for any passage that carries money, dates, or negation. It takes thirty seconds. If a vendor can't show you that side-by-side, that tells you something about how much they've thought about this problem.

A clean-looking PDF is not evidence that the text layer is clean. The image is the original. The text layer is a translation, and translations drift.

Five Checks Before You Trust Extracted Text

Run these in order. The whole sequence takes under ten minutes on a typical 40-page agreement, and it catches the overwhelming majority of scan-induced errors.

1. Page count and page order. A 42-page contract that extracts to 41 pages has a problem somewhere. Duplex scans love to drop a back page, and stapled exhibits love to reorder themselves.
2. Visual diff on high-risk passages. Compare the extracted text against the scan image for every dollar figure, every date, every negation, and every party name. Not the whole document, just those. They're the passages that decide outcomes.
3. Table inspection. Look at fee schedules, payment grids, and delivery tables side by side with the image. Check the totals against the rows.
4. Handwriting, stamps, and initials. OCR either ignores these or mangles them. A margin note reading "N/A per phone call 6/12" or a struck-through clause can change the deal. Research on document review guidance keeps returning to this point: manual verification of high-risk passages is the control that matters.
5. Lowest-confidence sweep. If your tool exposes confidence scores, sort ascending and review the bottom few percent. That's where your errors are hiding, and nobody reads the bottom of the list.

The rule underneath all five: the higher the legal, financial, or privacy stakes, the more the machine's output needs a human signature on it.

Why You Shouldn't Scan the Whole Filing Cabinet

There's a second cost to bad scans, and it's not accuracy, it's exposure. Blurry scans of documents stuffed with Social Security numbers, account numbers, and medical details create a permanent liability every time they get uploaded to a service.

Legal-tech guidance on AI document review consistently lands on the same controls: minimize what you upload, redact personal information before it leaves your machine, check where data is stored and processed, and confirm in writing that your submissions aren't used for model training by default. That's data minimization in practice, and it starts at the scanner.

So think about what you actually need. If you're reviewing a lease to find the renewal terms, you don't need the page with the landlord's wire instructions and your voided check stapled to it. Scan the pages in scope. Redact the fields that have no bearing on the question you're asking. Keep the originals in your own records, where they belong.

Contracts that matter can sit in a file for years. Per the IRS, some supporting records should be kept three years, and employment tax records four. Fine, keep them. Just don't broadcast them into every tool that touches your workflow.

The Scan Quality Gate: A Policy You Can Paste Into a Wiki

Most teams have no standard for how documents enter the review pipeline. They fix that in about fifteen minutes with a short written gate. Here's one that works, drawn from what actually goes wrong.

Minimum settings. 300 DPI or higher for anything with figures. Grayscale, not color, for text pages. Despeckle and auto-straighten on. No phone photos of pages unless there's genuinely no alternative, and if there isn't, shoot straight down with the page flat and the shadow off the fold.

Pre-flight checks. Confirm page count against the physical original. Confirm the text layer exists. Spot-check two dollar figures and two dates against the image before you upload anything.

What gets flagged for manual verification. Any document that is a photocopy of a photocopy. Any page with a fold line, staple shadow, or highlight over text. Any schedule or table. Any page with handwriting, stamps, or initials.

What never gets uploaded without redaction. Account numbers, government identifiers, medical details, and anything covered by a confidentiality obligation to a third party. Redaction standards for accessibility and document handling are widely published, the W3C's WCAG guidance covers text-in-images concerns, and the principle carries over: an image of text is a lossy substitute for text.

Required sign-off. The person who relies on an extracted figure is the person who verifies it against the source page. No exceptions, no delegation to the tool.

A quality gate is cheaper than a post-mortem. Every hour you spend on scan hygiene buys back a week of explaining how a number changed.

Your Scan Becomes the Record

Here's the part people miss. The scan you make today becomes the artifact you'll rely on in two years, in a dispute, an audit, a renewal negotiation, or an insurance claim. When the paper original is in a box in a storage unit you no longer visit, the PDF is the document.

That has a consequence. If your archived copy has a missing decimal, an inverted negation, or a scrambled schedule, you didn't just introduce a review error. You destroyed the audit trail back to the truth. Legal AI governance conversations keep circling this question of provenance, where a fact came from and whether you can trace it. Provenance starts with a legible scan, not a language model.

The tooling will keep improving. Summarizers will get better at flagging exceptions, at citing page numbers, at labeling uncertainty, at telling you when they don't know. But none of that helps if the characters underneath are wrong.

So the next time you feed a stack of paper into a contract review workflow, spend ninety seconds on the scan itself. Look at the fold lines. Check the totals. Ask where the text layer came from. The most expensive mistakes in this business rarely look like mistakes. They look like clean, confident summaries of documents nobody ever actually read.