Processing Scanned Documents with AI: Handwriting, Stamps, and Noise

Scanned documents are never clean

In a perfect world, every document would be a natively digital PDF with a crisp text layer. In reality, organizations still deal with mountains of paper. Signed contracts photocopied in the 1990s. Handwritten inspection forms from field teams. Receipts photographed on a phone. Government filings with rubber stamps over printed text.

These scanned documents carry noise that digital documents don't have: uneven lighting, skewed pages, coffee stains, faded ink, handwritten annotations over printed text, stamps that overlap with important data, creased paper that distorts characters.

Traditional OCR was built for clean printed text. Give Tesseract a high-resolution scan of a laser-printed page, and it performs well. Give it a photograph of a handwritten form with a rubber stamp across the date field, and the output is unusable.

The gap between what OCR handles and what real scanned documents look like is where most document processing projects fail.

The types of noise in real scanned documents

Every document processing team encounters the same categories of noise:

Handwriting. Annotations in margins, filled-in form fields, signatures, hand-corrected numbers. Handwriting varies enormously between individuals. A "7" that one person writes looks like a "1" to another.

Stamps and seals. Rubber stamps, notary seals, official stamps overlapping printed text. Red ink stamps over black printed text create an OCR nightmare -- the engine can't separate the layers.

Physical damage. Creases, tears, water damage, staple holes through text, tape marks, yellowed paper. Any of these can obscure critical characters.

Scanning artifacts. Skewed pages, uneven exposure, shadows from thick bindings, bleed-through from the reverse side, moire patterns from rescanning printed halftones.

Mixed content. A typed form with handwritten entries. A printed contract with a stamped date and a handwritten signature. A spreadsheet printout with manual corrections in pen.

Low resolution. Faxed documents, photocopied copies of copies, thumbnail-resolution exports from legacy systems. Characters blur together.

Traditional OCR handles each of these poorly in isolation and catastrophically in combination. A faxed copy of a handwritten form with a stamp is functionally unreadable to conventional OCR.

How AI processes noisy documents

AI document understanding approaches scanned documents differently from OCR. Instead of character-by-character recognition, it processes the entire document image as a visual scene and interprets the content.

Noise filtering. The model distinguishes between document content and noise. A coffee stain is recognized as non-content. A crease is identified as a physical artifact. The model focuses on the actual text and images while ignoring damage.

Layer separation. When a stamp overlaps with printed text, the model can often separate the two layers visually. It recognizes the stamp as a separate element and reads the underlying text independently.

Handwriting interpretation. Modern multimodal models handle handwriting significantly better than OCR engines. They use context (a number in a "Total" field is likely numerical, a name in a "Name" field is likely alphabetic) to resolve ambiguous characters. A "7" that looks like a "1" gets resolved by the fact that it appears in a column of dollar amounts where "$7,500" makes sense and "$1,500" doesn't.

Skew correction. The model reads skewed text without needing explicit deskewing. Whether the text is rotated 2 degrees or 15 degrees, the model adjusts its reading. This eliminates a preprocessing step that frequently introduces its own errors.

Context completion. When a character is partially obscured -- by damage, by a stamp, by a fold -- the model uses surrounding characters and semantic context to complete it. "Am_unt: $5,000" is clearly "Amount: $5,000" in a financial context.

Practical scenarios

Scenario: Box of old contracts.

A law firm digitized 200 contracts from the 1990s and 2000s. Some are clear photocopies. Others are faded faxes. Many have handwritten dates and signatures. Some have correction fluid over original text with new text written by hand.

With traditional OCR, the firm would need to:

1. Pre-process each scan (deskew, enhance contrast, binarize)
2. Run OCR and get heavily error-prone text
3. Manually review and correct the OCR output for every document
4. Extract the actual data (parties, dates, terms) manually

With docrew, the process is:

1. Point the agent at the folder of scanned PDFs
2. Request extraction of parties, effective dates, termination dates, and key terms
3. The agent processes each document, handling noise automatically
4. Review the output, focusing on flagged low-confidence extractions

The agent doesn't need preprocessing. It reads each document as-is, interprets the content despite noise, and extracts the requested data. Documents with severe damage get flagged for manual review rather than producing silently incorrect output.

Scenario: Field inspection forms.

A construction company collects handwritten inspection forms from job sites. Inspectors fill in checkboxes, write measurements, note deficiencies, and sign the form. The forms are photographed on-site and uploaded for processing.

The handwriting varies from inspector to inspector. Some write neatly; others don't. The photos sometimes have shadows, reflections, or partial framing.

docrew's agent reads each form, interprets the handwritten entries in context (a measurement field contains numbers with units, a deficiency field contains text descriptions), and produces a structured report per inspection.

Scenario: Historical financial records.

An accounting firm needs to digitize a decade of paper-based records: receipts, bank statements, tax forms. The documents span multiple formats, multiple levels of degradation, and include both printed and handwritten content.

Batch processing with docrew handles the variety. The agent classifies each document (receipt, statement, form), extracts the relevant fields based on document type, and produces a consolidated dataset. Damaged or unreadable documents are flagged rather than silently dropped.

Handwriting recognition in depth

Handwriting is the hardest category of scanned document content. It varies between individuals, within a single individual (the first page of a form is neater than the fifth), and by context (a hurried signature vs. a carefully printed address).

AI models approach handwriting with two advantages over OCR:

Contextual priming. The model knows what kind of content to expect in each field. A date field should contain a date. An amount field should contain a number. A name field should contain a name. This prior expectation dramatically improves accuracy on ambiguous characters.

Whole-word recognition. Instead of recognizing individual characters and assembling them into words, the model often recognizes entire words or phrases as units. This mirrors how humans read handwriting -- you don't decode letter by letter; you recognize the word shape.

For common handwritten content (numbers, dates, short text), AI accuracy significantly exceeds OCR accuracy. For extended handwritten text (paragraphs of notes), accuracy is lower but still serviceable for extraction of key information.

docrew handles handwriting by reading it in context. Tell the agent "extract the handwritten notes from the inspection form" and it interprets them as a human reader would -- inferring unclear characters from surrounding words and field context.

Stamps, seals, and overlapping elements

Official documents often carry stamps, seals, and markings that overlap with the primary text. A notary stamp across a signature line. A "RECEIVED" stamp over a date. An official seal overlapping the document header.

These overlapping elements create two problems: they obscure the underlying text, and OCR often reads the stamp text instead of (or mixed with) the document text.

AI models handle this through visual layer separation. The model recognizes that the red stamp text is a separate element from the black printed text underneath. It reads both independently and reports them separately.

In docrew, you can request: "Extract the contract date, ignoring the received stamp." The agent understands that the stamp is an overlay and reads the date from the underlying text, even if the stamp partially obscures it.

Quality thresholds and when to flag

Not every scanned document is recoverable. Some are too damaged, too low-resolution, or too degraded for any system -- human or AI -- to read reliably.

docrew's agent uses confidence assessment to handle this gracefully:

High confidence. Clear text, minimal noise. The agent extracts data normally.

Medium confidence. Some noise or ambiguity. The agent extracts data but flags specific fields with lower certainty.

Low confidence. Significant damage or illegibility. The agent reports what it can read and flags the document for manual review.

This tiered approach means you process hundreds of documents automatically, review dozens of flagged items, and manually handle only the few that are genuinely unreadable. The ratio depends on scan quality, but for typical business document collections, 80-90% process at high confidence.

From paper archive to searchable data

The end goal of processing scanned documents isn't just reading them -- it's making them useful. A box of scanned contracts is only valuable if you can search, filter, and analyze the content.

docrew turns scanned document collections into structured, searchable data. The agent extracts the information you need, normalizes it into consistent formats, and produces output ready for analysis or database import.

The entire process happens on your device. Your scanned documents -- which often contain the most sensitive information (original signatures, handwritten notes, official stamps) -- never leave your machine. The AI reads them locally, and the structured output stays local.

For organizations sitting on years of paper archives, this is the path from storage cost to data asset. Those boxes of old contracts contain valuable information about terms, obligations, and relationships. Getting that information out -- accurately, at scale, without uploading sensitive documents -- is now a practical operation rather than a theoretical one.