Scanned Invoice Data Extraction: Which Tools Work with Bad Scans
How to extract data from scanned invoices when the scan quality is poor. Covers what causes extraction failures, which tools handle bad scans, and practical tips for improving results.
Why Scanned Invoices Are Still a Problem in 2026
Despite the shift toward digital invoicing, scanned paper invoices are not going away. Construction companies, medical offices, government agencies, and many small businesses still generate paper invoices. Some vendors fax invoices (yes, in 2026). Others hand-deliver paper copies.
As a bookkeeper, you receive whatever your clients' vendors send. And what they send is often a scanned PDF that looks like it was photographed through a screen door.
The challenge of extracting data from scanned invoices is fundamentally different from processing clean digital PDFs. A digital PDF contains embedded text that software can read directly. A scanned invoice is an image, and the software has to figure out what the text says by analyzing pixel patterns. When the scan is clean, this works well. When the scan is crooked, blurry, or low-resolution, everything gets harder.
What Makes a "Bad" Scan
Not all scan quality issues are equal. Here are the most common problems, ranked by how much they affect extraction accuracy:
Severe: Low Resolution
Scans below 150 DPI (dots per inch) lack enough detail for reliable character recognition. Letters blur together, numbers become ambiguous (is that a 3 or an 8?), and thin lines in tables disappear. Most extraction tools need at least 200 DPI to work reliably. 300 DPI is ideal.
Moderate: Skew and Rotation
When a page is fed through a scanner at an angle, the resulting image is skewed. A 5-degree skew is manageable for most tools. Beyond 10 degrees, text detection starts to fail because the software cannot reliably identify text lines.
Moderate: Noise and Artifacts
Coffee stains, pen marks, staple shadows, scanner bed dust. These artifacts add visual noise that can confuse character recognition. A dark smudge near a dollar amount might be read as an extra digit. A line through text might cause the tool to skip that section entirely.
Mild: Uneven Lighting
Photographed invoices (taken with a phone camera instead of a scanner) often have uneven lighting. One corner is bright, the other is in shadow. This affects contrast and makes text harder to read in darker areas.
Mild: Faded Text
Thermal paper receipts and old dot-matrix printouts fade over time. Light gray text on white paper is difficult for OCR to distinguish from the background.
How Different Tools Handle Bad Scans
Basic OCR (Tesseract, Adobe Acrobat)
Open-source OCR engines like Tesseract and built-in PDF OCR in Adobe Acrobat work well on clean scans. On bad scans, accuracy drops sharply. Tesseract has no built-in preprocessing for skew correction or noise removal. Adobe Acrobat adds some preprocessing, but it is designed for general document OCR, not specifically for invoices.
These tools extract raw text but do not understand invoice structure. You get a block of text with all the characters the OCR could read, but no structured fields. You still have to find and extract the vendor name, invoice number, and line items yourself.
Template-Based Extraction Tools
Dext, AutoEntry, and similar tools combine OCR with template matching. They know where to look for specific fields based on vendor layout templates. On clean scans, this works well. On bad scans, the combination of OCR errors and positional drift (fields shifting due to skew or different page sizes) compounds the problem.
If the OCR reads "lnvoice #" instead of "Invoice #" due to a blurry scan, and the template is looking for the exact text "Invoice #", the field gets missed.
AI-Powered Extraction
Tools that use large language models (LLMs) for extraction handle bad scans better than template-based approaches, and here is why: the AI understands context. If the OCR layer reads "T0tal: $1,234.5O" (with zeros instead of O's), the AI still recognizes this as a total amount of $1,234.50 because it understands what a total looks like in the context of an invoice.
This contextual understanding is the key advantage when trying to extract data from scanned invoices with quality issues. The AI compensates for OCR errors by using surrounding context to infer the correct values.
SkipEntry uses AI-powered extraction that processes scanned invoices alongside digital PDFs. The same upload workflow handles both.
Practical Tips for Better Extraction from Scans
Even with AI-powered tools, you can improve results by addressing scan quality at the source.
For Your Clients
Set scanner defaults to 300 DPI. Most office scanners default to 150 or 200 DPI to save storage. Bump it to 300. The file size increase is minimal, and extraction accuracy improves significantly.
Use a flatbed scanner for wrinkled or folded invoices. Auto-document feeders work great for flat pages but struggle with folded, stapled, or wrinkled documents. A quick flatbed scan produces a much better image.
Scan in grayscale, not black and white. Black-and-white (1-bit) scanning forces every pixel to be either black or white. This destroys subtle details in faded text or lightly printed numbers. Grayscale preserves shading information that helps OCR accuracy.
For Phone Photography
If clients photograph invoices with their phone instead of scanning:
Use a document scanning app (Microsoft Lens, Adobe Scan, or similar). These apps automatically crop, deskew, and enhance contrast. The result is dramatically better than a raw phone photo.
Photograph in good, even lighting. Avoid shadows across the page. Natural daylight near a window works well. Overhead office lighting works too, as long as you are not casting a shadow on the invoice.
Hold the phone directly above the page, not at an angle. Perspective distortion makes extraction harder.
When Extraction Fails: A Recovery Workflow
No tool works perfectly on every scanned invoice. Here is a practical workflow for handling extraction failures:
Step 1: Check the confidence flags. Good extraction tools flag low-confidence fields. If the tool is not sure about a value, it should tell you. Review flagged fields first.
Step 2: Compare against the source. Open the original scan alongside the extracted data. Verify the flagged fields and spot-check a few others.
Step 3: Manual correction. Fix any errors in the extracted data. This is still faster than manual entry from scratch because most fields are correct. You are fixing 2 or 3 fields, not typing 15.
Step 4: If extraction fails completely, fall back to manual entry for that invoice. Some scans are genuinely unreadable. A 50-year-old thermal receipt that has faded to near-blank is not going to work with any tool. Enter it manually and move on.
Step 5: Flag the vendor. If a specific vendor's invoices consistently cause extraction problems, note it. You can either ask the client to request digital invoices from that vendor, or set up a manual entry process for just that vendor.
The Bottom Line on Scanned Invoice Extraction
Extracting data from scanned invoices is harder than processing digital PDFs, but modern AI-powered tools handle it far better than legacy OCR. The combination of improved preprocessing (deskew, noise reduction, contrast enhancement) and contextual AI understanding means that most scanned invoices can be processed automatically.
The practical approach is: use automation for everything, review the results, and manually handle the small percentage that fails. For most bookkeeping practices, that failure rate is under 5 percent of invoices when scan quality is reasonable.
Try SkipEntry with your scanned invoices. Upload a mix of clean digital PDFs and scanned documents. The free tier gives you 50 pages to test with. See how the extraction handles your specific scan quality, and decide from there.