Document extraction software works great until it doesn’t.
It handles clean, digital invoices just fine. But the moment a scanned receipt, fax, or phone photo enters the mix, things fall apart. The results are noisy, inconsistent, or flat-out wrong, and the manual work creeps back in.
Most extraction tools assume an ideal world where every document is a pristine PDF. In practice, that world doesn’t exist.
Talk to any operations team processing documents at scale and you'll hear some version of the same story. A trucking company we work with had six full-time employees doing nothing but manual data entry. The documents? Handwritten driver tickets, photographed bills of lading, carrier invoices that had been scanned and faxed multiple times. Their previous extraction tool couldn't touch any of it.
A CPA firm doing 3,500 compliance audits a year told us their biggest headache was scanned payroll documents. "They don't convert very well with other systems," as one of their accountants put it. They'd tried multiple tools before finding Lido. Same result every time.
A restaurant group processes invoices from local vendors that are often handwritten, sometimes in Vietnamese. A propane distributor deals with suppliers who, as their ops lead described it, "just like to handwrite everything." A healthcare company handles 200,000 documents per month, including handwritten doctor's orders and requisition forms.
None of these are edge cases. This is what document workflows actually look like. And in every one of these scenarios, their previous extraction tool that was supposed to eliminate manual work ends up creating more instead.
Traditional OCR was built to convert printed text to digital text. It works by recognizing character shapes and matching them to known patterns. This is fine on a high-quality scan of a typed document. It falls apart quickly when you introduce any of the following:
The result is that teams end up with two workflows: one for clean digital documents that the tool can handle, and a manual process for everything else. As volume grows, "everything else" grows faster.
Even if your tool handles basic scans, template-based data extraction creates a second failure point. Every new document layout — every vendor, every format variation — needs its own template or model. Scanned documents from different sources look different. A scanned invoice from Vendor A doesn't match the template you built for Vendor B, even if they contain the same fields.
One sales engineer we work with put it bluntly:
"Lido has extracted things that I cannot read."
That's the bar. If a human can squint at a blurry scan and figure out what it says, the tool should be able to do the same. Most can't.
Solving real-world document processing at scale requires a different approach to understanding documents.
First, the tool needs to interpret documents visually, not just run character recognition. Modern AI vision models can understand a document the way a human would — recognizing structure, context, and meaning rather than just matching character shapes.
Second, the tool needs to handle layout variability without templates. If every new scan requires manual configuration, you'll never keep up with the variety of inputs your business actually receives.
Third, handwriting support needs to work in practice, not just in demos. Real handwriting is messy, rushed, and abbreviated. Sometimes it's in a language other than English. We've successfully extracted handwritten Vietnamese invoices. If your tool can't handle that, it can't handle your real workflow.
Fourth, iteration needs to be free. Scanned documents are inherently variable. Getting extraction right sometimes takes a few passes. Tools that charge per attempt — including failed attempts — are penalizing you for their own limitations.
If scanned or handwritten documents are part of your workflow, don't let the vendor demo on clean samples. Bring your worst documents.
Upload the receipt that's been forwarded three times. The handwritten ticket photographed in bad lighting. The faxed invoice from 2019. If the tool can't handle your hard cases, it won't help you actually save time in the long run.
Test handwriting explicitly. Ask the vendor to extract from a handwritten document in your actual workflow. If they hedge or mention "custom training," that's a red flag.
Ask what happens when extraction isn't perfect. Do you pay again to iterate? Tools that charge for failed attempts are optimized for clean inputs, not real-world documents.
Check processing time on scans. Some tools slow down dramatically on image-heavy inputs. At volume, that creates a bottleneck.
Lido uses a custom blend of AI vision models, OCR, and LLMs to extract data from any document — including scanned PDFs, phone photos, faxes, and handwriting. No templates, no model training, no special configuration for messy inputs.
Disney Trucking freed six employees from manual data entry on handwritten driver tickets. YMI Jeans processes sales orders that are half handwritten. Kei Concepts extracts handwritten Vietnamese invoices with complex tax calculations. These aren't special configurations — they're the same tool, working on real documents.
If your current tool can't handle your scanned documents, the problem isn't your documents.
%20(1).svg)
