A PDF or a phone-photographed packing slip doesn't arrive as data, it arrives as an image. Here's how extraction actually works, and what has to happen before a field is trusted.
Key takeaways
Reading a PDF or scan combines optical character recognition with layout recognition, mapping the position of text on the page to what it means, so a number sitting next to "Total" is read as a total, not just as a number.
A document type that repeats, the same supplier's PO layout, the same carrier's bill of lading, gets more accurate over time as the system sees more examples of it. A one-off format from a vendor who's never sent you anything before starts from scratch every time.
Every extracted field carries a confidence score, not just an extracted value. Below a set threshold, that field routes to a person for review instead of being written to your system on the strength of a guess.
This is the actual safety mechanism, not an afterthought. It's also the reason a well-built exception queue matters as much as the extraction engine sitting in front of it.
Extracting a supplier name off a document is only useful if it's matched to the existing vendor record in your system, rather than creating a new, slightly misspelled duplicate. The same logic applies to matching a line item to a SKU already in your catalogue.
This is where the extraction has to talk to whatever you already run, Dynamics 365 Business Central, Sage 300, SAP Business One, QuickBooks Online, Procore, Jobber, Salesforce or HubSpot, rather than living in a separate tool nobody checks.
A faxed document or a poorly lit phone photo still gets processed, but more of its fields drop below the confidence threshold. That means more human review on that document, not fewer checks overall, which is the correct outcome.
The goal was never zero review. It's a review queue small enough that a person can clear it in minutes, built around documents that actually need the extra look instead of every document getting the same scrutiny.
A 30-minute call is enough to tell you whether AI pays for itself here.
