Line item data capture is the extraction of the detail from each individual row of a customer order: the product reference, description, quantity, unit of measure and price. It is the part of order capture that deals with what is being bought, as opposed to header information such as the customer, delivery address and PO number.
Header data is usually easy to capture. Line item data is where order automation projects succeed or fail.
What has to be captured per line
| Field | Why it is difficult |
|---|---|
| Product reference | May be the customer’s own code, a barcode, a brand name or a loose description |
| Description | Free text that varies by customer and by order |
| Quantity | Simple to read, easy to misread when columns shift |
| Unit of measure | Often implied rather than written |
| Price | May be absent, out of date or per case rather than per unit |
| Line notes | Easily missed, sometimes important |
Why line items are harder than headers
There are many of them. A header is captured once. A 60-line order means 60 chances to get something wrong.
Formats vary within one document. A single order can mix typed rows, handwritten additions and a note in the margin.
Columns move. Template-based extraction maps fields to positions, so a customer adding a column breaks it.
Customers use their own language. The same product may appear as a supplier code, a brand name, an abbreviation or a description.
Tables break across pages. Rows split over page boundaries are commonly dropped or duplicated.
Context matters. Deciding whether “12” means 12 cases or 12 units often depends on what that customer usually orders.
Capture is not the same as understanding
Reading the text on a line is only the first step. The captured text then has to be resolved into something your ERP can use:
- Extract the raw values from the document
- Match the product reference or description to a SKU
- Normalise the unit of measure to how you sell the item
- Resolve the price against the customer’s agreed pricing
- Validate the quantity against stock and the customer’s usual pattern
Tools that stop after step one leave the remaining work with a rep. The comparison of AI order processing and OCR covers where that gap usually shows up.
Measuring line item capture
- Line-level accuracy: share of lines captured with every field correct
- Document-level accuracy: share of orders where every line is correct
- Match rate: share of lines resolved to a SKU without help
- Review rate: share of lines sent to a person to check
Document-level accuracy is the number that matters operationally. At 98% line accuracy, a 50-line order still has roughly a two in three chance of containing at least one wrong line, so a rep still has to look at it.
How Ordana helps
Ordana captures and resolves line items rather than only extracting them. The Ordana Desk Agent reads orders in any layout with no templates to configure, so a customer changing their purchase order format does not break capture.
For each line, the agent maps the customer’s own codes and descriptions to your SKUs, works out the intended unit of measure using that customer’s ordering history, applies their agreed pricing and checks availability. Where a line is genuinely unclear, it asks the buyer rather than guessing. Ordana reports a 95% decrease in order entry errors for its customers, and the full order, not just the extracted data, is written into the ERP.
Schedule a demo to see line item capture run on your own orders.
Frequently asked questions
What is the difference between header and line item capture?
Header capture takes order-level details such as customer and delivery date. Line item capture takes the detail of each product row.
Why do template-based tools struggle with line items?
They rely on fields appearing in fixed positions, so new customers and layout changes need new templates.
What accuracy level is needed?
High enough that most orders need no review. Line-level accuracy alone can be misleading on long orders.
Can handwritten line items be captured?
Modern AI handles handwriting far better than older OCR, and should flag anything it cannot read with confidence.
← Back to all blogs