OCR vs invoice matching: extracting data is only the first step

OCR reads text from an image. Invoice extraction organizes that text into fields and rows. Invoice matching then connects those rows to the relevant order and compares their values. You can extract every number correctly and still get the comparison wrong by pairing unrelated items.

Some products combine these tasks. When evaluating one, check what happens after the table appears: can you trace each match, inspect the original values and resolve a difference?

A correct extraction with the wrong matches

Consider this fictional order for two items. Both documents use USD, eaches and prices before tax. No discounts or charges apply.

PO row	Item	Quantity	Unit price	Amount
10	R-10, small bracket	100 each	$8.00	$800.00
20	S-20, large bracket	50 each	$12.00	$600.00

The invoice lists the same items in the opposite order, with changed prices:

Invoice row	Item	Quantity	Unit price	Amount
1	S-20, large bracket	50 each	$11.20	$560.00
2	R-10, small bracket	100 each	$8.40	$840.00

Both totals are $1,400.00. Assume extraction returned exactly the rows above. Matching by row position would compare the small bracket with the large bracket and produce false quantity and price differences.

Pairing by the documented item codes within this order gives the useful comparison:

Match	Calculation	Difference to review
PO row 10 → invoice row 2, R-10	100 × ($8.40 − $8.00)	+$40.00
PO row 20 → invoice row 1, S-20	50 × ($11.20 − $12.00)	−$40.00
Overall amount	+$40.00 − $40.00	$0.00

The equal totals conceal two price changes. The documents do not establish whether either change was accepted. Check the relevant agreement or amendment before deciding how to handle them.

Download the example and review worksheet. It includes the source rows and a blank record for your own check.

Check the work at each stage

Stage	What to inspect	A failure to look for
Reading and extraction	Every page, item, quantity, unit and charge against the original	A faint decimal point is lost, or the last row is omitted
Document grouping	Supplier, buyer entity and order reference	Two different orders happen to use the same item code
Line matching	Item code, description, specifications and supported relationships	Reordered rows are paired by position
Comparison	Quantity, unit, price, currency and calculation basis	A case price is compared directly with an each price
Review	Original values, exact differences and the evidence for a decision	An accepted difference disappears from the report

An OCR confidence score, where a tool provides one, concerns its own recognition process. It does not prove that the item was paired with the right PO line or that a price change was authorized.

Handle cases that a simple lookup cannot settle

If supplier and buyer codes differ, inspect the specifications and any documented cross-reference. Keep the pairing unresolved when the evidence is insufficient. The different-SKU guide has a worked mapping example.

A repeated SKU may represent split delivery lines. Retain the individual rows before grouping their quantities and amounts; matching one row and dropping the rest would understate billing. Follow the split-line example for that case.

When the units differ, establish the pack size from a source record before calculating equivalence. Four cases can equal 48 each when the record says 12 each per case. A similar description alone does not establish that conversion. See the case-versus-each guide.

Decide what you need from the tool

If your task is to copy invoice fields into a spreadsheet, extraction may cover most of it. For comparisons across documents, test row matching and difference calculations as separate requirements. For a team handoff, also check whether another person can recover the evidence and distinguish open questions from completed decisions.

Use permitted examples with reordered rows, a missing item and a unit change. Save the first result before correcting it. A tidy output table is useful only if its rows account for the source documents.

LineRecon’s workflow includes reviewing the proposed comparison plan, inspecting differences with source evidence and recording human decisions. Its document-reconciliation example shows prepared fictional findings. That example is not an OCR accuracy benchmark or a promise that every supplier layout will be read correctly.

Common questions

Does a searchable PDF need OCR?

It may already contain a usable text layer. Check whether the text and table structure can be extracted accurately. Scanned pages and mixed image/text files need separate attention; copying visible text from one page does not establish that every page was read.

Is invoice matching the same as three-way matching?

Three-way matching includes receiving evidence as well as the PO and invoice. Microsoft’s invoice-matching documentation describes quantity checks against product receipts. A PO and invoice alone cannot establish what arrived.

Can ChatGPT perform these checks?

You can ask it to compare supplied documents, then verify the result against their sources. Our ChatGPT invoice-matching exercise includes a prompt and reference answer. Check the input format and the scope of any test before applying its result to your own files.