AI can copy values out of a document and show where each came from. Checking that they’re right, that nothing is missing and that the documents agree stays with a person. So do payments, release of goods and trade classification.
If it reads the code correctly, isn’t that enough?
No. Copying a field is different from knowing the field is right. A plausible classification code can still be the wrong one, and a confident score doesn’t settle it.
1. Capture what’s there. Flag what isn’t.
Value. Shipment reference, item, quantity, unit, weight, date or another agreed field.
Source. The document, its version and the exact place the value came from.
Status. Found, missing, ambiguous or conflicting. Keep calculated values separate from extracted ones.
Practical check: A required field can still be missing. Don’t turn a blank into a fact.
A fictional example: twelve cases, 120 units.
- Packing list
- 12 cases, pack size not stated.
- Invoice
- 120 individual units.
- Check
- They agree only if each case holds 10 units. Verify that; don’t assume it.
Practical check: A tidy table can conceal a unit mismatch.
2. Review the whole document.
Compare key fields and totals with the source. Confirm that no pages, rows or amended versions were missed. Test on blurred scans, split tables and conflicting documents, and decide which fields need review based on what they affect.
Practical check: Extraction doesn’t approve a payment, release goods or determine a trade classification.
3. A first use: prepare a review pack.
Start with approved test documents and a holding table. Link each proposed value to its evidence. Let the responsible operator correct or reject it before any authorized update. Record review time and errors against the current process.
Practical check: Don’t turn faster reading into unchecked action.
Read more: Using AI to Compare Supplier Quotes: Make Them Comparable First · AI Shipment Tracking: A Predicted ETA Is Not Proof of Delivery
The example is fictional. This is process guidance, not customs or tax advice.