The three components
- Extraction — pulling fields from the document, whatever its layout
- Validation — checking the result against your master data and internal arithmetic
- Routing — confident items proceed, uncertain items reach a person
Most disappointing implementations have the first and neither of the others, which is why they work in a demonstration and not in a month-end.
Validation is where accuracy comes from
- Does the supplier exist in your ledger?
- Do the line totals sum to the stated total?
- Does the tax figure match the rate and the base?
- Is the currency plausible for that supplier?
- Is this invoice number already recorded?
Extraction alone might reach 90%. Extraction with validation reaches the same accuracy but knows which 10% it got wrong, which is the property that makes automation safe.
What accuracy means here
Per field, not per document. A 95% per-field rate across twelve fields means barely half of documents are perfect, which is a very different conversation from “95% accurate”.
We report both, plus the straight-through rate, which is the number that determines how much work actually disappears.
Design for the awkward ones
Multi-page invoices, credit notes, scanned faxes, handwritten annotations, statements masquerading as invoices, and the one supplier who sends a spreadsheet. Every document set has these.
They will not be automated economically, and the system should identify them instantly and route them to a person rather than producing a confident wrong answer.
Typical results
| Document type | Straight through | Notes |
|---|---|---|
| Regular suppliers, consistent layout | 88–95% | The bulk of volume |
| Mixed suppliers, varied layouts | 70–85% | The usual overall figure |
| Scanned or photographed | 55–70% | Image quality dominates |
| Handwritten elements | Below 40% | Route to a person by default |