What the application actually does
- Accepts documents however they arrive — email attachment, upload, scanner, shared folder
- Classifies the type, because an invoice and a delivery note need different handling
- Extracts the fields you need, each with its own confidence figure
- Validates against your own data: does the supplier exist, does the total match the lines
- Routes: straight through, to review, or to quarantine
- Writes into your system with the source document attached
Accuracy by document type
| Document | Field accuracy | Main difficulty |
|---|---|---|
| Standard supplier invoice | 92–97% | Line items spanning pages |
| Purchase order | 93–97% | Non-standard product codes |
| Delivery note | 90–95% | Handwritten annotations |
| Bank statement | 95%+ | Merchant names |
| Handwritten form | 70–90% | Legibility |
| Photographed document | 80–92% | Angle, lighting, shadow |
Validation beats a better model
The cheapest accuracy improvement is not a stronger model. It is checking the output against data you already hold: does this supplier exist, does this purchase order number match an open order, is this date plausible?
Those rules catch errors extraction cannot detect on its own, and they make the failures intelligible, which matters for trust.
Measure straight-through rate, not field accuracy
Field accuracy is the vanity metric. What matters operationally is the proportion of documents needing no human touch at all, and time per document including review.
A system at 94% field accuracy might have a 60% straight-through rate, because one uncertain field sends the whole document to review. Improving the review design often beats improving the model.
Cost
A single document type with review interface and validation typically £10,000–£25,000. Each additional type is considerably cheaper once the pipeline exists — usually a third to a half of the first.