The problem
Extraction is the easy half.
Pulling fields off an invoice stopped being hard a while ago. Any competent model gets most of the way there, and anyone selling you extraction as the clever part is selling you the commodity.
The work that still takes a trained person is reconciliation. The invoice says 1,200 cartons and the packing list says 1,200 pieces. The consignee is a cold store rather than the buyer. The certificate was issued three days after the vessel sailed. Each of those is defensible once explained and expensive once missed, and none of them is visible in any single document.
So this is built the other way round. Extraction is plumbing. The product is the set of cross-checks between documents, the judgement about which mismatches actually matter, and the queue that puts the rest in front of a human with the reason already attached.
- Cross-document reconciliation, not field extraction
- Exceptions surfaced with the reason, not a confidence score
- The checks are the product; the model underneath is replaceable
How it works
Send a pack. Get back what disagrees.
By API, by watched mailbox, or through a review queue your team works in. Same checks either way.
- 01
Submit the pack
Invoice, purchase order, transport document, packing list, and any certificates. PDFs, scans or photographs. A pack rather than a file, because a single document cannot be checked against anything.
- 02
Fields are read and normalised
Quantities, weights, currencies, incoterms, dates, parties and goods descriptions, pulled out and put into comparable units so 1,200 cartons and 1,200 pieces are visibly not the same claim.
- 03
Documents are reconciled against each other
Every field that appears in more than one place is compared. Mismatches are classified by whether they give the buyer grounds to dispute, and ranked accordingly.
- 04
You get exceptions, not a report
A verdict per pack, and for each exception the two documents, the two values, and why it matters. Clean packs pass through without anyone looking at them.
What it handles
- 6 doc types
- Invoice, PO, transport document, packing list, certificates, contracts
- Pack level
- Checked as a set, because that is the only way discrepancies appear
- API or queue
- Machine-to-machine, watched mailbox, or a review UI
- 0 facilities
- Finance required to use it
Coverage depends on corridor and document quality. Handwritten annotations and poor scans reduce confidence, and the verdict says so rather than guessing.
What it checks
The reconciliations that actually catch things.
These are the comparisons that find real problems, in roughly the order they find them.
- 01 Quantity and unit of measure
- Invoice against packing list against transport document, normalised first. Pieces against cartons against pallets is the most common miss in the sector.
- 02 Amount against the order
- Invoice total reconciled to purchase order price times quantity actually shipped, so a part shipment invoiced in full is visible immediately.
- 03 Parties and consignee
- Buyer, seller and consignee compared across documents. Goods routed via an agent, warehouse or group entity is flagged as an anomaly to confirm rather than an error.
- 04 Dates and sequence
- Invoice date against shipment date against certificate issue dates, checked for an order of events that is actually possible.
- 05 Goods description drift
- The same goods worded differently on different documents. Both true, neither reconcilable, and the single biggest cause of avoidable delay.
- 06 Incoterms and risk transfer
- Whether risk had passed when the invoice was raised, which decides whether the receivable existed on that date.
Questions
What buyers of this ask.
Can we use this without taking any finance from you?
Yes. That is the point of selling it separately. Nothing about it assumes a facility, and there is no obligation to discuss one.
What accuracy should we expect?
We will run your own historic packs through it and give you the numbers on your documents rather than quote a benchmark. A figure from someone else's corridor, in someone else's document quality, is not information.
Does a person see everything, or only the exceptions?
Your choice. Most teams pass clean packs straight through and review only exceptions, which is where the time saving is. You can also route everything to a reviewer while you build confidence.
Which model does the extraction?
Several, chosen per document type, and it is deliberately not the interesting part. The reconciliation logic sits above whichever model reads the page, so we can change that underneath you without changing your results.
Can we run our own thresholds?
Yes. Which mismatch classes block a pack, which merely annotate it, and the tolerances on quantity and amount are all yours to set.
How does it price?
Per pack, with volume tiers. We would rather quote against your real monthly throughput than publish a rate card that turns out to be wrong for you.
Try it on real paper
Send us twenty packs you have already processed.
Ideally ones where something went wrong. We will show you what it catches, what it misses, and what it flags that your team would have waved through.
- Run against your own documents, not a demo set
- We will tell you what it gets wrong as well as what it gets right
- No facility, no commitment, no sales process attached