Classify
Split and name
A 412-page scan becomes 39 typed, ordered, correctly named documents in the right queues.
Smudged invoices, 400-page KYC bundles, faxed carbon copies, handwritten delivery notes, land records from 1974. CItructure reads all of it, works out what it means, and hands the result to something that finishes the job.
| Description | Qty | Rate | Amount |
|---|---|---|---|
| CR sheet 1.2 mm, IS 513 D | 18 coils | 62,400.00 | 11,23,200.00 |
| Galvanised strip 900 mm | 640 kg | 104.00 | 66,560.00 |
| Freight - Pune to Nashik | 1 | 9,200.00 | 9,200.00 |
| IGST @ 18% | 2,15,812.80 |
Anyone can pull a total off an invoice. The hard part is knowing that the handwritten note in the margin contradicts line three, that page 214 of the loan file is the amendment that supersedes page 40, and that this claim references a policy that lapsed in March.
We model the document the way a reviewer reads it: layout, hierarchy, references, and the relationships between values - then bind the result to your schema, not ours.
{
"invoice_no": "HS-2291-B",
"supplier": {
"name": "Harbour Steel & Alloys Pvt. Ltd.",
"gstin": "27AACCH1234F1Z5",
"matched_vendor_id": "V-10442"
},
"lines": [
{ "desc": "CR sheet 1.2 mm",
"qty": 18, "uom": "coil",
"grn_qty": 16,
"variance": -2 }
],
"total": { "amount": 1414772.80,
"currency": "INR" },
"margin_note": {
"text": "2 coils short - hold pmt, RK",
"handwritten": true,
"confidence": 0.74,
"contradicts": "lines[0].qty"
},
"citations": [ page 1 · bbox … ],
"route": "human_review"
}
A 412-page scan becomes 39 typed, ordered, correctly named documents in the right queues.
Contract v4 against v7: what changed, who benefits, what needs legal's eyes.
Grounded assistants that answer with the clause, or say plainly that they don't know.
PII and PHI removed at ingest, before storage, before inference, before anything leaves.
Nobody has trouble with a clean PDF. Production is a faxed carbon copy with a rubber stamp over the invoice number, photographed at an angle on a loading dock in poor light - and then a delivery note somebody wrote by hand.
Each one is an API call on its own if that is all you need. Most customers chain the lot.
OCR, handwriting, layout analysis, table structure, checkbox and signature detection.
Page-level typing, bundle splitting, duplicate detection and automatic file naming.
Your schema, filled. Nested tables, line items, references and relations between values.
Arithmetic checks, master-data lookups, three-way match, policy and threshold rules.
Post it, route it, answer with it, or hand it to an agent that closes the exception.
A new type is normally days, not months - there is no template to draw and no rules to write. This is simply where we have already done the work.
| Document family | Typical difficulty | What makes it hard | Target straight-through rate |
|---|---|---|---|
| Invoices & credit notes | low | Thousands of layouts, multi-page line items, tax variants | 94% |
| Purchase orders & GRNs | low | Matching across three systems that disagree | 96% |
| Bills of lading & delivery notes | medium | Carbon copies, stamps, handwritten quantities | 88% |
| KYC & identity packs | medium | Mixed bundles, photo IDs, regional scripts | 91% |
| Loan & credit files | high | 400+ pages, amendments that supersede earlier terms | 82% |
| Insurance claims | high | Photos, estimates, policy references, handwriting | 79% |
| Contracts & agreements | high | Clause resolution, defined terms, version diffs | review-first |
| Clinical & lab reports | high | Ward notes, abbreviations, PHI that must never leave | review-first |
| Land & municipal records | high | Decades-old scans, regional scripts, faded ink | 71% |
| Customs & trade documents | medium | Multi-language, HS codes, inconsistent forms | 89% |
Straight-through rate = processed with no human touch at the confidence threshold most customers pick. Everything else routes to a reviewer rather than guessing.
This is where document AI usually stops and where the work actually starts. The extraction is the input to an agent that finishes the task - the same agent runtime described on the solutions page, pointed at paperwork.
Invoice, PO and goods receipt reconciled automatically. Short supply raises the debit note, holds payment and emails the vendor - under an approval gate you define.
Claims, KYC and pre-authorisation packets assembled from a pile of scans, cross-checked, and returned with an explicit list of what is missing.
A grounded assistant over everything you have processed. It answers with the page, the clause and the paragraph - or says plainly that it does not know.
We come back in two weeks with a working prototype on your own data, a measured per-field accuracy number, and an honest list of what it still cannot do.