Document intelligence
Extraction, classification and validation of the documents your clients send you, with a human review queue that your own people work.
What it does
Every document that reaches your delivery floor is classified, read and checked before it lands in the system of record. Each field carries a confidence figure. Fields the machine is unsure of, fields that fail a rule and a random sample go to a person before delivery.
The reviewer is yours. They see the source region beside the extracted value, approve or correct it, and the decision is written to the audit log with their identity. Corrections feed the next evaluation run.
Nothing here trains a model for anyone else. The evaluation set, the prompts, the rules and the extraction service transfer to you at acceptance.
Processes we have scoped this line for
- Invoice intake and validation against purchase orders and receipts
- Claim forms and attachments: classification, completeness check, triage
- KYC packs: document type, field extraction, mismatch flags against the application
- Loan files: page splitting, classification, a conditions checklist
- Remittance advice: extraction, denial categorisation, work-queue routing
- Freight invoices and bills of lading: charge fields matched to rate cards or bookings
Section 2 — what it does
Yellow means the machine read it. Red means a person touched it. A hairline means one system ends and another begins.
Healthcare documents in the United States usually need a business associate agreement between your client and you. Lawyer review before scoping.
What a reviewed document looks like
Section 3 — a reviewed document
The end state of the sequence on the home page. Thirteen fields read, one corrected by a person, delivered.
Meridian Freight Pvt Ltd0.98
Plot 14, Bhiwandi Logistics Park, Thane 421302
GSTIN 27AAECM4821K1Z5
Bill to
Halcyon Retail Group, accounts payable, Pune
| Description | Qty | Unit | Amount |
|---|---|---|---|
| Container handling, 20ft | 120.99 | 850.00 | 10,200.000.98 |
| Customs documentation | 30.99 | 420.00 | 1,260.000.96 |
| Freight surcharge, fuel | 1 10AK | 1,240.00 | 12,400.000.96 |
Amounts in USD. Synthetic document.
Human review
Line 3 quantity read as 1 0.61. Does not reconcile with the line amount.
AK reviewing
Corrected to 10. Reconciles.
Your system
SAP, accounts payable
13 fields delivered. 1 corrected by a person.
How it runs
| Step | Stage | Who | What happens |
|---|---|---|---|
| 1 | Intake | System | Documents arrive from mail, scan, portal or API, in the format your client already sends. |
| 2 | Classification | Machine | Document type and page splits, with a confidence figure per page. |
| 3 | Extraction | Machine | Header and line fields, each with a confidence figure and the source region. |
| 4 | Validation | Machine | Arithmetic, format and reference checks against purchase orders, rate cards or applications. |
| 5 | Review queue | Person | Low-confidence fields, failed checks and a random sample, worked by your reviewers. |
| 6 | Delivery | System | Structured output to your ERP, claims platform or ticketing system over the interface it already has. |
| 7 | Audit log | System | Every extraction, correction and export attributable to a person or a system. Extract available on request. |
Section 4 — how it runs
Steps 2 to 4 are the machine. Step 5 is a person. Steps 1, 6 and 7 are system boundaries.
What the pilot builds
One process, two weeks, a fixed fee. Acceptance criteria are agreed before day one. Your reviewers work the queue. The report, the system and the code transfer to you.
| Process | What the pilot builds | What acceptance measures |
|---|---|---|
| Invoice intake and validation | Extraction of header and line fields; validation against purchase orders and receipts; exception routing; review queue | Field accuracy and exception rate on your own sample |
| Claim form and attachment intake | Classification of claim documents; extraction of claim fields; completeness check; triage to the right queue | Classification accuracy, completeness-check accuracy, exception rate |
| Identity and address document check | Document type classification; field extraction; mismatch flags against the application; review queue | Field accuracy and mismatch-flag precision |
| Loan file indexing and conditions | Page splitting; document classification; extraction into a conditions checklist | Classification accuracy, checklist completeness |
| Freight invoice audit | Extraction of shipment and charge fields; matching against rate cards or bookings; discrepancy flags | Field accuracy, match rate, exception rate |
Section 5 — the pilot
Targets for every measure are set in the pilot scope from your own sample. placeholder — replace with the agreed floor.
The pilot fee is fixed and quoted in the scope. placeholder — replace with the pricing model.
What you receive
- A pilot report: field accuracy and exception rate on your sample, with the evaluation set that produced them
- The extraction service, its prompts, rules and schema, pinned to a version
- The review queue your reviewers used, with the audit log
- Transfer of all of it at acceptance, and the operations retainer if you want it run
- We never contact the BPO's client.
- The BPO sets the price.
- Everything we build transfers to the BPO.
Section 6 — what you receive
Everything we build transfers to the BPO. Policy 09 gates every model change on your evaluation set.
We do not contact your clients. Ever.