Most invoice OCR APIs hand you text and call it a day. The DocPro Invoice Agent identifies the invoice type, pulls header fields and line items into a schema, validates tax math and PO match, flags duplicates, and returns JSON with a confidence score on every field. Your AP team stops keying. Your auditor stops asking how a number got booked.
Invoices look standard until you actually have to post a thousand of them on a Wednesday. These are the failure modes that send the work back to a human.
Some put the GSTIN in the header. Some hide it next to a stamp. Some attach a credit note that subtracts from page 3. Template-based OCR cracks the first time a vendor redesigns their letterhead.
Line items wrap across pages. Sub-totals appear mid-table. Some rows are descriptions, not products. OCR returns a CSV that looks right and is wrong.
OCR can pull a total of 1,75,584 with high confidence. That does not tell you whether the tax math is right, whether the PO has that line, or whether you have already paid this invoice last quarter.
Header, line items, taxes, references, and footer terms. Each field comes back with a name, a value, and a confidence score. Mouse-over highlights below show what maps to what.
Invoice number, issue date, due date, PO reference, vendor name and address, vendor GSTIN, currency, place of supply. Tagged separately, so an "INV-2026/" prefix doesn't end up in the date column.
SKU, description, HSN/SAC, quantity, unit of measure, rate, discount, line tax, and amount — one object per row. Continuation rows are joined to their parent.
Subtotal, CGST, SGST, IGST, cess, rounding adjustments, and grand total. Each tax with its rate and taxable amount, so reconciliation does not have to back-calculate.
Bank details, payment terms, IRN and QR payload on e-invoices, delivery address, contact, signatures. Captured but optional — your schema decides what travels forward.
The Invoice Agent is the work the finance team would rather not do — keying, matching, chasing — handled in a single API call.
Tax invoices, proforma, credit notes, debit notes, utility bills, freight invoices. New vendors land in the system without a template build.
Multi-page tables, continuation rows, mid-table sub-totals, mixed currencies. The agent maps rows to a clean schema, not a raw CSV.
Invoice to PO, invoice to PO to GRN, plus contract or service confirmation. Mismatches return the reason, not a generic fail.
Cross-checks vendor, invoice number, date, and total against the last 24 months. Catches the same invoice resubmitted with a new scan ID.
Recomputes CGST, SGST, IGST, and cess against subtotal. Validates GSTIN format and checksum. Flags rate-vs-HSN mismatches.
Every field carries a confidence score. Auto-post above your threshold. Send the rest to the review queue with the original document attached.
Low-confidence invoices land in a web UI. Reviewers correct fields side-by-side with the source. Corrections feed back into your tenant's tuning loop.
Pre-built handlers for SAP, Oracle, NetSuite, Tally, and Microsoft Dynamics. Webhooks for the in-house ledger nobody else has heard of.
No black box. At every step you get the data, the rules that fired, and a confidence score you can act on.
POST a PDF, image, or zip to /v1/extract with agent=invoice. Single file, batch, or a webhook-driven drop folder.
The agent identifies the invoice type, locates header and line-item regions, and reads the text. Skew, rotation, and phone photos are handled by the pipeline.
Tax math, PO match, GSTIN format, duplicate check, vendor whitelist. Each rule reports pass, fail, or warning, with the field it touched.
JSON shaped to your schema, with a confidence score per field. High-confidence invoices auto-post. Low-confidence ones route to a reviewer.
The Invoice Agent is a REST API, not a portal your dev team has to wrap. Upload the file, pass the schema name, get back validated JSON. The rules live in your repo, version-controlled like any other contract.
The same API, three different jobs. Pick the one that sounds like your Monday.
Vendor invoices arrive over email, portal, and paper. The agent reads, matches, and posts the safe ones to the ERP. Reviewers only see the exceptions.
Pull 12 months of invoices into a clean dataset for GST reconciliation, ITC claims, or year-end audits. Every field is traceable to its source page.
Procurement, expense, and ERP vendors embed the Invoice Agent under their own UI. White-label, multi-tenant, per-tenant tuning when needed.
Invoices carry vendor PII, tax IDs, and bank details. The defaults reflect that, not your trust in us.
Controls mapped to ISO 27001 Annex A; ISO 9001 quality regime in place. Audit artefacts available under NDA.
Pick the region on AWS, Azure, or Google Cloud. On-premise and air-gapped options for regulated buyers.
A Data Processing Agreement is signed before any document moves. Your paper or ours — either is fine.
Customer invoices are never used to train foundation models. Per-tenant tuning happens only when you opt in.
Every extraction logs model version, rule set, reviewer actions, and timestamps. Exportable for internal and external audits.
TLS 1.2+ on the wire, AES-256 at rest. Customer-managed keys on enterprise plans.
If yours isn't here, a 30-minute call with an engineer is usually faster than an email chain.
No. OCR returns text and coordinates. The Invoice Agent identifies the invoice type, extracts header fields and line items into a schema, validates tax math and PO match, flags duplicates, and returns JSON with a confidence score per field. The OCR step is one part of a longer pipeline — and not the part that determines whether the output is usable.
PDFs (native and scanned), images (JPG, PNG, TIFF), multi-page scans, phone photos, and e-invoice JSON or XML payloads. Layouts include tax invoices, proforma invoices, credit and debit notes, utility bills, freight invoices, and the vendor templates from 2004 that nobody has updated since.
Header field accuracy is typically 97 to 99% on clean PDFs and 92 to 97% on phone photos or messy scans. Line-item accuracy depends on table complexity and falls in the 93 to 98% range once the agent is tuned to your vendor mix. The number that actually matters is how many invoices clear the auto-post threshold — that lands between 70 and 85% on most AP pipelines.
Yes. GSTIN format and checksum are validated. CGST, SGST, IGST, and cess are extracted separately. HSN and SAC codes are pulled at the line-item level for ITC matching. On e-invoices, the IRN and QR payload are parsed and reconciled against the document fields, so a tampered total gets flagged automatically rather than silently posted.
Two-way match (invoice to PO), three-way match (invoice to PO to GRN), and four-way match (adds contract or service confirmation) are all supported out of the box. Mismatches return the specific reason — short quantity, rate variance, tax mismatch — not a generic fail flag. The reasoner is configurable per tenant if your tolerance bands differ.
One REST endpoint. Upload the file, get JSON back, post the JSON to SAP, Oracle, NetSuite, Tally, Microsoft Dynamics, or your in-house ledger. Pre-built connectors exist for the common ERPs. SDKs ship for Python, Node.js, Java, and .NET. Webhooks are available for async batches and high-volume overnight runs.
In the cloud region you pick on AWS, Azure, or Google Cloud. On-premise and air-gapped deployments are available for regulated buyers. A DPA is signed before any invoice moves. Customer documents are never used to train foundation models — your data stays your data.
Yes. Send a batch of your own invoices to the sandbox and get JSON back within a working day. No credit card and no upfront sales call. If the output works on your documents, we talk pricing. If it doesn't, you still walk away with data you can compare against any other vendor.
Per invoice, with volume tiers. Pricing is indicative until we see your actual document mix — a phone-photo-heavy AP queue and a clean PDF EDI feed sit at different points on the cost curve. Enterprise plans bundle committed throughput, customer-managed keys, priority support, and a named solutions engineer.
The ones with rotated scans, missing line totals, the ones your team has flagged "do not auto-process". We run them through the Invoice Agent and send back the JSON. No deck, no sales call. Just output you can compare against what you have now.
