How to Set Up Rossum Invoice Capture for AP Automation
Rossum reads invoices and turns them into structured accounts-payable data. Setup is four moves: connect an inbox or upload channel, let it learn your supplier invoices, add validation rules for totals and tax, then push validated records to your ERP. A mid-size AP team gets to a usable line in about two weeks.
The two weeks are mostly spent on exceptions, not on the tool. Rossum handles clean single-page invoices well out of the box; the work is your multi-page, multi-currency, PO-less supplier invoices.
What does Rossum do differently from basic OCR?
OCR returns text. Rossum returns fields with confidence scores and line items — vendor name, invoice number, PO reference, tax breakdown, per-line quantity and price — and it uses a learned model per document type rather than a fixed template.
That distinction shows up on invoices with unusual layouts. A template-based tool needs a new template; Rossum usually gets there after seeing two or three examples.
Step 1: connect the capture channel
Settings → Queues → Email gives you a dedicated AP mailbox address. Forward ap@yourco.com to it, or point your existing AP mailbox rules at it.
If you use shared Outlook rules today, keep one rule that files an invoice copy into an archive folder. Rossum keeps documents, but an independent archive has saved more than one audit.
Step 2: train on your supplier mix, not on a demo
This is the part where projects either go well or stall. Take 30–50 invoices per high-volume supplier, ideally recent ones, and upload them into a training queue.
What to include: your three biggest suppliers by invoice count, plus any supplier whose invoices routinely get held for manual review. What to exclude: one-off weird invoices from the first pass.
Rossum’s extraction improves per document type, so the sequence matters. Train the 80% first, then handle the long tail.
Step 3: validation rules that catch what the model can’t
Extraction confidence is not correctness. Three rules catch most of it:
- Totals check — line item sum plus tax equals the invoice total, within a 0.02 tolerance for rounding.
- PO match — invoice PO reference exists in your ERP and matches the committed amount within tolerance.
- Duplicate check — vendor plus invoice number seen in the last 180 days.
Anything failing a rule goes to an exception queue with the failing rule named. Agents fix one field instead of hunting through a PDF.
Step 4: route by confidence and value
A rule that works well in practice: auto-approve documents above 95% field confidence under your materiality threshold, everything else to a human. Keep the threshold tied to money, not to percentage — a 95%-confident $40,000 invoice still deserves eyes.
Rossum exposes both confidence and totals on every document, so this is a queue filter, not custom code.
Step 5: export to the ERP
Native connectors exist for NetSuite, SAP, Microsoft Dynamics, Oracle, Coupa and Sage, plus a generic API. Choose the connector your ERP team will support, not the one with the nicest screenshots.
Two practical notes on the export step:
- Map to the ERP’s existing vendor master. If Rossum creates new vendors freely, your master data degrades within a quarter.
- Export only validated documents. Pushing exceptions into the ERP to “fix later” moves manual work to a more expensive place.
Is Rossum’s pricing reasonable for a small AP team?
Rossum prices per document, not per seat, which changes the maths. A team processing 2,000 invoices a month is the usual entry point for enterprise pricing; below about 500 documents a month, the per-document model rarely beats a per-seat tool or sending it to a local processing bureau.
Ask for the cost per extracted document including exceptions. The licence number is meaningless without the exception rate, and exception rate is a function of your supplier mix, not of the tool.
Where Rossum struggles
Handwritten delivery notes and photos of crumpled paper receipts. The model is trained on scanned and PDF invoices, and quality drops off outside that. Photographed receipts belong in an expense tool, not AP capture.
Multi-page invoices with 200 line items also push validation time up — the extraction is fine, the reviewer isn’t.
Quick Summary
- Rossum extracts invoice fields plus confidence and line items, and improves per document type.
- Train on 30–50 invoices per high-volume supplier before expecting reliability.
- Three validation rules — totals, PO match, duplicate — catch most extraction errors.
- Route by confidence and value, not confidence alone.
- Priced per document; the model usually pays off above ~500 invoices a month.
Related guides
- Parse PDFs into RAG-ready JSON with Unstructured — if you need document text in a database rather than AP fields.
- Run Unstructured at scale for contract review — the long-document version of the same extraction problem.
