How to Cut Rossum Exception Rates on Messy Supplier Invoices
Rossum’s value is entirely a function of exception rate. At 8% exceptions, a 2,000-invoice month still leaves 160 documents for humans. At 25%, you’ve automated the easy half and made the hard half slower. Getting from 25% to under 10% is four specific changes, and only one of them is inside Rossum.
The counter-intuitive one: fix the supplier, not the model.
What counts as an exception?
In Rossum, an exception is a document that fails validation or falls below your confidence threshold and therefore stops for human review. Track it as a single number per month: exceptions ÷ documents received.
If your AP team can’t state that number, that’s the first job. Everything else is guessing.
Lever 1: get better invoices from suppliers
The biggest source of exceptions isn’t extraction quality — it’s invoices that were never machine-readable. Blurry scans, missing PO numbers, tax lines that don’t sum to the total.
Three fixes, in order of effort:
- Send suppliers a template with the fields you need, and a note that PO references must appear as a separate field, not in a free-text memo.
- Redirect high-volume suppliers to e-invoicing. Even 20% of volume moving to structured PDF or XML removes a chunk of exceptions permanently.
- Reject-and-return the worst offenders. A supplier whose invoices fail half the time costs more than their discount is worth.
Teams consistently underestimate this lever because it’s procurement work, not tooling work. It’s also the only one whose gains compound.
Lever 2: tune confidence per field, not per document
Document-level thresholds are blunt. Vendor name being uncertain is harmless if invoice number and total are solid. Line items being uncertain matters much more.
Raise the per-field confidence threshold on total, tax, and PO reference; leave vendor and description lower. Then route to a short-form review where the reviewer only fixes flagged fields.
What thresholds to start with?
Around 95% for money fields, 85% for descriptive fields is a sane start. Watch the two-week exception count, adjust by five points, re-measure. Don’t chase a single digit — the marginal reviewer minutes cost more than the exceptions you remove.
Lever 3: make the duplicate check smarter
A naive duplicate check on vendor plus invoice number fails in three ways:
- Supplier sends a corrected invoice with the same number. Legitimate duplicate number, different amount.
- Vendor name spelling varies (“Ltd” vs “Limited”, abbreviations).
- Credit notes share numbering ranges with invoices.
Normalise the vendor name against your master data before comparing, and compare amount too. A same-number, different-amount document is a correction, not a duplicate — route it to review with a “supersedes previous” flag instead of blocking it.
Lever 4: design the exception queue for one decision per document
The typical exception queue shows 40 fields and an agent who fixes the first visible error, re-submits, and finds the next one. Four rounds later, the invoice passes.
Split it:
| Queue | Contains | SLA |
|---|---|---|
| Missing PO | PO reference absent or unmatched | 1 business day |
| Amount mismatch | totals or tax fail | 4 hours |
| Low confidence | below threshold, all else passes | next batch |
| Unreadable | extraction failed outright | back to supplier |
The point is that an agent in the Missing PO queue has exactly one job. Batch them by supplier and one person resolves twenty in ten minutes with the right reference open in the ERP.
Does more training data always help?
No. Retraining on your exception documents teaches the model the messy layouts, which can pull overall accuracy down slightly on the clean ones. Retrain on corrected documents — the ones where a human fixed a field and confirmed it — not on the ones that failed extraction entirely.
Run retraining quarterly, not weekly. Each cycle needs enough new corrections to move the model, and weekly cycles just add noise.
What if the exception rate plateaus at 15%?
That may be the floor for your supplier mix. A business with 400 small suppliers sending ad-hoc PDFs will not reach 5%. The right response is to stop optimising the model and start consolidating suppliers, or to accept 15% and staff for it. Chasing the last ten points of automation on structurally bad input is how AP automation projects become expensive hobbies.
Quick Summary
- Exception rate = exceptions ÷ documents; learn yours monthly before tuning anything.
- Supplier-side fixes are the biggest lever — templates, e-invoicing, and rejecting chronically bad invoices.
- Tune confidence per field: ~95% on money fields, ~85% on descriptive ones.
- Duplicate checks need name normalisation and amount comparison, or corrections get blocked.
- Split exception queues by failure type so each reviewer makes one decision per document.
Related guides
- Set up Rossum invoice capture for AP automation — queue setup, training corpus and ERP export.
- Parse PDFs into RAG-ready JSON with Unstructured — when the problem is document structure rather than accounting fields.
