Invoice to cash: automating the back office without breaking the ERP
Document AI reads every invoice, matches it to the purchase order and receipt, and posts it — while exceptions route to humans with the evidence already assembled.
The problem
An industrial manufacturer processed a high volume of supplier invoices a month in eleven formats across four languages, with a shared-services team keying them into the ERP by hand. Roughly a fifth failed three-way match and went into an exception queue that ran two weeks behind. Late payments meant lost early-payment discounts; disputed invoices meant cash sitting still.
The hypothesis
Invoice processing is a well-bounded extraction and reconciliation problem. The judgment lives entirely in the exceptions. If extraction were reliable enough to make the clean majority touchless, the same headcount could clear exceptions in hours instead of weeks.
The build
- Extraction with confidence scoring — a document model reads header and line-item fields across formats and languages, returning a confidence per field rather than one score per document.
- Deterministic three-way match — invoice against purchase order against goods receipt, in ordinary code. Tolerances are configuration the finance team owns, not model behavior.
- Confidence-routed posting — every field above threshold and a clean match posts straight to the ERP. Anything below routes to a reviewer with the document, the extracted values, and the mismatch highlighted side by side.
- Exception assist — for genuine mismatches the system drafts the supplier query, citing the specific line and quantity in dispute.
Design choice that mattered: we never let the model adjust a tolerance or override a match rule. Finance controls are deterministic and auditable; the model's job ends at reading the page accurately. That division is what got this through internal audit.
Rollout
We began with a single supplier segment representing about a fifth of volume, ran it parallel to manual keying for four weeks, and compared field-level accuracy. Thresholds were then set per field — an invoice number needs near-certainty, a description does not. Volume expanded segment by segment over the following quarter.
Results
The large majority of invoices now post without a human touch. The exception queue moved from a two-week backlog to same-week, days-sales-outstanding improved by several days, and the team captured early-payment discounts it had been quietly forfeiting for years. Headcount did not change — the work did.
What we'd tell you
- Confidence per field, not per document. One uncertain line should not send a clean invoice to a human.
- Keep financial control logic in deterministic code. Auditors will ask, and "the model decided" fails.
- The prize is the exception queue, not the clean path. Measure the backlog, not just the automation rate.
- Roll out by supplier or segment, so accuracy problems stay contained and diagnosable.
More from the field.
Consumer lending, decided in hours
Early-warning credit risk
The 6 a.m. advisor brief
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.