Document intelligence for accounts payable: what actually works

Every accounts payable team has the same pile: supplier invoices arriving as PDFs, scans and photographs, in dozens of layouts, keyed by hand into Fusion Payables. Document intelligence promises to read them. It does. The reason many pilots stall is that reading was never the hard part.

Extraction with confidence, not extraction

Modern document models, including OCI Document Understanding and the general-purpose vision models, return each field with a confidence score. Treat that score as the product. A supplier name at 0.98 can flow straight through; a total at 0.71 goes to a person. The threshold is a business decision per field, and it should start conservative and loosen as the measured error rate earns it.

The fields that matter

• Supplier identity: name, tax or registration number, bank details if present. Matched against the Fusion supplier master, never trusted from the document alone.
• Invoice number and date, which drive duplicate detection.
• Purchase order number, which decides whether this is a PO-matched or a non-PO invoice and therefore who approves it.
• Line items and totals, reconciled against each other; a document whose lines do not sum to its total is a review item regardless of confidence.
• Currency and tax, which are where silent errors cost real money.

Design the review queue first

The review queue is where the automation rate is won or lost. A reviewer should see the document and the extracted fields side by side, with low-confidence fields highlighted, one keystroke to accept, and the correction fed back as a labelled example. If reviewing takes longer than keying, the team will key. Build the queue as a product, measure time per document, and treat every correction as training data or as a rule.

Matching belongs in rules, not in the model

Two-way and three-way matching against purchase orders and receipts is deterministic and already exists in Fusion. Let the model extract and the rules match. Tolerances, hold reasons and approval routing are configuration, and configuration is auditable in a way a model's judgement is not.

Posting to Payables

Accepted invoices reach Fusion through the Payables REST API for individual documents or through FBDI import for batches, with the original document attached and the extraction confidence stored as a descriptive flexfield so audit can see what was automated. Fusion's own Intelligent Document Recognition covers the straightforward email-in path; a custom pipeline earns its place when you need your own thresholds, your own queue or sources beyond a mailbox.

Measure the right things

• Straight-through rate: the share of invoices posted with no human touch.
• Field-level accuracy on a held-out sample, re-measured monthly, because suppliers change their templates.
• Review time per document, which tells you whether the queue is a product or a punishment.
• Exceptions by cause: unknown supplier, missing PO, total mismatch. Each one points at a fix upstream, usually in supplier onboarding.