How does AI invoice processing actually work - extraction, validation, coding, routing?

OCR converts a document image into raw text - it reads characters but doesn't understand them, so it can't tell an invoice number from a phone number. Machine-learning extraction adds learned recognition of which text is which field, trained on labeled examples, and degrades when it meets layouts unlike its training. LLM-based understanding reasons about the document the way a person skimming it would - inferring that a number in a footer is the invoice number from context - which handles novel and messy layouts far better. The strongest systems layer all three: OCR to read, ML and LLMs to understand, and structured rules to validate the result against your ERP.

Each prediction carries a confidence score, and a threshold decides its fate: high-confidence values are suggested and pre-filled for quick confirmation; low-confidence values are left for a human to enter or are flagged for review. Well-designed thresholds are the whole game - set too low, the system suggests garbage confidently and the human stops trusting it; set too high, everything routes to a human and the automation does nothing. The right design routes human attention to exactly the fields and invoices where judgment is needed and lets the routine flow through with a glance.

Header fields (invoice number, date, total) appear in predictable places and there are a handful of them; line items are a variable-length table where every vendor formats columns differently, descriptions are free text, and one logical line can wrap across rows. Extracting them means inferring table structure, not just reading values - which is why line-level extraction lags header extraction in accuracy and why many "AI" tools quietly only do headers. Line-level capture matters because matching, coding, and price-creep analysis all depend on it.

The system predicts coding by learning from your history: for each line, it finds how similar past invoices from the same vendor were coded and proposes the most likely account, department, class, and location. Predictions improve as the system accumulates your corrections, and a vendor change re-triggers prediction because vendor is the strongest coding signal. It's pattern-matching on your own decisions, not a universal rule - which is why it gets sharper the more consistently your team codes.

Validation runs after extraction and before posting: duplicate checks compare the invoice against history at multiple points; PO matching reconciles invoice lines to purchase-order and receipt data, including at line level despite inconsistencies; and tolerance tests flag price or quantity variances beyond a set band. Extraction tells you what the invoice says; validation tells you whether to trust it - and surfaces the ones that fail for human investigation.

Reasoning-based models generalize to unseen layouts because they infer field meaning from context rather than matching a stored template, so a new vendor or a redesigned invoice doesn't break them the way template OCR breaks. For poor-quality scans and photos, systems fall back to vision-based extraction that reads directly from the image. Accuracy is lower on degraded documents than on clean digital PDFs, which is why measuring on your actual document mix matters.

Yes, within limits: the system learns routing from organizational structure and the history of who approved similar invoices (by vendor, amount, department, GL coding), then predicts the likely approver. It works well where past routing was consistent and needs configuration plus enough history to perform - it's learning your patterns, so it can't predict a routing rule you've never demonstrated. Humans confirm and can override.

Invoices arrive through multiple channels (email, upload, portal, EDI, CSV), get converted to structured records via OCR plus AI, are enriched with vendor/PO detection and coding predictions, run through validation (duplicate, matching, tolerance), route to approvers, and post to the ERP after human approval - with the ERP remaining the system of record throughout. Each stage feeds the next, and corrections at any stage feed back to improve future predictions.

Stampli AI runs this full pipeline inside the invoice workflow. It evaluates structured, ERP-aligned fields across requests, invoices, and receipts; extracts header and line-level data; auto-detects vendors and PO numbers; predicts GL coding from the organization's historical patterns; and predicts approvers from organizational structure and past behavior - with a multi-model orchestration layer selecting the best prediction per field. Every prediction carries a confidence score, and only high-confidence values surface as suggestions; the AP user reviews, corrects, and approves before anything posts to the ERP. On average, Stampli AI performs 87% of the field-level work across 2,700+ unique fields, with humans handling validation and the exceptions that need judgment.