Document AI at Scale: How Sam Classifies, Extracts, and Acts on 20+ Document Types

The document chaos problem

Suppliers send shipping documents in every format: PDF BOLs, Excel packing lists, email confirmations, scanned invoices, portal exports, photography of physical documents. Operations teams manually open each one, extract relevant data, and enter it into the TMS or OMS. For supplier-managed shipments, they must also create shipment records manually. This bottleneck delays visibility, introduces errors, and scales linearly with supplier count.

The three-stage pipeline

Stage 1: Classification

Sam classifies incoming documents into 20+ supply chain document types: Bills of Lading, Commercial Invoices, Packing Lists, Proof of Delivery, CBP 7512 customs forms, Arrival Notices, Lumper Receipts, Delivery Receipts, and more. Classification is not rule-based (checking for keywords in headers). It uses document structure analysis, field pattern recognition, and content semantics to identify the document type even when formatting varies across suppliers.

Stage 2: Extraction

Once classified, the extraction pipeline pulls structured data: PO numbers, quantities, carrier names, BOL numbers, ship dates, tracking numbers, weight, dimensions, and document-specific fields (customs entry numbers, temperature readings, pharmacist credentials). Each extracted field carries a confidence score. Fields below the confidence threshold are flagged for human review with the AI's best guess pre-populated, reducing the human task from data entry to validation.

Stage 3: Downstream invocation

Extracted data triggers downstream actions automatically. A BOL with a carrier name and tracking number triggers shipment creation in FourKites. A POD with delivery confirmation triggers invoice unblocking in AP. A customs document with a compliance gap triggers an alert to the import team. The document does not just get processed; it drives the next operational step.

The RAG feedback loop

Validated corrections feed back into the extraction model through a retrieval-augmented generation (RAG) pipeline. When a human corrects a field that Sam extracted incorrectly, that correction becomes a training signal for future documents from the same supplier, format, or document type. The system gets measurably better with every document processed. This is not periodic retraining. It is continuous learning from production corrections.

Why this matters

A Fortune 100 CPG company eliminated 15,000 hours per year of manual document processing with Sam. 95%+ extraction accuracy across PDF, Excel, and email formats. But the real innovation is the pipeline: classify, extract, act, learn. Each stage feeds the next, and the feedback loop means the system improves with every document.