SAP BTP Document Information Extraction (DOX) & Business AI Processing Guide

1. The Evolution of Document Processing in SAP

For decades, processing incoming PDF invoices, purchase order confirmations, and freight documents in SAP relied on costly third-party add-ons or rigid OCR tools requiring hundreds of hours of manual coordinate template training.

SAP BTP Document Information Extraction (DOX) transforms this paradigm:

• Pre-trained Domain Models: Out-of-the-box support for invoices, purchase orders, payment advices, and delivery notes in over 40 languages.
• Generative AI Foundation Models: Leverages LLMs to extract unstructured header and line item data without template configuration.
• Clean Core Compliance: Interacts purely via released REST APIs and creates business documents using released ABAP Cloud interfaces (e.g. I_SupplierInvoiceTP) without database modifications.

2. Architecture of an Intelligent Document Pipeline

A complete enterprise document pipeline consists of four distinct phases:

1. Ingestion: Inbound documents arrive via email (BTP Mail Service), SFTP, or Fiori upload and are staged into an object store.
2. Extraction Job: Binary payload is submitted to DOX (POST /document/jobs) specifying client ID and schema version.
3. Enrichment & Master Data Matching: Extracted supplier tax IDs, IBANs, and purchase order references are enriched against S/4HANA master data views (I_BusinessPartner, I_PurchaseOrderItemAPI01).
4. Straight-Through Processing (STP) or Exception Workflow: If OCR confidence meets the business threshold (e.g. >= 0.85), documents post directly via RAP EML. Incomplete or ambiguous invoices trigger a human-in-the-loop task in SAP Build Process Automation / My Inbox.

3. Handling Asynchronous Job Callbacks

Since OCR and LLM extraction take several seconds to complete, synchronous blocking calls must be avoided. The pipeline supports two reliable mechanisms:

• Webhook Callback: DOX notifies an exposed SAP CAP or Kyma endpoint when extraction reaches status DONE.
• Application Job Poller: In ABAP Cloud, an Application Job polls GET /document/jobs/{id} at exponential intervals until the worklist is populated.