Intelligent Document Processing

An agent that ingests incoming documents (invoices, claims), extracts structured data, validates it, and pushes it to downstream systems (ERP, CMS) - replacing a manual processing team. The key business KPI is STP rate (Straight-Through Processing): the percentage of docs processed without any human touch. Day-1 STP is typically ~70%; after 6 months of corrections and retraining it reaches 90-95%.

“Defense in depth - extraction errors caught by validation; validation failures caught by human review. No single check failure lets a corrupt record reach the ERP. Queue-based and idempotent - every step retryable safely. Immutable audit trail from day one - not added later.”

Incoming Docs (email, SFTP, portal, scanner)
  → [1] Ingest + Dedup + Queue (Pub/Sub)
  → [2] Classify  (what doc type is this?)
  → [3] OCR / Text Extraction  (image → text + layout)
  → [4] LLM Extraction  (text → structured JSON fields)
  → [5] Validation
       ├── PASS  → [6] Downstream Adapter → ERP/DB → Audit: DELIVERED
       └── FAIL  → [7] Human Review Queue → Corrections → Re-validate → Downstream
                         ↓
                   Corrections = labeled training data → fine-tune → higher STP
  → [8] Audit Log (immutable, every state transition)

• 1Ingestion - Multi-channel adapters (IMAP, SFTP, REST API). SHA-256 hash for deduplication. Write to Pub/Sub queue - decouples ingestion rate from processing rate. Store raw files in immutable GCS/S3 (never delete originals - audit source of truth).
• 2Classification - Vision-language model (Gemini Pro Vision / Document AI) on first page → { doc_type, confidence }. If confidence < 0.85 → human classifies before extraction. Wrong classification = garbage extraction downstream.
• 3OCR - Digital PDF → extract directly (PyMuPDF). Scanned PDF / image → Google Document AI (returns text + bounding boxes + table structure). Layout-aware OCR is critical - column/table structure tells you which zone a value belongs to.
• 4Intelligent Data Extraction - Highest Value, Highest Risk - Use forced JSON output (Gemini response schema / Anthropic tool use) - never parse free-form LLM text for financial data. Hybrid approach: rules/regex for known vendor templates (cheap + fast), LLM only for novel layouts. Table line items → use OCR table structure directly, not LLM text parsing.
• 5Validation - Two Layers - Syntactic: required fields present, dates valid, math checks (line items = subtotal + tax = total), duplicate invoice number. Cross-system: PO matching vs. ERP, vendor master check, amount threshold (>$50k = human approval). 3-way matching = invoice vs. PO vs. goods receipt.
• 6Human Review Queue - Original doc rendered side-by-side with extracted fields. Highlight the failing field + its source bounding box on the doc image. Every correction captured as { field, model_prediction, human_correction } → training data. Re-validate after correction before downstream push.

• 1Extraction errors that pass validation - Mitigated by 3-way matching + math checks (defense in depth) - internal consistency alone is not enough.
• 2Human review queue overload - Mitigated by queue-based autoscaling on queue depth (Kubernetes HPA) + SLA alerting when queue exceeds threshold.
• 3Model drift as vendor formats evolve - Mitigated by per-vendor accuracy monitoring + correction flywheel that continuously fine-tunes on human corrections.

Google Stack: Document AI (OCR + form parsing) → Gemini Pro (LLM extraction) → Cloud Pub/Sub (queue) → Cloud Run / GKE (workers, autoscale) → BigQuery (audit + analytics) → Looker (ops dashboard)

Azure Stack: Azure Blob Storage + Azure Service Bus (ingestion/queue) → Azure AI Document Intelligence (OCR, 300+ languages, layout + table extraction) → Azure OpenAI GPT-4o (LLM extraction, forced JSON) → Azure Logic Apps / Power Automate (downstream ERP adapters) → Azure AI Search (vendor template matching) → Azure Human-in-the-Loop (review queue) → Log Analytics Workspace (audit log)

AWS Stack: S3 + SQS (ingestion/queue) → Amazon Textract (OCR + AnalyzeDocument, native table/form extraction) → Amazon Bedrock / Claude (LLM extraction) → AWS Step Functions (workflow orchestration) → AWS Lambda (downstream adapters) → Amazon Augmented AI / A2I (human review queue) → DynamoDB + S3 Object Lock (immutable audit log)

Other Options: UiPath Document Understanding + OpenAI (enterprise RPA path) | Microsoft Power Automate AI Builder (low-code, Office 365 native) | Rossum + SAP (finance-specific, pre-built ERP connectors) | Reducto + Anthropic Claude (developer-first, high-accuracy extraction API)