Zero-Touch PDF & Excel Order Intake Pipeline for Enterprise ERP Systems

1. The Baseline Bottleneck & Operational Reality

The enterprise is a national B2B wholesale distributor of industrial hardware and engineering components, processing 1,500 purchase orders monthly across an inventory catalog of 14,000 SKUs. A dedicated four-person Customer Operations team spent their working hours manually transcribing data from customer emails into the enterprise ERP suite (Comarch ERP XL and Subiekt GT).

Wholesale contractors and industrial buyers submitted purchase orders in dozens of non-standard formats: digital vector PDFs, multi-tab Excel workbooks (XLS/XLSX), scanned invoices with physical stamps, and raw table exports. Registering a single document required:

• Manually decoding buyer descriptions and cross-referencing buyer internal codes with internal master EAN indices,
• Validating contract pricing, volume tier discounts, and customer-specific price lists tied to tax identification numbers,
• Querying real-time credit limits and open accounts-receivable balances in the ERP accounting ledger,
• Manually re-typing line items line-by-line into the native ERP Sales Order entry form.

Average cycle time was 14 minutes per order. The team spent 350 staff hours every month purely re-keying data. Operator fatigue triggered errors in 4.8% of all orders (mistyped EAN digits, confusion between single units and master cartons, or mismatched discount formulas). These discrepancies led to costly credit-note adjustments, delayed warehouse pick-and-pack operations, stalled cash collection, and friction with Tier-1 accounts. Total latency from email receipt to warehouse fulfillment averaged 28 hours, burning 38,500 PLN (€8,900) monthly in repetitive labor overhead.

2. System Architecture & Engineering Layers

I engineered and deployed a containerized, sovereign intake pipeline built on Python 3.12, Pydantic v2 schemas, and an isolated PostgreSQL staging buffer to prevent direct, unvalidated writes to the core ERP database:

graph TD
  subgraph Ingestion["1. Ingestion & Pre-processing"]
    A[B2B Inbound Mailboxes / Wholesalers] -->|IMAP Webhook / n8n Core| B[PDF & Excel Attachment Extraction]
    B --> C[Azure Document Intelligence / Table OCR]
  end

  subgraph Extraction["2. Deterministic LLM Extraction & Validation"]
    C --> D[Python 3.12 Extraction Engine]
    D -->|Structured JSON| E[Pydantic v2 Schema Validator]
    E -->|Schema Mismatch| F[Langfuse Telemetry & Auto-Retry]
  end

  subgraph Staging["3. SQL Staging Buffer & Business Rules"]
    E -->|Validated JSON| G[(PostgreSQL Staging Buffer)]
    G --> H[Business Rule Engine]
    H --> I[Customer SKU to EAN Cross-Mapping]
    H --> J[Tier Pricing & Contract Matrix Verification]
    H --> K[Credit Limit & Balance Check]
  end

  subgraph Decision["4. HITL Decision Gate"]
    H --> L{Validation Condition?}
    L -->|Price Delta >2% OR Limit Exceeded| M[Exception Queue: Slack Alert + Operations Dashboard]
    L -->|100% Match - 88% Volume| N[Comarch ERP XL API / Subiekt GT API]
  end

  subgraph Execution["5. Transactional ERP Execution"]
    N --> O[(ERP Production DB: Sales Order ZS)]
    O --> P[WMS Picking Queue: Instant Warehouse Dispatch]
    M -->|1-Click Account Manager Authorization| N
  end
        

Architectural Layer Breakdown:

• Layer 1: Ingestion & Pre-processing: Inbound B2B mailboxes are monitored via self-hosted n8n orchestrator webhooks. Inbound attachments trigger SHA-256 cryptographic hashing for instant deduplication, followed by Azure Document Intelligence extraction to resolve complex spatial grid coordinates and multi-line invoice tables.
• Layer 2: Deterministic Extraction with Pydantic v2: Python 3.12 microservices process the extracted layout using language models constrained by strict JSON schema enforcement. Order headers (tax ID, PO number, delivery window, target shipping address) and line items (buyer part number, product description, ordered quantity, unit of measure, net price) are deterministically parsed. Any typing mismatch triggers Langfuse telemetry logging and an automated retry loop.
• Layer 3: SQL Staging Buffer (PostgreSQL) & Business Rules Engine: To safeguard production ERP integrity, extracted datasets land in an isolated PostgreSQL staging buffer. The engine runs:
  • Fuzzy association matching buyer part numbers against historical master EAN catalogs with a 98% confidence threshold,
  • Real-time price recalculation against customer-specific contract discount matrices stored in the ERP database,
  • Automated credit solvency checks against open receivables and trade credit limits.
• Layer 4: Decision Gate & Human-in-the-Loop Routing: Clean orders without discrepancies automatically pass straight into the production ERP pipeline (Zero-Touch). Orders displaying price variance exceeding 2%, unmapped buyer SKUs, or credit limit flags are routed to an exception queue, triggering a Slack alert with a 1-click mobile authorization dashboard.
• Layer 5: Transactional ERP Execution (Comarch ERP XL / Subiekt GT API): Approved records are committed via official ERP transactional APIs. The system registers official Sales Order (ZS) documents, secures inventory reservations, and pushes picking tasks to warehouse terminals in real time.

3. Performance & Business Impact (KPI Table)

4. Autonomy Boundaries: The Human-in-the-Loop Safeguard

To guarantee complete data integrity and prevent AI hallucinations, the pipeline maintains strict operational boundaries between automated execution and human oversight:

100% Autonomous AI Pipeline Tasks:

• Parsing multi-page tables across complex PDFs and spreadsheets while retaining column hierarchies,
• Extracting buyer corporate entity details via VAT/tax ID and pulling live ERP customer master records,
• Mapping buyer internal part numbers to internal master EAN indices via association matrices,
• Mathematical recalculation of line net sums, applicable VAT rates, gross totals, and tier volume discounts,
• Automated creation and status commitment of Sales Order documents for 88% of inbound traffic.

Tasks Requiring Human Authorization (12% Exception Volume):

• Discrepancies exceeding 2% between quoted buyer pricing and master contract ERP matrices,
• Insolvency risk flags: credit limit breaches or overdue invoice settlements,
• Novel inventory part numbers lacking prior mapping in the cross-reference registry,
• Handwritten or unstructured contract addendums requiring legal or commercial sign-off.

5. Deep Dive: High-Stakes Edge Case

6. Technology Stack & Data Sovereignty

The entire pipeline runs on dedicated, sovereign server infrastructure owned by the client. Sensitive commercial pricing, customer catalogs, and trade terms are never shared with public multi-tenant SaaS clouds. The enterprise maintains 100% code and database ownership with zero per-document licensing fees and zero vendor lock-in.

7. Executive Testimonial