Phase 3: AI Document Processing & Data Extraction · 50 min · n8n · OpenAI API · LangChain

Email Parsing and Classification with LLMs

15,000 emails a day is a software problem, not a people problem.

Hiring signal: Email automation is the #1 use case companies ask for in automation interviews. Being able to describe the C.H. Robinson pattern (classification + extraction + routing) and the Experian pattern (semantic intent matching + RAG) shows you understand email automation at production scale.

What you will learn

Use LLMs to classify incoming emails by intent (order, complaint, inquiry, support, invoice)
Extract key fields from unstructured email bodies with LLMs
Design routing logic: classification result → workflow selection → human queue for edge cases
Apply the C.H. Robinson pattern: 7 minutes/email → automated classification + extraction + order creation

The Problem

C.H. Robinson, one of the largest logistics companies in the world, receives 15,000 shipment-related emails per day. Each email takes a human 7 minutes to process: read, understand the intent (load offer, tracking update, complaint, invoice), extract relevant data (BOL number, origin, destination, rate), and create the right record in their TMS. That's 1,750 hours of manual work per day.

They automated it with AI: LLMs classify each email by intent, extract structured data, and route to the appropriate workflow. The result: 5,500 orders automated per day, 600+ hours saved daily. This is the pattern every email-heavy organization needs to learn.

Email automation is the #1 use case in AI workflow automation
Every company has an inbox. Every inbox has repetitive processing. The C.H. Robinson pattern (classify → extract → route) and the Experian pattern (semantic intent matching + RAG) are the two production-proven approaches. Master these and you can automate any email-heavy process.

The Concept

The C.H. Robinson Pattern

┌──────────────────────────────────────────────────────────────┐
│  EMAIL AUTOMATION PIPELINE                                    │
│                                                              │
│  [Email Received]                                            │
│       ↓                                                      │
│  [Parse: Extract subject, body, attachments, sender]        │
│       ↓                                                      │
│  [AI: Classify Intent]                                       │
│  "Is this a load offer, tracking update, complaint,         │
│   invoice, or other?"                                        │
│       ↓                                                      │
│  [Switch: Route by classification]                           │
│  ├── Load Offer → [AI: Extract BOL, origin, dest, rate]    │
│  │               → [TMS: Create order]                      │
│  │               → [Email: Confirm]                         │
│  ├── Tracking   → [AI: Extract BOL, status, location]      │
│  │               → [TMS: Update shipment]                   │
│  ├── Complaint  → [AI: Extract issue, priority]            │
│  │               → [Slack: Alert manager]                   │
│  ├── Invoice    → [AI: Extract vendor, amount, lines]      │
│  │               → [AP: Create invoice record]              │
│  └── Other      → [AI: Extract entities]                   │
│                  → [Queue: Human review]                     │
└──────────────────────────────────────────────────────────────┘

The Experian Pattern: Semantic Intent + RAG

Experian automated 35% of customer emails using a different approach:

[Email] → [AI: Semantic intent matching against known intents]
          → [RAG: Retrieve relevant policy/procedure from KB]
          → [AI: Generate response using retrieved context]
          → [Confidence check: auto-send or human review]

Pattern	Best For	Key Feature
C.H. Robinson	Transactional emails (orders, tracking, invoices)	Classification → extraction → system action
Experian	Customer service emails (questions, complaints)	Semantic matching → RAG → response generation

Classification Categories

Category	Example Email	Extracted Fields	Action
Order/Load	"Offering a load from Chicago to Dallas, $2,500"	Origin, destination, rate, equipment	Create order in TMS
Tracking	"BOL 123456 delivered at Dallas facility"	BOL number, status, location	Update shipment
Complaint	"My shipment was damaged, I want a refund"	Issue type, priority, order ref	Alert manager
Invoice	"Attached invoice for April freight charges"	Vendor, amount, line items	Create AP record
Inquiry	"What are your rates for reefer loads?"	Question type, relevant info	Auto-respond from KB
Other	"Happy birthday from the team!"	Entities	Human review

An email says "Hey, following up on BOL-789012 — the driver said he's 30 minutes out from the delivery location." How should this be classified and what should be extracted?

This is a tracking update — the sender is providing a status update on an existing shipment. Extract BOL number (789012), status (en route), and ETA (30 minutes). The action would be to update the shipment status in the TMS with the new ETA. It's not a complaint (no dissatisfaction), not an inquiry (not asking a question), and not "other" (it has clear intent).

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Build It, Use It, Ship It, Exercises, Key Terms, Common Pitfalls — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.