AI-Powered Email Processing System for a Fashion Retailer

Turning chaotic inboxes into structured, automated order & inquiry workflows

Retail operations look deceptively simple from the outside: customers email, staff respond, orders get booked, stock gets updated. Under the hood, it’s usually an unstructured mess—especially for small or mid-size fashion brands that haven’t fully automated their digital operations. The result is slow responses, lost orders, inconsistent inventory updates, and frustrated customers.

This project is a practical proof-of-concept that shows how LLMs can be deployed as real workers in this workflow. Not abstract magic. Actual operational logic.

The goal:

Automatically read incoming emails, understand whether they are order requests or product inquiries, extract structured data, check stock, update inventory, and generate responses—end to end.

Everything runs through a modular chain of LLM-based “micro-agents,” each responsible for a small, verifiable task. This keeps the system interpretable instead of becoming a mysterious black box.

Github Repo:

Business Context: Why This Problem Exists

A fashion retailer receives two types of recurring emails:

  1. “Do you have this in size M?”
    — product inquiries

  2. “I want to order 2 beige linen shirts.”
    — order requests

These land in a common inbox. Humans spend hours manually:

  • reading each email

  • determining intent

  • checking product catalog and stock

  • preparing responses

  • adjusting inventory

  • logging orders manually in sheets / db

The friction multiplies with scale.
The mistakes multiply even faster.

An LLM-based system solves exactly these bottlenecks:

  • no missed orders

  • real-time stock updates

  • consistent responses

  • structured logs for auditing

  • scalability without hiring more staff

The system is not meant to be a chatbot. It is meant to be an intelligent back-office worker—fast, predictable, and fully auditable.

Technical Overview: How the System Works

The design intentionally avoids monolithic “one giant prompt” architectures. Instead, each function is a narrow expert—an independent agent.

Processing happens sequentially, not in batches, to maintain inventory correctness. When two emails request the same product, sequential order ensures one stock update happens before the next evaluation.

Everything flows through a deterministic pipeline:

flowchart TD A[llm_classify_email] --> B[email-classification log] B --> C{Is Order Request?} %% Order Request Path C -->|Yes| D[llm_extract_order_items] D --> E[process_order_requests_pipeline] E --> F[llm_generate_order_response] F --> G[order-response log] %% Product Inquiry Path C -->|No, Product Inquiry| H[answer_inquiry_with_stock - RAG and stock lookup] H --> I[inquiry-response log] %% Final Step G --> J[Write all logs] I --> J[Write all logs]

This structure makes the system easy to reason about, easy to debug, and easy to extend.

Potential Future Extensions

  • Multi-store inventory sync
    • If stock is distributed across warehouses.
  • Price negotiation automation
    • Common in boutique fashion retail.
  • CRM integration
    • Tag customers by intent, lifecycle, and purchase history.
  • Predictive insights
    • Incoming email patterns → demand forecasting
    • (email-driven signals are underutilized gold)
  • Full order booking
    • Push confirmed orders directly into ERP or Shopify.

The pipeline already supports extension because each agent is isolated.