Making email-based order processing reliable with LLMs

Rather than starting from scratch, we conducted a detailed technical review of the existing system and focused on the specific components that were causing failures in production.

Prompt engineering refinement. We redesigned the prompts to guide the LLM toward precise order extraction, reducing ambiguity and improving accuracy. Structured output schemas were introduced so each LLM response adhered to a predictable, machine-readable format.

Semantic product matching. We integrated transformer-based semantic alignment to match product names accurately, even when customer phrasing differed from the database. A final LLM-based validation step evaluates context before confirming the match, significantly reducing mismatch errors.

Attachment processing. We extended the system to extract order data from both email bodies and their attachments, closing a gap that had previously required manual handling.

Email classification. An intent classification step was added to determine whether incoming content related to a new order, a cancellation, or another request — ensuring only relevant content reaches the extraction pipeline.

Order extraction accuracy was substantially improved. Product matching errors were significantly reduced. The system now handles a wide variety of input scenarios – including email attachments, multi-language content, and inconsistent formatting – without requiring manual intervention for the vast majority of cases.

The collaboration demonstrated how incremental, focused improvements – combined with deep LLM expertise – can transform a promising prototype into an enterprise-grade production system.