Manufacturing · Bathroom and sanitary equipment
Email order processing made reliable enough for production
Radaway, an international bathroom equipment manufacturer, needed order data pulled out of unstructured customer emails at scale. A first LLM system was already in place and showed clear promise. We took it from there to a reliable enterprise system in three weeks, keeping everything that already worked.
Key results
Client: Radaway
An international bathroom equipment manufacturer serving customers across several European markets.
- Industry
- Manufacturing
- Use case
- Order extraction from email
- AI approach
- LLM with semantic matching
- Systems
- Email and the product database
- Languages
- Multiple European markets
- Engagement
- Refinement and production hardening
In short
What it takes to make extraction production-grade
- Radaway’s orders arrive the way customers write them: as email, in whatever wording and format the customer chose.
- A first LLM system already extracted the order data and worked in controlled conditions. At the scale of real customer mail, four things still needed work: order details, product matching against the database, attachments, and telling an order from anything else.
- We reviewed the existing system rather than replacing it and hardened exactly those four components: prompts, semantic product matching, attachment processing and intent classification.
- Manual intervention fell by 90 percent, product matching reached 95 percent and above, and orders arriving as attachments went from being handled by hand to being handled at all.
The starting point
The challenge
The LLM system worked in controlled conditions but lacked the structured output formats, validation logic and edge-case handling that reliable enterprise operation requires. Frequent manual intervention was still needed.
Radaway processes a high volume of orders that arrive through everyday customer communication, primarily by email. To modernise that, they had introduced an LLM-based system to extract the order data from incoming messages automatically.
The initial implementation showed clear potential. Exposed to real customer communication at scale, the limits became visible: order details were sometimes misinterpreted, product references did not always match the database, and email attachments, where many orders actually arrive, were not part of the automated process at all.
What was missing was the layer that makes a system dependable rather than promising: structured output formats, validation logic and edge-case handling.
The build
What we built
Rather than starting again, we ran a detailed technical review of the existing system and worked on the four components that decide whether it can run unattended.
Prompts rewritten for precision
We redesigned the prompts to steer the model toward exact order extraction and remove ambiguity. Structured output schemas mean every response now comes back in a predictable, machine-readable shape instead of prose that has to be parsed.
Matching what the customer wrote to what is in the database
Customers do not use catalogue names. We added transformer-based semantic alignment so a product is matched even when the phrasing differs, followed by an LLM validation step that checks the context before confirming. Mismatches dropped sharply.
Attachments, finally
The system now extracts order data from both the email body and its attachments, which closes the gap that had been sending a whole class of orders to manual handling.
Deciding what is even an order
An intent classification step works out whether an incoming message is a new order, a cancellation or something else, so only the relevant content reaches the extraction pipeline.
What changed
The results
Before
Frequent manual intervention. Product mismatches, inconsistent extraction, and attachments handled separately by a person.
After
A reliable production pipeline. End-to-end extraction with minimal manual oversight, attachments included.
Extraction accuracy improved substantially and product matching errors fell sharply.
The system now handles a wide range of input without a person stepping in: attachments, multi-language content and inconsistent formatting.
Taking a promising system the last stretch to production is its own discipline, and it is usually cheaper than starting again: three weeks of focused work rather than a rebuild.
Questions about this project
Order extraction from email. The LLM system worked in controlled conditions but lacked the structured output formats, validation logic and edge-case handling that reliable enterprise operation requires. Frequent manual intervention was still needed.
Less manual intervention in order entry: –90%. Product matching accuracy after refinement: 95%+. From technical review to a production-ready system: 3 wks. Coverage of orders arriving as email attachments: 0 → 100%.
LLM with semantic matching. Technology used: Large Language Models, Prompt engineering, Transformer-based semantic matching, Structured output schemas, Email classification, Attachment processing, Python.
Technology used
Talk to the people who built this
theBlue.ai comes back within one business day.
More case studies
Automotive · leading luxury manufacturer
Planning data from SAP BW, asked for in plain language
Read more
Logistics · Fr. Meyer’s Sohn
Routing data pulled from thousands of shipping emails a day
Read more
Telecommunication · Arvato Bertelsmann