AGL Assistant handles a car parts dealer's email inquiries: it sorts them, asks for missing details and drafts both the inquiry to the supplier and the reply to the customer in their language. The team only approves the drafts. It has been in production since 2026.
A client project I worked on as part of a team at BeCode, mainly on the AI part, from supplier selection to draft replies.
Problem
Every question about a part's price or availability means two conversations: one with the customer and one with the supplier, who must not learn who the customer is. Suppliers sat in separate Slovak, Czech and Croatian spreadsheets, and part and car data was retyped by hand.
Solution
A background process reads mail from every inbox. OpenAI models assign each inquiry to one of 8 types, check that it contains everything needed, and read part numbers from attached screenshots and PDFs. An unclear inquiry gets a question about the missing details. A complete one becomes an inquiry to a supplier picked from a shared database.
Results
- 8 inquiry types, each with its own completeness rule and next step
- Drafts in 6 languages: Slovak, Czech, Croatian, Slovenian, German, English
- About 9,900 sent emails from the history set the tone of new drafts
- Each draft is compared with the reply actually sent, and only weak ones go to a person
Stack
- NestJS API and background worker
- PostgreSQL, Redis
- Next.js interface for the team
- OpenAI models for sorting, extraction and drafts
What I learned
A draft that can score itself saves review time. The team doesn't read every draft, only the ones that drift from how the company usually replies.