Bellus needed to detect real orders inside varied emails and attachments, extract usable fields, inspect AI output, and send structured results toward the operational workflow.
Email and attachments enter n8n.
Message body, files, and metadata are prepared.
Gemini extracts candidate order fields.
Operators inspect raw payloads and AI output.
Approved data moves to Google Sheets.
The system receives emails through a n8n webhook, parses content and attachments, classifies whether the email is an order, extracts order header fields and line items, stores the payload and analysis, and gives operators a dashboard for inspection.
Orders arrived as emails and attachments in different formats. The team needed to know what AI could extract, where it was uncertain, and whether the output could be trusted before downstream use.
The implementation included authentication, role-based access, session management, order/not-order classification, attachment handling, line-item extraction, order status tracking, payload inspection, dashboard statistics, and admin operations.
Operators can inspect raw payloads, extracted fields, confidence signals, processing state, failed records, and uncertain orders. Model updates, token limits, and retry behavior were part of the operating considerations.
The result is an operational workspace for email-order intake, with clearer visibility than raw JSON in a sheet and a safer path for moving extracted data toward internal systems.
Start with the inputs, extraction fields, review rules, and handoff target. Then build the workflow around that.
Scope an AI workflow