Context
Regional distributor with 600 B2B customers and around 200 orders per day. Orders arrived by e-mail, WhatsApp and as PDF purchase orders, in three different formats. Four people spent their mornings re-typing them into the ERP.
The problem
Each order took about four minutes to key in. Multiplied by 200, that is more than 13 hours a day. Data-entry errors caused wrong deliveries roughly twice a week, and the morning backlog delayed truck loading until 11 am.
What we built
We built an n8n pipeline that reads every incoming e-mail and attachment, extracts the order lines with a language model, matches customers and products against the ERP, checks stock, and creates the order. Anything ambiguous goes to a review queue in a simple web page with a one-click approve.
Confirmations are sent back to the customer by e-mail or WhatsApp, and a dashboard shows the day's orders, exceptions and loading status in real time.
What we learned
The review queue was the key. Aiming for 100% automation would have taken months; 85% automatic plus a two-second human check shipped in four weeks and is more trusted by the team.