From customer email to transport order: an AI pilot designed to be reused

Transport orders arrive the way each customer prefers to send them: order files exported from freight platforms, spreadsheets with a different layout for every customer, or a few lines of free text in an email. Someone has to read each one and type it into the transport management system (TMS).

The project had two aims: to find out whether AI could do the reading reliably enough to be trusted, and to build the result once, as a template that can be reused for future clients.

The system follows the shape of the work. A customer request becomes one or more orders — one truck, one order — and each order is divided into legs.

Validation is all-or-nothing by design. A request that reaches the TMS only in part is harder to repair than one that doesn't arrive at all: someone has to find out which orders made it and which did not.

• Reading: AI extracts the orders from emails and attachments, whatever their format.

• All-or-nothing validation: every order in a request is checked, and the request goes ahead only if all of them pass. Nothing half-valid reaches the TMS.

• Conversion: validated orders are converted into the XML that the group's TMS imports.

• Access: branded single sign-on through Keycloak, and an interface in Italian, German and English.

The TMS remains the system of record. The pilot delivers the XML format the TMS already imports, so the AI step sits in front of the existing system instead of inside it, and can run alongside the current process without changing it.

Observability is built in with OpenTelemetry, so each request can be followed from the email to the XML: when someone asks about an order, the answer is in the trace, not in someone's memory. The security baseline is mapped to ISO/IEC 27001:2022 controls and OWASP guidance, and 25 architecture decision records explain why each part is built the way it is.

This is a pilot. The first real orders were reviewed with the client in September 2026.

We will report results once they are measured in production, not before.

• Large language models
• Keycloak
• OpenTelemetry
• XML import for the TMS