Case Study · Freight forwarding · Greater Vancouver
Anonymized Case — A Greater Vancouver freight forwarder: AI sorting for the info@ mailbox, tested in shadow mode first
Before the AI was allowed to sort a single email, it ran in shadow mode beside the staff — and agreed with their own sorting 85% of the time.
Privacy note: the client is anonymized. No company name, email content, customer or shipment details are shown — only the approach and the measured result.
Overview
For a Greater Vancouver freight forwarder: AI sorting of the shared info@ mailbox, run first in shadow mode next to staff (85% agreement with their own sorting), plus a back-office data-layer design that connects the forwarding software and Microsoft 365 while staff keep the tools they use. Client anonymized.
A freight forwarder's info@ address catches a bit of everything, and someone has to read each message and decide where it goes before any real work starts. That sorting is repetitive, but getting it wrong is costly, so the question was not "can AI sort email" but "how do we know it sorts the way this team does". The answer was to let it prove itself first: in shadow mode the AI sorted every incoming message alongside the staff, without touching anything, and we compared the two.
The challenge
Everything sent to the shared info@ mailbox had to be read and sorted by hand before it reached the right person. The work was repetitive and took time from people who had shipments to handle. Behind the mailbox, the forwarding software and Microsoft 365 held related information in separate places, and any change had to leave staff with the tools they already know.
Our approach
Start with sorting, and start safely. The AI read each new message in the info@ mailbox and proposed a category, in shadow mode: its choices were recorded but not acted on, while staff kept sorting as usual. Comparing the two showed where the AI agreed with the team and where it did not. In parallel we designed a back-office data layer that connects the forwarding software and Microsoft 365, so information can flow between them without asking staff to change tools.
Results
In shadow mode the AI's sorting matched the staff's own sorting 85% of the time. The mismatches show exactly which kinds of email need clearer rules or should stay with a person. The data-layer design maps how the forwarding software and Microsoft 365 connect; staff keep working in the tools they already use.
Why shadow mode first?
An AI that moves email around on day one can lose a message nobody notices. Shadow mode removes that risk while producing the one number that matters before go-live:
- The AI sorts every incoming message, but nothing is moved or answered.
- Staff sort as usual; their choices are the reference.
- Agreement is measured on real mail, not a demo set: 85%.
- The disagreements become the work list — sharper categories, or messages that should always go to a person.
The decision to let the AI act is made from measured agreement, not from a sales demo.
The back-office data layer: connect the systems, keep the tools
Forwarding software and Microsoft 365 each hold part of the picture. The design adds a data layer behind them so the two can share information, instead of replacing either one. Staff keep their screens; the connection happens in the back office.
Technical breakdown
- Mail source: the shared info@ mailbox; the AI proposes one category per incoming message.
- Shadow mode: AI decisions are logged next to the staff's own sorting and compared, with no action taken on the mailbox.
- Back-office data layer (design): connects the forwarding software and Microsoft 365 without changing the tools staff use.
Reusable lessons
- Measure agreement with the people who do the job today before letting automation act.
- The mismatches are the most useful output of a shadow run: they show where the rules are unclear.
- Connect systems behind the scenes; asking staff to switch tools is the expensive part of any change.
Modules used
Want results like these?
Tell us about your business and we'll scope a paid proof of concept.