From 11 minutes to 38 seconds, per ticket.
Northgate Logistics moved 1,243 tickets a day through 18 dispatchers. Eight weeks after shipping their first Lumin agent, 94% of those tickets are auto-handled in 38 seconds — and the dispatchers spend their day on the 6% that actually need them.
A team running out of operators.
Northgate had grown 38% year-over-year for three years. Headcount couldn't keep up — they were hiring three dispatchers for every two that left, and the inbox queue kept getting longer.
Each dispatcher handled 70-90 tickets a day, taking 11 minutes per ticket on average. Most of the time was spent re-keying load details across email, the carrier portal, and their TMS.
Co-build the first agent, in three weeks.
We sat next to two dispatchers for an afternoon and watched what they actually did. Week 1, we mapped the agent in the Builder — Parse the email, Plan the carrier, Guardrail the margin, Post to the TMS, Write the reply, Log the receipt.
Week 2 was eval against the previous month's tickets. Week 3 it ran in shadow mode beside the dispatchers. By the end of week 3, the team flipped it live for one shipper, then five, then everyone.
Before. After.
We instrumented every ticket touch for the four weeks before kickoff. After eight weeks of Lumin in production, we re-ran the same measure on the same lane mix.
The auto-handle rate climbed from 12% (rules-based macros) to 94% (Freight Intake). Mean handling time dropped from 11 minutes to 38 seconds. Cost per ticket fell from €1.40 (loaded labor + tool cost) to €0.04 (cloud + tool calls).
“It’s not the AI we doubted — it’s the receipt.”
One ticket. Every byte.
Each ticket — auto or human — produces a signed JSON + PDF receipt. The auditor reads it instead of asking us what the agent did.