A regional logistics operator came to us with a familiar problem: a 40-person operations team drowning in shipment-status tickets, most of them asking questions the data already answered somewhere in their systems.

The starting point

Average first response time sat at just under six hours. Roughly 60% of incoming tickets were status lookups — "where is my shipment," "why is it delayed," "can I change the delivery window" — that required an agent to log into two separate systems just to answer.

What we built

Instead of a single chatbot bolted onto the help desk, we deployed a small stack of purpose-built agents:

  • A retrieval agent that pulls live shipment status across the carrier APIs and internal warehouse system in one call.
  • A resolution agent that handles delivery-window changes directly, within a defined policy envelope, without a human touching the ticket.
  • An escalation agent that recognizes disputes, damage claims, and anything emotionally charged, and routes those straight to a senior rep with full context attached — no re-explaining required.

The six-week build

Weeks one and two went entirely to mapping the existing workflow and connecting to source systems cleanly — most of the real engineering risk lives in integration, not in the agent logic itself. By week four we were running the stack in shadow mode alongside the human team, comparing its answers to real agent responses before it ever touched a live ticket.

68% of tickets never needed a human at all — they were resolved correctly on the first automated pass.

The results after eight weeks live

  • First response time dropped from 5.8 hours to under 100 minutes — a 71% reduction.
  • 68% of status and change-request tickets resolved with zero human involvement.
  • The operations team reallocated to the disputes and account-management work that actually needs a person.

Nothing about this required a huge model or an exotic technique. It required correctly separating "questions the data already answers" from "questions that need judgment," and building for each of those cases on its own terms.