Case study · Logistics · Customer support
From 30% to zero support backlog: how agentic AI transformed customer support for XpressBees.
XpressBees · Logistics and supply chain
Up to 15,000 emails a day, triaged by hand and in no particular order. We put an agent layer in front of the queue-and the backlog went to zero.
Zero
Support backlog, down from 30%
90%
Accuracy across 20 support categories
5x
Improved operational performance
50%
Efficiency boost for support representatives
The opportunity
Three to five thousand emails a day, with no way to tell which one mattered most.
XpressBees receives 3,000 to 5,000 customer emails on an ordinary day, and up to 15,000 in peak season. None of it arrived prioritised. Reviewing it was manual, and because manual review at that volume is inconsistent by nature, which query got attention first was effectively random-so time-sensitive inquiries surfaced late.
Non-Delivery Reports were handled inconsistently for the same reason. With no ticket workflow automation and no real-time visibility into what was in the queue, roughly 30% of it sat as backlog.
No structured prioritisation
Up to 15,000 daily emails arrived as one undifferentiated queue.
Manual review was inconsistent
Which query got looked at first was effectively random, so urgent ones surfaced late.
NDRs handled inconsistently
Non-Delivery Reports had no standard path through the queue.
No real-time visibility
Without a live view of the queue, roughly 30% of it stayed as backlog.
The impact
Every email routed, scored and drafted before an agent opens it.
100% of email through the agent layer
Nothing arrives untriaged: every incoming message is routed through the system rather than sampled by hand.
Prioritised, not first-come
Tickets are categorised by priority against 25+ recurring support categories, so time-sensitive queries surface first.
Context attached on arrival
Key entities are extracted and enriched with context, so the agent opens a ticket that already explains itself.
Drafted replies, human sign-off
A personalised draft is generated for agent review before sending-which is where the 50% efficiency gain comes from.
The stack
What it runs on.
- Model layer
- Amazon Bedrock foundation models
- Retrieval
- RAG-retrieval-augmented generation
- Training scope
- 25+ recurring support categories, 20 scored in production
- Human checkpoint
- Draft replies reviewed by a support agent before send
The road ahead
The queue is under control. Next is the rest of the journey.
With intake triaged and drafted automatically, the work moves outward: the same context that answers an email can resolve a shipment exception, and the same accuracy measurement extends to new categories as they appear.
- 01
More support categories in production
- 02
Proactive shipment exception handling
- 03
Context shared across voice and chat
Talk to us
Bring one process. Leave with an outcome architecture.
30 minutes on one function-intake triage, exception handling, or agent assist-mapped to the architecture behind it. No slides.