High-volume Retail Support
with Helpdesk Handover
A large home improvement retailer handles around 100K support conversations a month. A big support team was answering the same questions again and again. The AI agent now automates almost 80% of first-level requests and hands everything else to Zendesk as a ticket the team can act on straight away.
~100K
Conversations a month
~80%
Of first-level requests automated
AI
Title and summary written on each handover
1
Published integration guide
Stack
Tiledesk
Zendesk
OAuth
Webhooks
Make
Context
The problem
At around 100K conversations a month, a large share of the support team’s time went into the same few questions. Some requests can’t be answered by an agent at all, because they need a person to act. Those still had to reach the team with enough context to handle them.
01. The same questions every day
Order status, product availability and quote requests made up a large part of the volume.
02. Some cases need a person
Requests such as scheduling an in-field visit always need someone on the team to take them over.
Architecture
How it works: 4 layers
The agent answers what it can. Everything else becomes a Zendesk ticket, created automatically and filled in by the AI.
1
AI agent
First-level answers
Handles the repeat questions: order status, product availability and quote requests.
2
Handover
Anything that needs a person goes to Zendesk
Requests the agent can’t resolve, and cases like scheduling an in-field visit, are handed to the Zendesk ticketing system.
3
Integration
OAuth client and webhook
An OAuth client in Zendesk Admin Center and a webhook create the ticket automatically, classify its urgency and map the requester’s data onto it.
4
AI ticket content
Title and summary written by the AI
On handover the AI writes a ticket title and a summary of the conversation into the ticket, so the support agent has enough context to act.
Design decisions
Why each choice was made
Hand over a raw transcript, or a structured ticket?
A transcript makes the support agent read the whole conversation before they can start. A ticket with a clear title, a short summary and the requester’s data lets them act on it straight away.
Not chosen: raw transcript
Full history, but the team has to read it all to find the request.
Chosen: ticket with AI title and summary
The request is clear at a glance, with the conversation summarised and requester data mapped.
Direct OAuth integration, or Make?
The main integration uses an OAuth client and a webhook, with no extra platform in between. I also built a second version of the integration through Make, for teams that already manage their automations there and want to change the flow without code.
Alternative: Make scenario
Easier for a team to adjust visually. Adds one more platform to run.
Main version: direct OAuth and webhook
Fewer moving parts between the agent and Zendesk.
Why classify urgency at creation?
At this volume, tickets arriving in one undifferentiated queue would bury the urgent ones. The urgency is set when the ticket is created, so the team sees what matters first without triaging by hand.
Handover and measurement
Two paths for every conversation
Path A
Resolved by the agent
- Customer asks about an order, a product or a quote
- The agent answers in the conversation
- No ticket, no work for the team
Path B
Handed to Zendesk
- The request needs a person, such as an in-field visit
- A ticket is created with urgency and requester data
- The AI writes the title and conversation summary
- The support agent picks it up with full context
What I track on my assistants
- Total conversations and resolution rate
- Handovers to the team
- Answered and unanswered questions
- Knowledge base performance
Result
End to end
CustomerAsks a question
→
AI agentAnswers first level
→
HandoverOnly when needed
→
ZendeskTicket with title and summary
→
Support teamActs with full context
The agent automates almost 80% of first-level requests across around 100K conversations a month. Every case that reaches the team arrives as a clean ticket instead of a transcript. The integration was published as Tiledesk’s official Zendesk integration guide.
Takeaways
What I learned
- The handover is part of the product. The team only trusts the agent if the cases it passes on are easy to pick up.
- A short AI summary saves more time than a full transcript.
- Offering the same integration in two forms, direct and through Make, makes it usable for more teams.