Customer support · Production
Assembled builds Cal, an LLM-powered agent-assist product for customer support
The problem
Customer support AI tools have focused on deflection to cut contact volume, but this still leaves handle time unaddressed so cost per case remains high. Assembled saw an opportunity to make agents faster rather than replace them.
Workflow diagram · grounded in source
1
Agent needs answer or reply
trigger
“We start by helping agents quickly find answers and write replies.”
2
Cal pulls in sources
ai_action
“I built a very simple Slackbot prototype that could help by pulling in sources and answering questions. That was our first iteration of Cal.”
3
Cal generates reply draft
ai_action
“the team launched a new feature for alpha users that would automatically generate replies”
4
Agent interacts with output
human_review
“In the three days that followed, they re-engineered the feature to be more interactive.”
Reported outcome
Re-engineering the auto-reply feature to be more interactive increased Cal's usage by 50%.
Early users like Honeylove centered OKRs around increasing Cal's usage and improving its accuracy.
Reported metrics
Cal feature usage50%
Reported stack
OpenAIZendesk
Source
https://www.assembled.com/blog/a-conversation-with-the-team-behind-assembleds-big-bet-on-ai
Read source ↗Frequently asked questions
What did this team achieve with this AI workflow?
Re-engineering the auto-reply feature to be more interactive increased Cal's usage by 50%.
What tools did this team use?
OpenAI, Zendesk.
What results were reported?
Cal feature usage: 50% (source-reported, not independently verified).
How is this customer support AI workflow structured?
Agent needs answer or reply → Cal pulls in sources → Cal generates reply draft → Agent interacts with output.