Post
2200
What happens when you treat an AI support assistant like a product instead of a chatbot?
The Katana platform team had 5 engineers supporting 500+ production applications, with adoption doubling every year.
They built an AI assistant grounded in:
Slack history
Internal documentation
Custom instructions
Live platform API data, including logs and configuration
Today, it resolves 80%+ of support requests autonomously.
The interesting part is what happens around the model.
The team continuously improves the assistant based on real usage, monitors escalations and sentiment, keeps it updated as the platform changes, and maintains a human fallback when needed.
They also ran a two-week experiment with direct Slack support disabled before making the approach permanent.
The model is only one piece of the system. Context, feedback loops, observability, escalation paths, and human handoff are what make it work in production.
🔗 https://www.godaddy.com/resources/news/how-we-scaled-platform-support-10x-without-scaling-the-team
The Katana platform team had 5 engineers supporting 500+ production applications, with adoption doubling every year.
They built an AI assistant grounded in:
Slack history
Internal documentation
Custom instructions
Live platform API data, including logs and configuration
Today, it resolves 80%+ of support requests autonomously.
The interesting part is what happens around the model.
The team continuously improves the assistant based on real usage, monitors escalations and sentiment, keeps it updated as the platform changes, and maintains a human fallback when needed.
They also ran a two-week experiment with direct Slack support disabled before making the approach permanent.
The model is only one piece of the system. Context, feedback loops, observability, escalation paths, and human handoff are what make it work in production.
🔗 https://www.godaddy.com/resources/news/how-we-scaled-platform-support-10x-without-scaling-the-team