AI support grounded in business data
40-60 tickets a day in September 2026, with staff reviewing the proposed replies.
The problem
Support questions about book orders and products required information from several business systems. A useful answer had to reflect the underlying records.
My responsibility
At Independent Publishers Group, I co-designed the architecture and integrations of a RAG-based Freshdesk agent, working with the engineering team. My work connected the agent to business data and brought the workflow into daily operations.
The approach
The agent combines relevant support guidance with read-only access to operational systems. It prepares an internal note for a member of staff to review. This design makes the proposed answer inspectable before it reaches a customer.
Result and measurement
In September 2026, the system generated internal notes for approximately 40-60 tickets a day. This is a dated measure of operational use. It does not measure tickets resolved autonomously or staff hours saved.
The engineering work included evaluation of proposed answers and iteration on integration behavior. A useful production measure must distinguish answer quality, human corrections, coverage and cost.
Leadership perspective
Introducing an agent means deciding who reviews its output, who owns errors and how the team learns from corrections. Those decisions belong in the operating process alongside the software.