EdTech company, customer service
A Support Agent That Really Knows the Knowledge Base
As part of an AI transformation at an EdTech company, we backed customer service with an AI agent.
| Industry | Education |
|---|---|
| Services | AI Agents, AI Transformation |
Situation
Customer service answered a large share of its inquiries repeatedly from the same sources: help articles, internal notes, individual employees' experience. The knowledge existed but was scattered, and every inquiry cost time that was missing for more complex cases.
Task
An AI agent was to answer customer inquiries directly from the knowledge base, with real access to current content rather than prewritten answers, and with a clear line for when a human takes over.
Implementation
Knowledge base indexing
The existing content was prepared for access by the agent.
Answer agent
Answers inquiries based on current content, not from a rigid script.
Escalation path
Clearly defined line at which an inquiry goes to a human.
Feedback loop into the knowledge base
Questions the agent cannot answer surface gaps in the knowledge.
Integration into the existing support tool
The agent works inside the team's existing system, not as an extra tool.
Approach
We first worked with the support team to define which questions the agent may answer on its own and which must go to a human. That line mattered more than the technology behind it. Testing was done on real, historical inquiries, not constructed examples.
The knowledge base was treated as a living system: when the agent could not answer a question, that was a signal to update the content, not a failure of the agent. The support team was involved in this upkeep so the database stays current.
Result
The agent answers recurring inquiries productively in live operation, not as a pilot project. The support team has more time for cases that actually need judgment, and the knowledge base gets actively maintained because the agent surfaces its gaps.
Transferable lessons
The knowledge base is the lever
An agent is only as good as the content it can access. Invest there first.
Define the line first
What the agent may and may not answer has to stand before the build, not after.
The agent is part of the process, not a widget
It only works embedded in the existing support workflow, not alongside it.
First conversation: 30 minutes, concrete.
We look at where an AI agent would actually carry weight in your customer service.