An AI agent on first-line support
Before a cohort started, support was taking 900 requests a day. The agent answers from the knowledge base with a link to the source.
- Client
- Online school, 12,000 students
- Industry
- Education
- Year
- 2026
- Duration
- 4 weeks
On the days a cohort started, the support queue reached 6 hours. 70% of the questions repeated: access to lessons, payment, deadlines, refunds.
What we did
and why we did it that way.
Cleaning up the knowledge base
We collected the internal policies and found 40 contradictions between documents. That gained more quality than the choice of model did.
No answer without a source
If the agent cannot cite a source, the answer is not shown to the student—the request goes to a human.
Training in the shadows
For two weeks the agent only suggested answers to the operator. That produced the labelled data, and no student ever noticed it was there.
Injection testing
We ran AI red teaming before switching it on: we found a way to extract someone else’s email address and closed it at the permissions level.
What changed
and what we measured it with.
How it went
day by day and week by week.
- Week 1
Knowledge base
Collecting the documents and resolving the contradictions.
- Week 2
Agent and tools
Search across the documents, access to payment status.
- Week 3
Shadow mode
Suggestions to the operator, labelled data collected.
- Week 4
Red team and launch
Injection testing, switched on for 30% of traffic.
How it looks
in schematics.
These are screen schematics. We do not publish client interfaces without permission
A similar problem
on your side?
Describe it in the brief. In working hours we come back with an estimate of time and cost within two hours.