An AI support agent that resolved 62% of tickets before a human saw them

The challenge

Support handled roughly 900 tickets a month, and about two-thirds were the same fifteen questions. Response times stretched past a day during admission season, and prospective students dropped off while waiting.

Research

• Clustered 6 months of ticket history to find the true question distribution.
• Reviewed existing help content for gaps and contradictions.
• Set an accuracy bar and escalation policy with the support lead upfront.

Strategy

• Ground every answer in approved content; never let the model improvise policy.
• Escalate fast and cleanly when confidence is low.
• Measure deflection and satisfaction, not conversation volume.

Design decisions

• Persistent widget with clear 'talk to a human' escape at every step.
• Answers shown with source links so students can verify.
• Multilingual replies matching the student's chosen language.

Development

• Retrieval over an approved content index with strict grounding rules.
• Confidence threshold that opens a real ticket with full transcript attached.
• Rate limiting and abuse protection on the public endpoint.

How we worked

1. 1
   Content

   Rewrote and structured the knowledge base.
2. 2
   Pilot

   Internal-only rollout with human review on every reply.
3. 3
   Launch

   Public launch with live escalation to the ticket queue.
4. 4
   Tune

   Weekly review of failed answers and content updates.