---
title: "AI Chatbot for Customer Support — Case Study"
description: "How we built a RAG knowledge assistant that deflected repetitive support tickets and cut average response time for a SaaS support team."
canonical_url: "https://mercurystk.com/case-studies/ai-chatbot-for-customer-support"
last_updated: "2026-06-29T04:31:10.269Z"
---

<case-study-hero :breadcrumbs="[{"label":"Home","to":"/"},{"label":"Case Studies","to":"/case-studies"},{"label":"AI Support Assistant"}]">



</case-study-hero>

## The challenge

A growing B2B SaaS support team was answering the same product questions over and
over. Knowledge was spread across a help center, internal policy docs, and the
heads of a few senior agents. New hires took weeks to ramp, and response times
climbed during busy periods.

> The numbers below are representative of this type of engagement. Real figures
> are shared with prospective clients under NDA.

## What we built

We built a **retrieval-augmented knowledge assistant** that:

- Ingested the public help center, internal SOPs, and past resolved tickets
- Answered questions **with citations** back to the source document
- Enforced which content was customer-facing vs. internal-only
- Surfaced "answer gaps" — questions the knowledge base couldn't answer well

The assistant was embedded both in the customer help widget and in the agents'
internal console.

## The approach

1. **Content audit** to find authoritative, up-to-date sources
2. **Ingestion pipeline** with scheduled re-syncing
3. **Retrieval + re-ranking** tuned against a labeled evaluation set
4. **Guardrails** so the assistant defers instead of guessing
5. **Analytics** on deflection, satisfaction, and content gaps

## Results

By grounding answers in real content and measuring quality before launch, the
assistant deflected a meaningful share of repetitive tickets and cut first
response times — while giving agents faster, cited answers.

## Stack