The challenge
The client's support team was spending most of its day answering the same handful of questions. Where is my shipment, how do I change a setting, why did this status update. Ticket volume was climbing faster than the team could hire, and response times were slipping. They wanted to deflect routine questions without subjecting customers to a frustrating, scripted chatbot.
Our approach
We proposed a retrieval augmented generation agent, often shortened to RAG. Rather than letting a model guess from general knowledge, the agent answers only from the client's own help docs, product guides, and past resolved tickets. Every answer is grounded in a real source, and the agent links to it.
- 1We gathered and cleaned the knowledge base: help centre articles, internal runbooks, and a representative sample of resolved tickets.
- 2We split that content into passages and stored them in a vector database, so the agent could find the most relevant pieces for any question.
- 3We built the agent so each reply is generated from the retrieved passages, with a citation back to the source.
- 4We added a clear handoff: when confidence is low or the topic is sensitive, the agent routes the customer to a human with the full conversation attached.
What we built
- A chat widget embedded in the customer dashboard, matched to the client's brand.
- A retrieval pipeline that pulls the most relevant documentation for each question.
- Source citations on every answer, so customers and agents can verify the response.
- A confidence threshold and human handoff, so the agent never bluffs through a hard case.
- An analytics view showing deflection rate, common questions, and gaps in the documentation.
The results
Within the first two months the agent was handling a large share of routine questions on its own, and the support team shifted its time toward complex cases that genuinely need a person.
- A meaningful drop in tier-1 tickets reaching human agents.
- Faster first response, since the agent replies instantly at any hour of the day.
- A useful side effect: the analytics surfaced weak spots in the docs, which the team then improved.
Because every answer cites a source, the agent also became a quality check on the documentation itself. Questions it could not answer well pointed straight to gaps worth fixing.
A note on accuracy
We designed the agent to say I am not certain, let me connect you to the team rather than invent an answer. For a support tool, a confident wrong answer is worse than an honest handoff, and customers ended up trusting the system more because of it.