How we cut a client's support load by 70% with an AI agent

A growing SaaS client came to us with a familiar problem: their support inbox was drowning. Two staff spent most of their day answering the same questions — how to reset a password, where to find an invoice, how a particular feature worked. It was slow for customers and expensive for the business. Within eight weeks we shipped an AI support agent that now resolves around 70% of incoming questions without a human ever touching them. Here's how we built it.

Start with the data, not the model

The temptation with AI is to jump straight to a chatbot. We started somewhere less glamorous: the knowledge. We gathered the client's help docs, past support tickets and product FAQs, cleaned them up, and split them into small, well-labelled chunks. Good retrieval beats a bigger model almost every time — an accurate answer grounded in real documentation is far more valuable than a fluent guess.

Retrieval-augmented generation (RAG)

We used a retrieval-augmented generation pipeline built with LangChain. When a customer asks a question, the system finds the most relevant chunks of the client's documentation, then passes them to Claude with instructions to answer only from that context and to say when it isn't sure. This keeps answers accurate, on-brand and current — update a doc, and the agent's answers update with it.

• Embeddings index over help docs and resolved tickets
• Semantic search to pull the most relevant context per question
• Claude generates a grounded answer with citations back to the docs
• A confidence check routes uncertain questions to a human

Keep a human in the loop

The goal was never to remove people — it was to free them from repetitive work. When the agent isn't confident, or a question involves billing or account changes, it hands off to the support team with the conversation and suggested answer already prepared. Agents spend their time on the genuinely hard cases, and customers get instant answers to the easy 70%.

The results

Within a month of launch, the agent was deflecting roughly 70% of incoming questions, median first-response time dropped from hours to seconds, and the support team stopped dreading Monday mornings. Just as importantly, the client can see exactly what the agent answers and where it hands off, so they trust it — and keep improving the docs behind it.

Takeaways if you're considering this

AI support agents work best when they're grounded in your real documentation, honest about uncertainty, and paired with a clean human handoff. Measure deflection and satisfaction from day one, and treat the knowledge base as a living product. If your team is answering the same questions over and over, there's a strong chance a well-built agent can give most of that time back.