Case study
AI customer-support email agent
A retrieval-augmented agent that turns a support inbox into reviewed, ready-to-send draft replies
Project overview
Our client's support inbox was full of repetitive questions that still deserved fast, personal answers. We built an agent that reads each incoming email, identifies the customer, retrieves the right answer from the brand's own FAQ knowledge base, and writes a polished reply — saved as a draft for a human to review and send with one click. Because every reply is grounded in the customer record and real FAQ material, answers are accurate and personal. Drafts are ready within minutes of an email arriving, saving the team around six hours of support work every week.
The challenge
Problem: A high volume of repetitive support questions (enrollment, schedules, policies) was eating hours every week, and answer quality drifted depending on who replied. The goal was to handle the repetitive 80% instantly and on-brand — without ever auto-sending a machine-written email to a customer.
How it works
Triggered by the email itself, the agent grounds every reply in the customer record and the FAQ knowledge base before drafting.
Technical implementation
Retrieval-augmented generation (RAG)
Brand FAQs and policies are embedded with Cohere and stored in a Pinecone vector database. For every email, the most relevant passages are retrieved and fed to the model, so answers are grounded in real source material rather than guessed.
Context pre-fetch from the system of record
Before the model runs, the workflow normalises the sender's address and looks up their enrollment record in Google Sheets, so the reply already knows who the customer is and what they signed up for.
On-brand drafting with Gemini
A Google Gemini agent composes the reply in the client's tone of voice, using the retrieved FAQ context and the customer's record as grounding.
Human-in-the-loop by design
Nothing is sent automatically. Every response is saved as a labelled Gmail draft, so a person reviews and sends with one click — keeping quality control and a human signature on every message.
Key features
- Reads every inbound support email automatically
- Identifies the customer and pulls their enrollment record
- Retrieves the best-matching FAQ answers from a vector store
- Drafts a polished reply in the brand's voice
- Saves it as a labelled Gmail draft for one-click review
- Leaves edge cases for a human instead of guessing
Need a system like this?
Tell us where work gets stuck — in one call you'll hear what it takes to remove it.
Impact
Technologies used
Business value
- Faster first-response times on support email
- Consistent, on-brand answers every time
- No email slips through the cracks
- Staff time refocused on genuinely tricky cases