.svg)
Harmon Advisory was triaging, classifying, and drafting every support response by hand. Volume was growing and response times were slipping. We built an AI support system that reads every incoming email, finds the right answer in internal documentation, and has a reviewed draft ready before a human opens the inbox.
Harmon Advisory handles a high volume of inbound support queries through Gmail. Their support team was spending most of the day reading, classifying, and drafting responses manually — the majority of which were variations of questions their internal documentation already answered. The answers existed. Finding them quickly enough was the problem.
The team needed a system that could read incoming emails, understand what was being asked, retrieve the relevant answer from existing documentation, and prepare a draft — without a human doing the first three steps. They also needed the knowledge base to stay current automatically whenever a document was updated. Both problems required different builds. We built both.
Triage was manual and inconsistent — different team members categorised the same type of query differently. Response time varied based on who was available, not how urgent the email was. Internal documentation was spread across Google Docs that no one had time to search during a busy support shift. And whenever a document was updated, someone had to manually re-brief whoever was handling support that day.
Our team was reading the same types of emails and writing the same types of responses every single day. That work is gone now. They spend their time on the cases that actually need a person.
We documented the full flow from email arrival to response — every triage decision, every knowledge source, and every case where genuine judgment was required versus cases where the answer was clearly in the documentation. This gave us a precise picture of where automation could own the process and where humans needed to stay involved.
We built an n8n workflow triggered by Gmail that extracts clean text from incoming emails and runs it through an OpenAI classifier. Support queries go to a LangChain agent that performs semantic search across a Supabase vector database built from the company's full documentation set. The agent composes a draft response grounded in verified internal knowledge and creates it directly in Gmail — ready for one-click send or a quick edit before it goes out.
A separate workflow monitors Google Drive for file creation and update events. When a document changes, the pipeline re-ingests it — deletes outdated rows, reads the updated content, chunks it, generates embeddings via OpenAI, and inserts them into Supabase. New knowledge is available to the response agent the moment a document is saved in Drive. No manual re-ingestion, no re-briefing.
Manual triage steps. Every email is read, classified, and routed within seconds of arrival.
Increase in support volume handled per team member without adding headcount or extending response times.
Knowledge base maintenance. Document updates are reflected in the support system instantly with no manual process required.
