E-commerce & Retail
Harbor Supply Co.
A support agent that answers in seconds and knows when to stop talking
Harbor Supply Co. sells marine parts online — and two support staff were drowning in 400+ tickets a week, most of them repetitive. We built a RAG support agent grounded in their help center and live order data, deployed to site chat and email, with a handoff design that gets a human involved the moment one is actually needed.
Two people, four hundred tickets, and the same twenty questions
Harbor Supply Co. runs a marine-supplies store with a catalog full of parts that must fit specific engines, hulls, and fittings — which means customers ask a lot of questions before and after buying. Volume had grown past 400 tickets a week, handled by a support team of two. First responses were taking a day and a half at peak, and the team's ticket triage had become the job, with actual problem-solving squeezed into whatever was left.
When we analyzed the queue, roughly 70% of tickets fell into three repetitive buckets: order status, returns and exchanges, and fitment questions already answered somewhere in the help center. The remaining 30% genuinely needed a human — a warranty dispute, a damaged shipment, a customer mid-repair on a boat that won't start. The brief wrote itself: automate the 70% without ever fumbling the 30%.
Six months of real tickets beat anyone's intuition about what customers ask
We started with data, not design workshops: six months of ticket history — several thousand conversations — clustered by intent and resolution. That corpus told us exactly which questions recurred, which help-center articles actually resolved them, and where the docs were silent, contradictory, or out of date. About a dozen articles were rewritten before a line of agent code existed, because a retrieval system grounded in wrong documentation is just a faster way to be wrong.
The ticket history also became the exam. Before launch, the agent was evaluated against hundreds of real historical tickets with known-good resolutions, scored on whether it answered correctly, cited the right source, or — equally important — correctly declined and escalated. It went live only when it beat the bar on all three.
Grounded in the help center, wired into the order system
The agent runs on frontier LLMs behind a LangChain orchestration layer, with the help center, returns policy, and fitment guides embedded in Pinecone for retrieval. Every answer is generated from retrieved passages and quotes its source — if the documents don't contain the answer, the agent says so and escalates rather than improvising. It never invents policy, because it is structurally prevented from answering policy questions without a retrieved policy passage to stand on.
For order questions, retrieval isn't enough — so a Node.js tool layer gives the agent read-only access to the live order API. "Where's my order?" gets answered with the customer's actual tracking status, not a paragraph about shipping in general. The whole thing deploys through Intercom on both site chat and email, so the agent meets customers in the channels they were already using.
The moment it should stop being a bot, it stops being a bot
Most support bots fail at the exit, trapping frustrated customers in loops of "I didn't quite get that." We treated the handoff as the core feature and designed the triggers explicitly: retrieval confidence below threshold, a customer asking for a human in any phrasing, sentiment turning negative, two failed clarification attempts, or any topic on the always-escalate list — warranty disputes, damage claims, anything involving a safety-critical part. Any one trigger ends the agent's turn immediately.
The handoff itself is where the day-and-a-half response time used to hide, so we made it seamless: the human agent receives the full conversation, the customer's order context, and a one-line summary of what's needed — the customer never repeats themselves. The agent doesn't pretend the transfer isn't happening, either; it says a person is taking over and sets an honest expectation for when. Customers forgive being escalated. They don't forgive being looped.
Two-thirds of the queue resolves itself, and the humans got their jobs back
Within the first two months, 68% of tickets resolved end-to-end with no human touch — almost exactly the repetitive share the ticket analysis predicted, which is what happens when you size the automation to the data instead of the demo. Median first response fell from a day and a half to under 60 seconds, around the clock, including the weekend mornings when boat owners actually discover problems.
Post-chat CSAT settled at 4.6/5, and — the number we watched most nervously — satisfaction on escalated conversations rose too, because handoffs now arrive with context instead of a cold start. The two-person support team stopped triaging and went back to solving the hard 30%, which is both the work that needs them and the work they're good at.
From answering questions to preventing them
The agent's conversation logs have become Harbor Supply's best product-content research: every question it couldn't answer is a documentation gap with a frequency count attached. The help center now gets updated from that list monthly, which steadily shrinks the escalation rate without touching the agent at all — the cheapest accuracy improvement available.
Next on the roadmap: proactive order-delay notifications, so the agent tells customers about a late shipment before they ask, and retrieval over the full product catalog to handle deeper fitment questions pre-purchase. Harbor Supply owns the stack end to end, and the models behind it are swappable as better ones ship. The queue, meanwhile, stays quiet — which around here counts as a feature.
Built With
I was ready to hate a chatbot. What sold me was watching it give up gracefully — the moment a question got hairy, a customer was talking to my team with the whole story already on screen. Our two support folks finally do support again.Owner · Harbor Supply Co.
More Case Studies
Want Results Like These?
Tell us what you're building — or what's breaking. We'll come back with a concrete plan and a fixed price.
Book a Free Consultation
At Axomble, our expertise lies in creating brand loyalty. We are dedicated to forging connections that lift your name and fame, making you the world's leading brand.
© Copyrights 2023-2026 | Axomble | All Rights Reserved