← All case studies
Case 02· E-commerce & Retail· 2025–2026· RAG & Support AI

Eight years of support history, turned into instant answers.

A privacy-first RAG knowledge system for The Lighting Outlet — staff search, agent copilot, and (eventually) customer chat. Running on hardware they own.

Client
The Lighting Outlet
Sector
E-commerce lighting retail
Engagement
Design, build & rollout
Timeline
Phase 1 built · Phase 2 deploying
§ 01 The brief

The problem

Eight years of customer support answers, product knowledge, policy decisions and workarounds — all locked inside a help-desk tool and the heads of a handful of senior agents.

Every new starter spent weeks repeating questions their colleagues had already solved a hundred times. Customers waited longer than they should for answers to things the team already knew cold. A "proper" chatbot was tempting, but the off-the-shelf options all wanted the data in someone else's cloud — a non-starter for a business that values its customer data.

The approach

We built an on-premise RAG system: tickets, product data, website content and policy docs extracted into a structured knowledge base, served to staff through a streaming AI search tool and, progressively, a Gorgias copilot and customer chat widget.

No model fine-tuning, no cloud vendor holding the data. PII is scrubbed at ingest and again on output. Every answer is grounded in real records with confidence scores and a review queue when the model isn't sure. When the engagement ends, TLO keeps the hardware, the data, and the code.

Depth and privacy.

~80k
Tickets of history
made searchable
11k+
Products
understood
100%
On-premise
zero cloud
0
Per-seat
SaaS fees
§ 02 Approach

Built on the client's hardware — not somebody else's cloud.

Every part of the system runs on a small server inside TLO's office. Customer data, support history, product information — none of it leaves the building. No per-seat SaaS fees, no vendor holding the crown jewels hostage, no surprise price rises at renewal.

The AI is grounded in real records. Every answer cites the ticket, article or policy it came from. When the system isn't confident, it flags the question for a human rather than guessing. When policies change, the knowledge updates instantly — no retraining cycle.

§ 03 Inside the system

Where the AI shows up for the team.

Staff

Ask anything

A plain-English search tool for the support team. Type a question, get an answer in seconds — grounded in real tickets, policies and product info. Every response cites where it came from, so staff can verify before acting.

In the inbox

Drafts in Gorgias & Slack

The AI works inside the tools agents already use. A sidebar inside Gorgias drafts responses based on the open ticket; a Slack bot answers questions without leaving chat. Low-confidence answers go to a manager review channel before anything goes out.

§ 04 Delivery

Three phases — each one useful on its own.

01

Staff search Built

The internal knowledge tool, deployed and in use by the support team. Trains nobody, helps everybody — new starters answer like veterans from day one.

02

Agent copilot Rolling out

AI drafts the reply inside Gorgias; the agent reviews and sends. Keeps the human in charge, takes the repetitive drafting off their plate.

03

Customer chat Planned

A customer-facing chat — only switched on once confidence and guardrails are proven across thousands of internal answers first.

"Eight years of answers are finally searchable — and none of it sits in someone else's cloud."

— Placeholder client quote · The Lighting Outlet

Got a similar problem?

If your team's knowledge lives in a help desk, a shared inbox, or the heads of your best agents — we'd like to hear about it. Initial conversations are free.

Start a conversation