We create a live “visitor brain” by piping behavioural events (scroll depth, dwell time, exit intent) from Google Analytics 4 and your e-commerce backend into a lightweight feature store in Snowflake. We enrich every session with catalogue, CRM and campaign metadata so the system knows whether someone came from a décor guide, a Facebook ad, or a loyalty-program email. These real-time signals allow the trigger engine to greet high-value shoppers before they drift away.

We embed a React chat widget that talks to GPT-4-level LLMs through a tight system prompt: friendly, brand-safe, and concise. We connect a small Retrieval-Augmented-Generation (RAG) index of top FAQs, shipping rules, and return policies so the bot always answers with verified facts. We start on 10 % of traffic to measure tone, latency, and engagement. The first win we notice is a spike in proactive greetings—silent visitors finally say “Hi.”

We transform the bot into a personal shopper: we embed every SKU with a CLIP or text-embedding vector and drop them into vector database. When a visitor asks “Does this sofa come in blue?” the LLM issues a function call that returns visually similar items plus in-stock variants. The response feels instant and curated—driving higher average order values.

We introduce an abandoned-cart notification flow. If a shopper adds items worth >$50 and shows exit intent (closing tab, idle for 30 seconds), the bot fires a friendly “Still thinking it over? Here’s 10 % off and free shipping.” If the visitor has already left, trigger an email or SMS within 15 minutes, populated by the LLM with personalised copy and product images.

We export the first 50 k chat transcripts and fine-tune the model so it speaks brand’s language and cuts wordiness by ~18 %. We reinforce the agent with a reward model that prioritises higher cart values and penalises hallucinations.

We roll out to 100 % of web and mobile traffic, turn on 24/7 coverage, and plug conversation data into Klaviyo or Braze for lifecycle marketing. We schedule monthly retraining and KPI dashboards so the system keeps learning while human staff focus only on complex edge cases—cutting live-chat labour costs by half. Future channels (WhatsApp, voice, in-room AR) can reuse the same back-end with minimal tweaks.