An AI sales agent that actually closes the order

An AI sales agent is a conversational assistant trained on your entire catalog, policies and brand voice — designed to close orders, not deflect tickets. A traditional chatbot follows scripted decision trees and routes to a human when it hits a dead end. Zubby's agent uses retrieval-augmented generation (RAG): it semantically searches your products and docs at every turn, recommends specific SKUs, answers shipping and returns questions with inline citations, and can add items to cart inside the chat thread. The result: 4.8% storefront conversion versus the 1.4% Shopify default.

Yes — disclosure is on by default and configurable per store. We don't pretend to be human, but shoppers consistently rate the conversations as more helpful than a typical chatbot because the agent answers their real question instead of falling back to 'I'll connect you with someone'. If a question genuinely needs a human, the agent hands off cleanly with full conversation history.

On install, the catalog-sync worker pulls your full product graph from Shopify (via Bulk GraphQL) or WooCommerce (via the Zubby plugin) and embeds every product into a 1536-dimension pgvector index with HNSW. Webhooks keep it fresh: a price change or new SKU reaches the agent within minutes. Policy docs, FAQ articles, and sizing guides are uploaded once and embedded the same way.

Both. Product recommendations render as visual cards inside the chat thread — image, name, price, stock — with a one-click 'Add to cart' button. On Shopify, the agent uses the storefront cart API; on WooCommerce, the plugin proxy. Variant selection (size, color) happens inline so the shopper never leaves the conversation.

Brand voice is a first-class setting. Configure tone (warm/expert/playful), formality, emoji policy, sign-off, and house phrases. The system prompt is composed per store from those values, and the recovery-tuned variants used in emails inherit them. The same voice runs across the widget, recovery emails, journey messages, and any white-labeled output.

Median first response is 1.4 seconds at p50, under 3 seconds at p95. Retrieval and rerank run in parallel; generation streams tokens to the widget as they arrive. The agent runs on a multi-provider routing layer (OpenAI, Azure OpenAI, Anthropic, Gemini) ordered by health and priority so a single provider outage doesn't take you offline.

Yes — 25 languages, auto-detected from the shopper's session. The agent, the product card labels, the recovery emails and the policy citations all switch in lockstep. You can override the detected language per audience rule if you need to (e.g., always Japanese on JP traffic regardless of browser locale).

Three layers. First, the system prompt forbids answering policy or inventory questions without a retrieved source. Second, guardrails block competitor mentions, medical claims, financial advice, and a configurable blocklist of phrases. Third, product cards inject as structured data — the model can't invent a price, stock level or shipping window because it's filling slots, not writing them.