AI provider & model
The AI Setup tab is where you decide which language model answers your shoppers and how the agent behaves on your storefront. Find it under Settings → Behavior → AI Setup, or go to /app/<storeId>/settings/ai. The page has two cards: Brand voice and AI conversion flows. The chat model itself is routed automatically by the platform — there is no model picker.
Which model answers shoppers?
You don’t pick the model, and that’s deliberate. Zubby routes every chat to the best available model on the platform and fails over automatically if a provider is degraded, so you always get a fast, healthy model without watching provider status pages.
Model routing is managed for you
chatcapability to a concrete model based on each provider’s status, health, and priority, and re-routes within seconds if the primary provider returns errors or breaches latency budgets. Provider and model selection is a platform-level control, not a per-store setting — so there is no model dropdown on this page. You still control cost exposure on AI budget.Brand voice (inline)
This card shapes how the AI sounds without leaving the AI Setup page. It has four controls plus a live preview:
- Tone profile— Professional, Friendly, Technical, or Luxury. Defaults to Friendly.
- Greeting style— Standard, Concierge, or Fast and direct.
- Emoji usage— No emoji, Light emoji, or Playful emoji.
- Custom instructions— a free-text box for specific guidance, e.g. “keep replies short, reassure about shipping speed, avoid medical claims, and push premium bundles before entry products.”
The dark preview panel on the right shows your configured agent name and welcome message and a sample exchange so you can sanity-check the tone before saving. Click Save AI setup to persist.
Two brand-voice surfaces
AI conversion flows
The third card is a set of toggles plus two dropdowns and an A/B test. Each toggle turns a capability on or off for your store:
- Page-aware context— let the AI adapt to product, cart, collection, and support page context.
- Variant-aware recommendations— use the selected variant, pricing, and stock signals in recommendations.
- Guided selling— allow the widget to run quiz-like product finder flows.
- Lead capture— capture the shopper’s email when intent is strong but the sale isn’t complete.
- Human handoff— offer support or sales escalation when AI confidence is low.
- Bundle recommendations— promote bundles, pairings, and attachments when relevant.
- Smart discounts— allow the AI to use merchant-approved incentive playbooks.
- Browse abandonment— track high-intent browsing for later follow-up capture and recovery.
The two dropdowns below the toggles fine-tune behavior:
- Guided selling style— Product finder, Concierge, or Diagnostic.
- Handoff destination— Support, Sales, or Concierge. Where escalations route.
Finally, the Opening message A/B test block lets you split visitors across two greetings (Opening message A and Opening message B) to improve open-to-conversion performance. Enable the switch, fill both messages, and click Save conversion flows.
Each card saves independently
Related
For the full request pipeline behind these settings — retrieval, tools, and guardrails — read How the AI works. To author the agent’s do-not-say rules and persona, see System prompt. Handoff behavior is covered in Human handoff.