How the AI works

Zubby’s AI isn’t a plain chatbot. It’s an agent: a language model wrapped in a retrieval pipeline, a function-calling tool registry, and a guardrail layer that keeps it inside the lanes you set. This page explains what happens between a shopper typing a message and the AI replying - so you know exactly what to tune when an answer is wrong.

Merchant Ask: store questions and reviewed plans

Merchant Ask is separate from the shopper assistant. A signed-in user must have access to the selected store. It answers using connected dashboard metrics over 7, 30 or 90 days and can propose campaign plans. Requests do not consume shopper conversation credits. Daily caps are 10 on legacy Starter, 25 on Growth and 100 on Pro; Enterprise has no numeric daily cap in this endpoint. Data availability, provider routing and other applicable controls still apply.

Review a proposed campaign before approval. Approval and sending are different steps: inspect the resulting unsent broadcast or paused journey, audience, channel, consent and schedule before separate launch or activation. An illustrative question such as “What changed in recovery over 30 days?” is not a tested customer result.

Actions depend on the store platform

Catalog and policy answers depend on synchronized or pushed data and completed indexing. Cart actions require an explicit shopper request and a supported storefront adapter; an emitted cart action or event is not proof that an item reached checkout. Verify the actual cart and variant in a controlled storefront session.

WooCommerce cart behavior depends on the paired plugin and bridge. Verified shipping-address changes need the applicable setting and explicit confirmation. Refunds, cancellation and order edits can require human tickets rather than direct WooCommerce writes. Existing Shopify writes require the relevant scopes, enabled action, verified identity and confirmation; new Shopify connections remain coming soon. Custom stores need product-URL or adapter fallbacks and cannot inherit native cart or order-write support from catalog ingestion alone.

For implementation, follow WooCommerce setup or Custom API setup on Pro or Enterprise. For evaluation and allowances, see Pricing. Read-only MCP exposes search_products, get_product, get_store_info and search_policies; it does not perform checkout or inherit REST write scopes.

The five stages of a reply

1. Context assembly. Zubby gathers everything it knows about this turn: the page the shopper is on, their cart, their identity (if logged in), their language, and your merchant layers - brand voice, rules, locked facts, playbooks, and capability toggles. All of it is composed into a single system prompt for this turn.
2. Retrieval. If the question needs catalog or policy knowledge, we embed the shopper’s message into 1,536-dimensional vectors and run a hybrid search across your products, policies, FAQs, and pages. The top hits become the agent’s “context window”.
3. Tool selection. The model decides whether to answer directly or call one of its tools: search products, compare, get bundle recommendations, get order status, check inventory, get size recommendation, get smart offer, get store policy, search knowledge, add to cart, create ticket, and more. Tools use the connected platform or the store-scoped data Zubby has received. Custom-store facts are only as fresh as your latest API push.
4. Drafting. The model writes the reply grounded in the retrieved snippets and tool outputs, under hard grounding rules: product facts, prices, stock, and policy terms may only come from retrieved context or a tool result - never from the model’s memory.
5. Guardrail pass. As the reply streams, a guardrail scans it against the global safety filter, your vertical’s blocked-claim pack, and your own blocked phrases. A violating reply is aborted mid-stream and replaced with a safe fallback.

Retrieval

Retrieval is the most consequential stage. If the right snippets do not appear, the intended behavior is to ask for clarification, use another tool, or offer a human handoff. An unsupported answer is a defect to review, not a substitute for a missing source. Three things shape retrieval quality:

• Catalog completeness - every product synced with rich descriptions, attributes, and at least one image alt text.
• Knowledge base depth - your policies, FAQs, and long-form pages indexed. See Knowledge base.
• Embedding freshness - when a product changes, Shopify / Woo webhooks trigger a re-sync and re-embed automatically.

Grounding & auditability

Facts must come from somewhere. The agent is only allowed to state a product or policy fact that appears in the retrieved context or a tool result from this turn - if the fact isn’t there, it calls a tool or says it will confirm. These are controls, not a guarantee: AI can still produce an unsupported statement. You can audit this per conversation: the detail view in Inbox → Conversations shows each reply’s tool calls and the sources behind it, and the Most referenced by the AI panel on the Knowledge page ranks documents by citation count.

When the AI hits a question it can’t ground, the moment is flagged and lands in the Knowledge Gaps inbox (Training) so you can close the hole.

Tools

Tools are functions the agent can call to get fresh data or take an action. What the agent is allowed to do is controlled from Settings → AI Setup: capability toggles cover guided selling, lead capture, bundle recommendations, smart discounts, and more. The money-action tools - cancel order, modify order, refund order - are per-store opt-ins that default to off; until you enable them, those requests route to a human ticket. Address changes are the exception and are on by default, with verified ownership, eligibility, and confirmation checks. Execution differs by platform; see the platform action matrix. You can also define your own tools under Settings → Custom Tools.

Two tools depend on what your store has connected:

• Product reviews (Growth and above, Reviews module). For quality, fit, and “is it good?” questions the agent cites the product’s average rating with its rating count, the themes buyers praise and complain about, and a few short quotes without reviewer names. It is instructed to use only retrieved ratings and prioritize your policies and locked facts over reviews. Review answers for accuracy. See Review-grounded answers.
• Subscriptions (ReCharge connected). Signed-in shoppers can see their subscriptions and next charge date. Skipping, rescheduling, pausing, resuming, or cancelling needs the Manage subscriptions switch, which is off by default, and every change needs the shopper’s confirmation. See Subscriptions in chat.

Guardrails

Two layers:

1. Universal. A global blocked-claim filter (medical cures, “FDA approved”, guaranteed results), vertical-specific pattern packs for high-risk categories, and prompt-injection defense - everything inside shopper messages and retrieved content is treated as data, never as instructions. Always on.
2. Brand-specific. Your “do not say” list (Block Claim rules), tone overrides, and approved-offer discount limits. Edit them on the Rules & Voice page; the persona itself lives in System prompt.

AI service availability

Zubby manages the underlying AI service; it is not a store setting. The service supports chat, search indexing, and recovery generation. Merchants configure agent behavior, approved knowledge, rules, and budgets in the store workspace rather than managing the underlying service configuration.

Availability depends on the wider service, so check the support team or contact support when replies or indexing remain unavailable.

The service can route requests to another eligible provider when one is unavailable. A working alternative, credentials, budget, and the required capability must still be available. Replies and indexing can fail or be delayed; failover does not guarantee uninterrupted service.

How to fix a wrong answer

1. Open the conversation in Inbox → Conversations and look at the reply’s tool calls - the fix depends on which stage failed, so diagnose before you edit anything.
2. No source was retrieved? The knowledge is missing - add a knowledge base entry, or review and correct an AI-drafted one from the gaps inbox before approving it.
3. A source was retrieved but it’s outdated? Fix the document itself; edits re-embed automatically. Promote hard facts to Locked Facts so they’re quoted verbatim.
4. The facts were right but the framing was wrong? That’s a voice problem - adjust your brand voice or add a tone / block-claim rule under Rules & Voice.
5. Wait for the embedding job to finish, then re-ask the question in the widget and inspect the cited source to confirm the fix landed.

Worked example: correct a delivery answer

Illustrative test, using a fictional store: the policy says dispatch takes two business days, but the agent promises next-day delivery. Open the exchange in Inbox → Conversations, inspect its source, and distinguish dispatch from carrier transit time in the approved policy. After indexing, ask both “When will you dispatch?” and “Will it arrive tomorrow?” A useful result states the supported dispatch window and declines to guarantee an unsupported arrival date. Also test a destination outside your shipping area. Keep the test question and source with your review notes; a fluent answer alone does not verify the correction.

If the source is correct but the answer is still wrong, record the conversation, time, and expected source for support. If no replies are available, check AI Budget, source-indexing status, and troubleshootingbefore repeatedly changing the prompt.

What good replies look like

Audit-worthy AI replies share three properties: they’re grounded in at least one real source, they answer the actual question rather than the adjacent question, and they offer a clear next action (product link, return policy, escalate). Review conversations against those three axes in Inbox → Conversations; the weekly digest surfaces patterns worth fixing.