AI Shopping Assistant: Scope and Handoff

An AI shopping assistant is a chat window on your storefront that answers questions in plain language, finds products and, in some versions, builds a cart. The risk is not that it will be dull. It is that it will say something confident and wrong about your shipping, returns or a product, and a shopper will act on it. This post sets out what a shopping assistant on Shopify should do, what it should refuse, how the handoff to your team should work, and what a 2024 ruling about an airline's chatbot means for a store owner. It ends with a pre-launch checklist.

A chatbot is part of your store's voice

In February 2024 a British Columbia tribunal decided Moffatt v. Air Canada (2024 BCCRT 149). A customer asked the airline's website chatbot about bereavement fares, and the chatbot told him he could apply for a reduced fare after travelling. Another page on the airline's website said the opposite. Air Canada argued that the chatbot was a separate entity responsible for its own actions. The tribunal rejected that. In the words quoted by McCarthy Tetrault, "While a chatbot has an interactive component, it is still just a part of Air Canada's website," and "It should be obvious to Air Canada that it is responsible for all the information on its website." The tribunal also found the airline "did not take reasonable care to ensure its chatbot was accurate," and it was not persuaded that customers should have to double-check the chatbot against another page.

That is a Canadian tribunal decision on a small claim, not US law, and we are not suggesting an identical result in a US court. The useful lesson is practical. A shopper will treat the assistant's answer as your answer, and "the policy page says otherwise" is a weak defense. In the US, the Federal Trade Commission put a related idea more broadly when it announced Operation AI Comply in September 2024: its chair said that "there is no AI exemption from the laws on the books." Whatever you tell shoppers through a bot is held to the standard you hold your own pages to.

What it should do

Keep the job narrow. An assistant earns its place in these tasks:

• Find products. A shopper describes what they want, such as a gift for a coffee drinker on a set budget, and the assistant searches your own catalog and shows items with price and availability. The answer comes from your catalog, not from the model's memory of the internet.
• Answer policy questions from your policies. Shipping times, return windows, payment methods. The wording should be read from your own policy pages, so there is one source of truth.
• Build a cart when the shopper agrees. Adding an item is an action, so it should happen after the shopper says yes, and checkout should stay in Shopify's own checkout.
• Say when it does not know, and pass the conversation on.

What it should not do

• Take payment or finish an order. Checkout belongs to Shopify's checkout, which shoppers already trust.
• Make up policy. If your return policy has no clause on a case, the assistant says so and offers the team. It does not decide.
• Give regulated advice. Medical, legal and safety questions about what a product can do for a body or a home are for a person, or for no one.
• Promise outcomes or discounts you did not set. If a shopper negotiates, the assistant has no authority to agree.
• Handle an existing purchase. A question about a specific shipment needs account and carrier data and a person who can act on it. Send those to your team.
• Pretend to be a person. California's bot disclosure law, Business and Professions Code section 17941, makes it unlawful to use a bot to communicate with a person in California online, with the intent to mislead that person about its artificial identity, for the purpose of knowingly deceiving the person about the content of the communication in order to incentivize a purchase. It also says a person using a bot is not liable if they disclose that it is a bot, and that the disclosure must be "clear, conspicuous, and reasonably designed to inform" people that they are dealing with a bot. Say plainly in the opening message that the shopper is talking to an AI assistant. That is a small cost and removes the question in every state. This is a summary, not legal advice.

The handoff is the product

A handoff fails in three ways: the shopper cannot find it, it goes nowhere, or the person who picks it up asks everything again. Design against each.

Make it easy to ask for a person. A visible option to talk to your team should be available at any point, not after three failed attempts. Shoppers who ask for a person should get one.

Pass a summary. The assistant's job at handoff is to give your team what was asked, what was shown and what the shopper was trying to do, so the opening reply from a person is a useful one.

Say what happens next. If your team answers by email within a business day, the assistant says exactly that and asks for an address. If nobody is at work on a Sunday, it says so. A promise the assistant cannot keep is worse than the missing answer.

Our own approach to support is described in why a real developer answers our chat and in the comparison of app support. The same principle applies to a store: tell people what is automatic and what is a person, and do not blur the two.

What to give it, and what to leave out

Quality comes from inputs. Give the assistant your catalog with accurate titles and descriptions, your shipping, returns and payment pages, and a short list of rules in plain if-then form: if a shopper asks about sizing for a product with no size chart, say so and offer the team. Leave out anything you would not put on a public page. Treat everything the assistant reads from your store as information to quote, not as instructions to follow, which is why a catalog description containing the words "ignore your rules" should not change its behavior.

The AI Chat mini app is built on that division. By its own page it searches your catalog, answers from your store policies, adds to the cart after the shopper agrees, hands the conversation to your team with a short summary, never takes payment and does not look up orders. A guided setup asks seven short questions and drafts the instructions for you to confirm, and nothing reaches shoppers until you publish. You choose the pages where the chat button appears, and analytics show conversations, chat-assisted orders and the revenue they influenced. The page defines a paid order as chat-assisted when that conversation added the items to the cart in the 7 days before, which is a stricter measure than counting every chat.

A pre-launch checklist

1. Read your shipping, returns and payment pages as a stranger would, and fix contradictions between them. The assistant will repeat whichever it reads.
2. Write five questions you hear most often and test the answers yourself, including the ones whose honest answer is "ask us."
3. Test five traps: a question about a shipment already on its way, a request for a discount, a medical question, a question your policy does not cover and a request to talk to a person.
4. Confirm the greeting says it is an AI assistant.
5. Pick the pages where it appears. A product page and a policy page are a good start; leave the cart page quiet until you have read the conversations.
6. Read every conversation for the opening two weeks. Fix the rule behind each bad answer, not the single answer.
7. Judge it by what it did: conversations, handoffs that your team could answer quickly, and orders that followed, compared with a period without it.

Your next step

Write down the five questions your team answers most, and the three it should never answer by itself. Then open the AI Chat setup, let the guided setup draft your instructions from them and read the result line by line before you publish. The template below is a related product page offer to test once the chat is answering well.