Most support inboxes are the same twenty questions, asked hundreds of times. Where is my order. How do I return this. Do you ship to Canada. Which size. An AI agent can answer those in seconds, at 3 a.m., in the customer’s language, and leave the hard cases to a person.
A bad one does the opposite: confident wrong answers, loops, and no way out. The difference is almost entirely in the setup.
What a good agent does
| It should | It shouldn’t |
|---|---|
| Answer from your written policies, product pages and FAQs | Make up policies or guess |
| Look up a customer’s order status and tracking | Show order details without checking it’s the right customer |
| Explain returns, sizes, shipping times and product use | Promise refunds, discounts or delivery dates you don’t offer |
| Collect the details of a problem (order number, photo) | Argue with an upset customer |
| Hand over to a person, with the whole conversation | Trap people in a loop with no way out |
| Say it’s an AI assistant | Pretend to be a person |
What it needs
1. Your knowledge, written down. Shipping, returns, warranty, sizing, product care, payment and the questions customers really ask. If a policy only lives in someone’s head, the agent can’t follow it. This step alone usually improves human support too.
2. Live data. To answer “where is my order”, the agent needs read access to orders and tracking, through Shopify’s Admin API, your help desk or your 3PL. Before showing anything, it checks the customer is who they say: the email on the order plus the order number, or a signed-in session.
3. Limits. A written list of what it may do alone, what needs a person’s approval, and what it must never touch. Money, account changes and anything legal go to a person at first.
4. A hand-off. One tap to reach a person, and a rule that the agent hands over by itself when it isn’t sure, when the customer asks, or when the customer is upset. The person should see the whole conversation, so the customer never repeats themselves.
5. Its voice. The same tone as your brand: short, warm, plain. Write a few model answers and let the agent learn from them.
The rules
Say it’s AI
The EU AI Act’s transparency rules have applied since 2 August 2026. AI systems that interact directly with people must be designed so people know they’re dealing with an AI, unless it’s obvious from the context. A clear first line (“I’m the store’s AI assistant; you can ask for a person at any time”) covers it, and it’s good practice in every market.
You own what it says
In Moffatt v. Air Canada (2024), the British Columbia Civil Resolution Tribunal held Air Canada responsible for what its website chatbot told a customer about bereavement fares, and rejected the idea that the chatbot was responsible for its own words. Assume the same applies to your store: the agent speaks for you. That’s why it answers from your written policies and hands over when unsure.
Data and privacy
- Give the agent only the data it needs to answer. Read-only order access is enough for most stores.
- Tell customers in your privacy policy that an AI tool handles support chats, and which provider runs it.
- In the EU, people have the right not to be subject to decisions with legal or similarly significant effects made only by automated means (GDPR Article 22, with exceptions such as a contract or explicit consent). Keep a person in the loop for decisions like refusing a refund or closing an account.
- Pick providers that don’t train their models on your customers’ conversations, and that can keep data in the region you need.
Don’t overclaim
Don’t sell the agent as something it isn’t, to customers or in your own marketing. The FTC has acted against businesses that made deceptive claims about AI.
Rolling it out
- Read the inbox. Export the last three months of tickets and sort them by question. The top ten usually cover most of the volume.
- Write the answers. One clear page per topic. Fix the policies that turn out to be unclear.
- Start in draft mode. The agent writes replies, a person approves and sends them. Correct it where it’s wrong.
- Go live on the easy topics. Order status and simple policy questions first, with the hand-off always on.
- Read the conversations every week. Look for wrong answers, missed hand-offs and new questions. Update the knowledge, not just the prompt.
- Widen slowly. Add actions (address changes, small refunds) one at a time, with limits.
Measure it on what matters: how many conversations it resolves without a person, how many it hands over, and the customer rating of both.
Off the shelf or built for you
Ready-made agents are quick to switch on and good for stores whose questions are mostly order status and standard policies. Shopify Inbox, Shopify’s free chat app, has an opt-in AI agent that chats with shoppers on your store. Help desks such as Gorgias and Zendesk sell AI agents that connect to Shopify orders. (Shopify’s Sidekick is different: it’s an assistant for the merchant in the admin, not for customers.)
A custom agent is worth it when the answers live in several systems (Shopify plus a 3PL, a subscription app, an ERP or a B2B price list), when you need the same agent on the site, email and WhatsApp, or when the off-the-shelf tools can’t follow your rules. We build those on your own stack, trained on your brand, with the hand-off and limits above.
Questions
What can an AI support agent answer for an online store?
The questions that fill most inboxes: where is my order, how do returns work, do you ship to my country, what size should I pick, is this in stock, how do I use the product. It answers from your own policies, product data and order system, and hands everything else to a person.
Do I have to tell customers they are talking to AI?
In the EU, yes: since 2 August 2026 the AI Act requires AI systems that talk with people to make clear they are AI, unless it's obvious. Elsewhere it's still the right call. Name the agent as an assistant, say it's AI in the first message, and offer a person at any time.
Is my store responsible for what the AI says?
Treat it as if you are. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline responsible for wrong refund information its website chatbot gave a customer. Keep the agent to your written policies and have it hand over anything it isn't sure about.
Should the AI give refunds or discounts on its own?
Not at first. Let it explain the policy and prepare the case, and let a person approve money. Once you trust it, allow small, rule-based actions, like a refund under a set amount on a damaged-item claim with a photo.
Do I need a special app, or can it be built?
Both work. Help desk tools for Shopify, such as Gorgias and Zendesk, sell AI agents you switch on. A custom agent makes sense when the answers need your own systems, like a 3PL, a subscription app or a B2B price list, or when you want it in email, WhatsApp and the site at once.
Sources
- EU AI Act Service Desk: Article 50, transparency obligations
- Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII)
- GDPR Article 22: Automated individual decision-making
- Shopify Help Center: Shopify Inbox
- Gorgias Help Center: Intro to AI Agent
- Zendesk: AI agents
- FTC: Crackdown on deceptive AI claims and schemes (Operation AI Comply)