AI customer service for online stores: what to automate, and when a human must answer

Most messages to an online store ask the same few things. Where is my order? Can I return this? Will it fit? How long does delivery take to Germany? AI customer service for ecommerce is good at exactly those questions, provided it answers from your real order data and written policies, and passes everything else to a person without a fight.

This guide is for store owners on Shopify, Amazon and other channels whose inbox is growing faster than their team, whether you are weighing a Shopify AI chatbot app, a helpdesk add-on or a custom agent. It covers what to automate, what a human should keep, how to ground answers in store data, how to design the hand-off, which channels to use, how to measure the result and what drives the cost. Third-party figures are as of October 2026. If you are still deciding what to automate across the whole business, start with my e-commerce automation guide.

I build AI agents, WhatsApp and Telegram bots, and automation for Shopify stores and marketplace sellers. My own site has an AI assistant called Wren, which I use as an example below.

What AI customer service for ecommerce handles well

The best candidates are frequent, factual and answerable from data you already hold:

• Order status. An order status chatbot answers "Where is my order?" from live order and tracking data, after a simple identity check such as the order number plus the email address on the order.
• Returns and exchanges. Whether an item can be returned, how long the customer has, who pays postage and how to start. If you use a returns portal, the AI can send the customer straight into it.
• Product questions. Sizes, materials, care, compatibility and what is in the box, answered from your catalogue rather than guessed.
• Shipping. Delivery times and costs by country, order cut-off times and which carriers you use.
• Stock. Whether a variant is available, and taking a back-in-stock request.
• Collecting details for a person. Photos of a damaged item, the order number and what the customer wants, so your team starts with a complete case.

What a human should answer

Some conversations should reach a person quickly, with the AI's help limited to gathering details:

• Damaged, wrong or missing items when the customer is upset, and anything that could become a dispute or chargeback.
• Refunds, credits or exceptions outside your written policy.
• Complaints and anything emotional.
• Safety, allergy or health questions, where a wrong answer could hurt someone.
• Wholesale, bulk or custom orders.
• Anything the AI cannot answer from your data. A confident guess costs more than a short wait.

Customers agree. In a survey of 3,566 customers published on 4 August 2026, Gartner found that 87% say companies using generative AI for customer service must provide access to a human agent. The most common thing that would change a reluctant customer's mind was being able to switch to a person if needed, and Gartner's analyst advised against using AI as "a mandatory first step for every issue".

Gartner is also confident about automation: in March 2025 it predicted that agentic AI will resolve 80% of common customer service issues without human intervention by 2029. The two findings fit together. Automate the common questions well, and keep the door to a person open for everything else.

Grounding answers in your store data

A store's support AI needs two kinds of knowledge, and they are handled differently.

Fixed knowledge is your returns policy, shipping table, size guides, care instructions and FAQ. Write it down once, keep it current, and tell the AI to answer only from it. When a policy changes, you edit the source and the AI follows.

Live data is orders, tracking, stock and return status. The AI reads it through your platform's API at the moment of the question, such as Shopify's Admin API or Amazon's Selling Partner API, with read-only access. It never infers an order status from the conversation.

A few rules keep this safe:

• Verify before revealing. No order details until the customer has matched an order number with its email address or postcode.
• Take numbers from the source. Return windows, fees and delivery times come from the written policy, word for word.
• Admit gaps. If the answer is not in the data, the AI says so and offers a person.
• Treat marketplaces separately. Amazon buyer messages run inside Amazon's own messaging system and rules. A drafting assistant, with a person sending, is the cautious route there.

Wren, the assistant on my own site, works on the same principle. It answers only from the content the site publishes. If the answer is not there, it says it does not know and points the visitor to WhatsApp, email or the contact form, so they reach me directly. It does not invent prices, dates or promises. A store's assistant should behave the same way with orders and policies.

Designing the hand-off

The hand-off is where most support bots lose customers' trust, so design it first, not last.

• A visible way out. Offer "talk to a person" in every conversation, not only after the AI has failed three times.
• Clear triggers. Hand over automatically on words like refund, damaged or chargeback, when the customer asks the same thing twice, when the AI has no source for an answer, or above an order value you choose.
• Context travels with the case. A short summary, the order number, what the customer wants and what the AI has already said. The customer never repeats themselves.
• Honest timing. Out of hours, the AI tells the customer when a person will reply, using your real working hours, and never promises more.
• Drafts for staff. On slower channels, the AI can draft a reply that a person checks and sends. It is a good halfway point while you build trust in it.

Channels: website chat, WhatsApp, email and apps

• Website chat catches questions before purchase, such as sizing or delivery dates, when the answer decides the sale.
• WhatsApp is where many customers already message businesses. It has its own rules on when a business may message first and on pre-approved templates. I build WhatsApp and Telegram bots, using Meta's official platform for WhatsApp, and the same AI can sit behind them.
• Email suits longer, slower questions and works well with drafts a person approves.
• Your own app, if you have one, can host support with the customer already signed in, which removes the identity check. See building an app if that is on your roadmap.

Build the knowledge, rules and hand-off once and attach each channel to them, so a policy change reaches every channel at the same moment.

Measuring whether it works

AI customer service for ecommerce is easy to launch and hard to judge by feel, so decide the measures before launch and look at them weekly:

• Resolution rate, defined strictly. The customer got an answer and did not contact you again about the same thing within a few days.
• Hand-off rate and quality. How many cases go to a person, and whether they arrive complete.
• Time to first reply and time to resolution, compared with before.
• Wrong answers found in review. Read a sample of conversations every week. Each wrong answer points to a source document to fix.
• Satisfaction, measured separately after AI and human conversations.
• Cost per resolved conversation, including model usage, so you can compare options honestly.

What it costs: per-resolution tools vs a custom agent

Helpdesk tools now commonly price AI by outcome. At the time of writing, Intercom's Fin pricing page lists $0.99 per outcome, where a resolution means no further help was requested after Fin's last answer. Gorgias lists AI Agent interactions at $0.90 each on annual plans or $1.00 on monthly plans, with an allowance included per plan and overage at $1.50. Prices change, so check the pages yourself.

Cost per AI resolution is easy to model: multiply it by your monthly resolutions. It suits a store that already runs on that helpdesk and mostly gets standard questions.

A custom agent makes more sense when:

• order data sits in several channels, such as Shopify, Amazon and TikTok Shop, and one answer needs all of them;
• you want WhatsApp on your own number, with your own rules;
• your returns or warranty rules are unusual;
• your volume makes per-outcome fees a large monthly line.

A custom agent costs a build plus running costs: model usage, hosting and maintenance. The build depends on the systems and channels it connects to, the languages it handles and how many actions need approval. I break those drivers down in what an AI agent costs to build. I quote AI agents and bots from a short brief rather than a price list. If the hard part is getting order and stock data out of several channels in the first place, that is integration work of the kind on my e-commerce automation page.

FAQ

Can an AI agent issue refunds on its own?

It can be built to, but I would not start there. Refunds move money and are hard to undo, so the safer design is for the AI to collect the details, check them against your written policy and prepare the refund for a person to approve with one click. Once it has produced accurate drafts for a while, you might let it handle small, clear-cut cases within a limit you set, with every action logged.

Will customers be annoyed talking to an AI instead of a person?

Some will, if it blocks them. Gartner's August 2026 survey found that 87% of customers say companies using generative AI for service must offer access to a human, and the option to switch to a person was the most common thing that would make reluctant customers try AI. Answer common questions instantly and show the route to a person from the first message, and customers who just want a tracking update get it faster.

Which customer service questions should an online store automate first?

Start with order status. It is often the most frequent question, and it can be answered read-only from live order data after a simple identity check. Next come returns policy, shipping times and costs, and product facts such as sizing, all answered from written sources. Leave disputes, damaged items, exceptions and anything emotional to a person, with the AI collecting the details first.

If you want AI answering your store's customers, send me a short brief: the channels customers use to reach you, the platforms that hold your orders, your ten most common questions and who handles hand-offs today. Send it through the contact form, see the AI agents for business service, or read how a WhatsApp chatbot for business works if that is where your customers are.