AI agents for ecommerce: after-sales, repurchase, abandoned carts and 24/7 support

In short: an AI agent for an online store is a program that talks with your customers over web chat, SMS, email, Instagram or WhatsApp and, unlike a scripted bot, looks up and operates your store: it finds the order, reads the shipping status, starts an exchange, builds an offer and sends the payment link. The four jobs where it pays off most are after-sales (where is my order, exchanges, returns, reviews), repurchase (replenishment, cross-sells and personalized offers), abandoned carts and after-hours customer service. What decides whether it works is not the AI model but how well it is connected to your catalog, your stock and your shipping.

This article walks through each of those jobs with example conversations, the data the agent needs, what you should not automate, how to measure it and what it costs.

What an AI agent for ecommerce is and what it is not

Three things with similar names coexist in the market:

All three are useful. An agent does not replace the other two; it organizes them. The practical rule for choosing: if the answer depends on data from an order or a customer, you need an agent; if it is the same for everyone, a fixed answer is enough.

An AI agent for an online store has three parts:

1. The channel: your website chat, SMS, email, Instagram or WhatsApp.
2. The AI model: the one that understands the message and decides the next step.
3. The tools: the connections to your store (Shopify, WooCommerce, BigCommerce or a custom ecommerce), your payment processor, your shipping carrier and, if you have one, your ERP or inventory system.

The third part does the work. An agent with no connection to your store can only repeat your FAQ, and for that something cheaper is enough.

1. After-sales: the job that frees the most hours

In an online store, most messages arrive after the purchase. The most common question in any industry is some version of "where is my order?"

Where is my order

What the agent does: it identifies the customer by phone number, email or order number; looks up the order in the store; checks the status with the carrier; and answers with the real data, not a template.

An example conversation:

Customer: hey, I ordered running shoes on Monday and nothing has shown up

Agent: Hi Jessica. I see your order 4812 from Monday: one pair of Trail running shoes, women's size 8. It left our warehouse on Tuesday and the carrier has it at the distribution center near you. The estimated delivery date is tomorrow. Do you want me to text you when it goes out for delivery?

What data it needs: orders and customers from the store, the tracking number and access to the carrier's API (or to a shipping platform that groups several carriers).

When it hands off to a person: if tracking says delivered and the customer says they did not get it, if the delay passes a set number of days, or if the customer asks to talk to someone.

Exchanges and returns

What the agent does: it checks that the order is within the return window, asks the reason, checks whether the new size or color is in stock, generates the return label or points to a drop-off location and records the case.

The rule that matters: the agent applies your policy; it does not interpret it. If the policy says "30 days, unworn, with tags," the agent asks exactly that. Exceptions (defective product, repeat customer, gift) are written as rules or handed off.

What not to automate at first: cash refunds. The agent can prepare the refund so a person approves it with one click. When volume and track record justify it, you can enable automatic refunds under a certain amount.

Reviews

What the agent does: a few days after delivery, it asks how the product arrived. If the answer is positive, it sends the link to leave a review. If it is negative, it does not ask for a review: it opens a case and treats it as after-sales.

That detour is the difference from a mass "leave us a review" email: the unhappy customer gets attention before writing publicly, and the happy one gets the review request at the right moment.

2. Repurchase: selling again to someone who already bought

Selling again to a customer who already trusts you usually costs less than winning a new one with advertising. The problem is that doing it well by hand requires looking at each customer's history, and nobody has the time.

Replenishment

For products that run out (pet food, supplements, coffee, cosmetics, supplies, filters), the agent estimates when each customer should be running low and writes before that happens.

Agent: Hi Mark, 28 days ago you picked up the 30 lb bag of adult dog food. If it lasts like the last one, you have a few days left. Want me to set up the same order? It is $54.90 today with free shipping. I will drop the payment link right below.

If the customer answers "yes, but the 15 lb bag," the agent switches the variant, recalculates and sends a new link. If the customer says "already bought," it records that and adjusts the next reminder.

Cross-sells

The agent recommends what complements the previous purchase, using rules you define: after a coffee maker, the filters and the coffee; after a mattress, the pillow; after a bicycle, the helmet. The difference from a generic email is that the recommendation comes from the customer's real order and respects what they already bought.

Personalized offers with a payment link

When a customer asks about something you do not have, or hesitates over the price, the agent can offer an alternative within margins you set (for example, up to 10 percent off a second purchase, or free shipping above a certain amount) and send the payment link right in the chat. The customer pays without browsing the store again.

What data it needs: order history per customer, catalog with stock, discount rules and the ability to generate payment links from your processor (Stripe, PayPal, Shop Pay or another) or prebuilt carts in the store.

The limit: messages you start are subject to channel rules. On WhatsApp, outbound messages are marketing templates approved by Meta and charged per message, and the customer has to have opted in. On SMS, marketing messages generally require the customer's prior express consent under the TCPA. This is not legal advice, so confirm the specifics with your lawyer or your SMS provider. An agent that writes too much burns budget and gets blocked. The WhatsApp rules are in AI agent for WhatsApp: what it handles.

3. Abandoned carts

The classic automation sends a fixed sequence: a reminder after an hour, another after a day, a coupon after three days. It works, and many platforms include it.

What an abandoned cart AI agent adds is the conversation when the customer replies. The most common reasons for abandonment are unanswered questions:

• "How long will it take to get to my zip code?"
• "Will this charger work with my model?"
• "Can I pay in installments?"
• "What if it does not fit?"

A fixed message cannot answer that. An agent can, because it checks the shipping cost by zip code, the compatibility on the product page, the financing options currently offered and the return policy.

The last row is usually the most valuable in the medium term: if 30 percent of those who reply ask about shipping, the problem is not fixed with messages, it is fixed by showing the shipping cost earlier in the store.

4. Customer service around the clock

A large share of pre-sale questions arrive at night or on weekends, when nobody is there to answer. A shopper who does not get a reply within minutes often buys somewhere else. This is the case for an AI agent for an online store working as your after-hours team, and for AI customer service on Shopify or any other platform it can connect to.

The agent handles pre-sale questions using your catalog as the source:

• Availability of size, color or variant.
• Total price with shipping to their address.
• Payment methods and financing.
• Comparisons between two products in your store.
• Recommendations based on what the person says ("it is for my dad, who runs marathons").

And it closes: it builds the cart or sends the payment link.

Which channels fit a US store

• Web chat. The default for most stores. It runs on your own site, so there is no third-party approval process, and it can read the cart the shopper is looking at.
• SMS. Good for order updates and replies to customers who texted you first. For outreach, remember the consent point above.
• Email. Fits slower, longer conversations such as returns and after-sales, and it is where most stores already receive their support requests.
• Instagram. The connection goes through Meta's official API with a professional account, and the store can reply within a time window after the customer's last message. It works well for answering direct messages and comments that ask for a price; it is not meant for cold outreach to strangers.
• WhatsApp. Worth adding if your store has Hispanic or international customers who already use it as their default way to message a business. It requires the official WhatsApp Business API through Meta or an authorized provider. Since January 15, 2026, Meta does not allow general-purpose AI assistants on that API, but it does allow business agents that support, sell and send notices about their own products, which is exactly this case.

What to automate and what not to

Handoff is part of the design, not a failure. A good agent hands off with a summary of the case, the order identified and what was already tried, so your team does not start from zero.

What needs to be ready in your store

Before building the agent, check four things. If any of them fails, the agent will repeat the mistake faster.

1. A catalog with complete attributes. Sizes, dimensions, compatibility, materials. If the answer is not on the product page, the agent does not have it.
2. Reliable stock. If you also sell in a physical location or on a marketplace, stock has to be synced. An agent that offers something sold out creates the worst kind of complaint.
3. Written policies. Return window, conditions, who pays for return shipping, delivery times by region. If they depend on "whatever the manager says," write them down.
4. Access to your systems. Shopify, WooCommerce and BigCommerce all have APIs for orders, products and customers. The large carriers do too. If your stock lives in a spreadsheet, fix that first.

These are the same points that shopping agents in ChatGPT or Gemini look for before recommending your store. The full list is in how to prepare your store for AI shopping agents, and the bigger picture is in what is agentic commerce.

How to roll it out without risk

An order that works for most stores:

1. Week 1: after-sales in read-only mode. The agent answers "where is my order" and pre-sale questions. It does not take actions. Your team reviews the conversations every day.
2. Weeks 2 and 3: actions with confirmation. The agent starts exchanges and prepares refunds, which a person approves.
3. Week 4: repurchase and carts. Outbound messages are turned on for a small group of customers who agreed to receive them, and you measure.
4. After that: expand. More types of inquiries, more channels, more autonomy where the numbers justify it.

Conversation review never ends: every week during the first month, then monthly. It is how you find weak answers before your customers do.

How to measure whether it works

Measure before you start. Without a baseline (how many after-sales messages you get per week, how long your team takes, how many carts you recover today) you will not be able to tell whether the agent changed anything.

What it costs

Two paths, with market reference ranges as of October 2026:

On top of that come the running costs: usage of the AI model (it grows with conversation volume), any messages your store starts on WhatsApp or SMS, and hosting. Those costs depend mostly on conversation volume and how many messages the store starts, and they are worth estimating from your real numbers, not an average. If you are weighing the wider picture across your business, AI agents for business: what they actually automate is a good companion, and so is software integration cost if the hard part is connecting systems.

If your store runs on Shopify with no unusual integrations, start with the platform's apps. A custom agent is justified when you need to cross data that no app crosses (your ERP, your own shipping, customer-specific pricing) or when volume makes paying per conversation cost more than building.

Common mistakes

• Connecting the agent to a messy catalog. The agent does not fix data: it repeats it.
• Giving it freedom with discounts. Without a written cap, the agent gives in to any customer who insists.
• Messaging every customer. Repurchase works with people who agreed to receive messages and at the moment it makes sense. The rest ends in blocks, complaints and, on WhatsApp, an account with a bad reputation.
• Hiding that it is an agent. Say so in the first message and always offer the option to talk to a person.
• Not reviewing conversations. Most problems show up in the second and third weeks, with cases nobody anticipated.

Frequently asked questions

What can an AI agent do in an online store?

It can answer order status, handle exchanges and returns within your policy, ask for reviews, recommend products, recover abandoned carts, notify customers when it is time to reorder and send payment links, over web chat, SMS, email, Instagram or WhatsApp, around the clock.

Does an AI agent work with Shopify, WooCommerce and BigCommerce?

Yes. All three platforms have APIs for orders, products and customers, which is what the agent needs to query. Custom-built ecommerce sites also work, as long as they have an API or an accessible database.

What is the best AI customer service option for Shopify?

It depends on your volume and how much is unusual about your operation. For a small store with standard flows, start with a Shopify app. If you need the agent to cross data from your ERP, your own shipping or customer-specific pricing, a custom agent connected to the Shopify API is the better fit.

Will an AI agent replace my support team?

Not entirely. It takes the repetitive questions that are answered with data, and your team handles the claims, the exceptions and the sales that need judgment. In practice, the same team handles more orders.

How does an abandoned cart AI agent recover carts?

It sends the reminder like any automation, but when the customer replies with a question (shipping, size, financing, compatibility) it answers using your store data and resends the cart or a payment link. It also records the reason for abandonment.

Can it take payment inside the chat?

It can send a payment link generated by your processor (Stripe, PayPal, Shop Pay or another) or a cart already built in your store. The payment happens in the processor, not in the chat.

What happens if the agent makes a mistake?

That is why actions involving money (refunds, discounts) have caps or human approval at first, answers come from your data and not from the model's memory, and conversations are reviewed every week.

Can I text customers with an agent?

Yes, but SMS marketing generally requires prior express consent from the customer under the TCPA. Replying to a customer who texted you first is different from starting outreach. This is general information, not legal advice, so check your obligations with your lawyer or SMS provider.

How long does it take to be up and running?

The working demo is built in 24 to 48 hours. A first production version with after-sales usually takes two to four weeks, depending on how many systems need to be connected.

If you want to see an after-sales and repurchase agent answering with your store's data, we build a free working demo in 24 to 48 hours. You can book a call or see how we build AI agents for business. You can also see how the AI customer service agent for ecommerce and the AI repeat purchase agent work.