AI for online shops: automated support, orders and returns

In an online shop, most customer messages are the same: "Where's my parcel?", "Do you have it in a medium?", "How do I return this?", "When will I get my money back?". Each one takes two minutes. Multiplied by dozens or hundreds a day, they become a full-time job. And at busy times, exactly when you are selling the most, replies slow down and customers get annoyed.

An AI system for your online shop takes over these repetitive conversations, using real data from your shop, and leaves people only the cases that genuinely need them. Below you will see what it can actually do, how it connects to your platform and what you need to bear in mind with returns.

What AI can take over in an online shop

Order status

The customer types their order number or phone number, and the agent replies with the real status: confirmed, being prepared, handed to the courier, with the AWB tracking number and tracking link. The data comes straight from your shop platform and the courier, not from guesswork.

Product questions

Sizes, materials, compatibility, stock, delivery times. The agent answers from your product pages and can recommend alternatives when something is out of stock. At Neurotech we use Claude, the AI model developed by Anthropic, which handles freely worded questions well, typos included.

Returns and exchanges

The agent checks whether the order is eligible, explains the steps, generates the return form or label (if your platform and courier allow it) and logs the request. Your team sees returns in a list, not in dozens of emails.

Order changes

Changing the address or payment method, or cancelling, as long as the order has not yet been dispatched. The agent checks the status and either makes the change or sends it to a person for approval.

Recovering abandoned baskets

With automation, a customer who left items in their basket gets a helpful message: an answer to a likely question, delivery information or help completing the order. No pressure and no spam, with marketing consent respected.

Confirmations for cash-on-delivery orders

For cash-on-delivery orders, an automated confirmation message or call before dispatch reduces refused parcels. An AI voice agent can call, confirm the address and that the customer will be available, and record the result in the system.

Recommendations and upselling

An agent that knows what the customer bought can naturally recommend the complementary product: the phone case, the shoe care kit, the replacement filter. Not a generic banner, but a reply that fits the conversation. As a real benchmark from our portfolio: for a beauty salon, a custom app together with the website brought in around 40% more revenue from upselling. The sector is different, but the principle is the same: the right offer at the right moment.

Which channels it works on

• Website chat, for pre-purchase questions.
• WhatsApp, the preferred channel for many customers for order status and returns. A WhatsApp chatbot picks up conversations and continues them at any time.
• Email, where AI can sort messages and reply to the standard ones.
• Phone, for customers who prefer to call.

What matters is that they all use the same data and the same rules. The customer gets the same answer wherever they ask.

Returns: what the agent needs to know

In Romania, as a general rule, online buyers have the right to withdraw from the contract within 14 days without giving a reason, under Government Emergency Ordinance (OUG) 34/2014. There are exceptions set out in law, and the exact details (for example, when the period starts and which products are excluded) should be checked with your lawyer or adviser.

For an AI agent, this means:

• A clearly written, legally checked returns policy becomes the source the agent answers from.
• The agent does not invent exceptions or refuse a return on its own. Borderline cases go to a person.
• The refund times given to the customer are the ones in your policy, not estimates.
• Every return request is logged with its date, so you can later prove what happened.

How it connects to your platform

The agent needs access to data: orders, products, stock, delivery statuses. Most online shop platforms have integration interfaces, and courier companies usually offer parcel tracking through integration. We connect the agent to these sources through n8n, and conversations and requests are saved in a Supabase database on servers in the European Union.

If your shop runs on a custom-built platform, or you want a new shop, we can also build the website side with integration designed in from the start.

What implementation looks like

1. Analysing conversations. We look at your last few hundred messages and group them by type. Usually a handful of categories cover most of the volume.
2. The knowledge base. Delivery, returns and warranty policies, the size guide, frequently asked questions. You approve every answer.
3. Integrations. Your shop platform, the courier, possibly your invoicing software.
4. Handover rules. What must always go to a person: complaints, faulty products, angry customers, large amounts.
5. Testing. Your team plays the role of difficult customers. We adjust.
6. Launch on one channel, then expand.
7. Weekly monitoring in the first month: which questions did not get a good answer and what needs adding.

What a conversation sounds like

Customer: Hi, I ordered a jacket yesterday and I'm not sure I put the right address.

AI agent: Hi! Could you give me your order number or the phone number you ordered with, please?

Customer: 0722...

AI agent: I've found your order. The address on it is Strada Observatorului, Cluj-Napoca. The order is still being prepared, so we can change it. Would you like to change anything?

The example is illustrative, but it shows the idea: the agent identifies the customer, sees the real order and only acts while it is allowed to. If the parcel had already left, it would have explained the courier's options to the customer.

What to measure after launch

• The percentage of conversations resolved without a person stepping in.
• Average time to first reply.
• The number of refused cash-on-delivery parcels, before and after confirmations.
• The reasons conversations end up with a person.

What it costs

Payment is made in instalments tied to deliverables. Monthly costs (infrastructure, usage, maintenance) depend on the volume of conversations.

The return calculation starts from three figures: how many hours a day your team spends on support, how many cash-on-delivery parcels are refused each month and how much abandoned baskets are worth. Even a small reduction in refused parcels or a few recovered baskets a week quickly shows up in the numbers.

Common mistakes

• A bot with no access to data. A chatbot that just repeats FAQs without seeing the customer's order annoys more than it helps.
• No route to a person. A customer with a faulty product wants to talk to someone. Handover must be quick and include the full conversation history.
• Vague policies. If your returns policy is unclear to people, it will be unclear to the AI too. Clarify it first.
• Too many marketing messages. Recovery automations should be few, useful and sent with the customer's consent.
• Launching just before peak season without testing. Put the system live a few weeks before your sales peaks, not on the day.

A quick checklist

• Do you have clearly written returns and delivery policies?
• Does your platform support integration (orders, products, stock)?
• Does your courier provide tracking data through integration?
• Do you know which types of message you get most often?
• Have you decided who picks up the cases the AI hands over?

Next step

Want to see how much of your shop's support could be handled automatically? Book a free 30-minute call. We will look together at the types of message you receive and tell you honestly where it is worth starting.