Every webshop gets the same questions all day long. Where is my order? Can I exchange this for a larger size? Do you ship to Ireland? A customer service chatbot exists to answer those questions the moment they are asked — at two in the morning as readily as at two in the afternoon.
What has changed in the last few years is not the idea but the quality. The old generation of chatbots followed a decision tree and fell over the moment a visitor phrased something unexpectedly. Modern AI chatbots read the question in plain language and answer from your own content. That is a different product wearing the same name.
This guide explains what a customer service chatbot is, what an AI agent can realistically handle, how to build one on your own shop's information, what it costs, and — the part most vendors skip — where the AI should stop and fetch a human colleague.
What is a chatbot for customer service?
A customer service chatbot is software that holds a conversation with your customers and answers their questions without a person typing the reply. It usually lives in a chat widget in the corner of your website, so the visitor never leaves the page they are on.
The useful distinction is not chatbot versus no chatbot, but how the bot decides what to say:
- Rule-based chatbots follow a script you draw out in advance. Button, button, answer. Predictable, but it only covers the paths you imagined.
- AI chatbots use a language model to interpret free-typed questions. They handle phrasing you never anticipated, including typos and half-sentences.
- An AI agent goes one step further: as well as answering, it can perform a task — look up an order, check a delivery status, open a support ticket — and then continue the conversation with the result.
Most shops today want the third kind, because a large share of customer queries are not really questions about policy. They are requests for a specific fact about one specific order.
How AI customer service chatbots use natural language processing
Under the bonnet, AI chatbots use natural language processing to turn a sentence into meaning rather than matching keywords. The important consequence for you is practical: you no longer maintain a list of trigger phrases. You maintain the underlying information, and the AI works out which part of it answers the question in front of it.
The current generation is built on generative AI, so it composes an answer in its own words instead of pasting a canned block of text. That is a large improvement in how a conversation reads, and it introduces one risk worth taking seriously: a model that composes fluently will also compose confidently when it is wrong. Anything you deploy should be grounded in your own content and should be willing to say it does not know. We go deeper into that in training a chatbot on your own data.
What an AI chatbot can actually handle
In an online shop, the realistic list of what chatbots handle well is longer than people expect, because so much of the daily inbox is repetition. Typical work an AI customer service chatbot takes off your desk:
- Delivery times, shipping costs, and which countries you ship to.
- Return and exchange windows, and how to start a return.
- Order status, once it can read your order data.
- Payment methods, invoices, and VAT questions.
- Product details — materials, sizing, compatibility, care instructions.
- Stock and restock questions, and pointing to an alternative product.
These are the common customer questions that arrive dozens of times a week and take a human two minutes each. Handing them to AI is not about replacing your customer service team; it is about giving them back the hours those two minutes add up to. Our article on chatbots for e-commerce goes through these shop-specific cases in detail.
Where the AI should hand over
An honest answer to what AI cannot do matters more than an ambitious one, because every over-promised chatbot ends up damaging the customer experience it was bought to improve.
Hand over to a person when the conversation involves a complaint with feeling in it, a decision outside your published policy, a refund that needs approval, a damaged or missing parcel, anything legal, or simply a customer who asks for a human. Also hand over on the quieter signal: the AI is unsure. A chatbot that says "I am not certain about this one, let me get a colleague" keeps trust intact. One that guesses spends it.
Veyra is built around that handover — the assistant answers what it can support from your content and escalates the rest, with the full conversation attached so the colleague does not ask the customer to repeat themselves. The mechanics of doing this well are in chatbot to human handover.
Chatbot or live chat — do you need both?
You need both, and in practice they are the same window. Live chat is a human typing in real time. A chatbot answers instantly but within limits. Run them separately and you get the worst of each: a bot with no exit, or a live chat inbox that is empty at 23:00 when half your orders are placed.
Run them as one channel and the visitor never has to know which they are talking to. The AI takes the first message, resolves the routine ones, and passes the rest along. We compare the two properly in chatbot vs live chat.
Which chatbot is best for customer service?
The best customer service chatbot is the one trained on your own content that admits when it does not know — not the one with the longest feature list. Every serious chatbot platform can produce a fluent sentence now; the differences that matter show up in the awkward cases.
Concretely, judge any chatbot solution on six things:
- Grounding. Does it answer from your pages, policies and product data, or from general internet knowledge? Only the first is safe to put in front of customers.
- Honesty. What does it do when the answer is not in your content? Silence and a handover beat a confident invention.
- Handover. Can a colleague step into the same conversation, with history, without the customer starting again?
- Order lookup. Can it read order data securely, with some verification of who is asking?
- Setup cost in your time. An afternoon of pasting content, or a six-week project with an implementation partner?
- Pricing that survives growth. Check what happens at ten times your current volume before you commit.
The big customer support platforms are excellent if you already run a support desk with a team on it. For a webshop with one to five people answering messages between everything else, they are a heavy fit — you pay for a workflow engine you will not configure. Smaller AI platforms for customer service are lighter and land faster. Veyra sits in that second group deliberately; you can see the scope on our features page.
How to create a chatbot for customer service?
You create a customer service chatbot by giving it your existing content, embedding it on your site, and correcting it for a week — no development work required with modern tools. The whole first pass is an afternoon.
Step 1 — gather what you already have
Your shipping page, returns policy, FAQ, terms, and the answers you have typed a hundred times in email. That last group is the most valuable and the least written down. Ten questions with a proper answer each is enough to start; you are not building an encyclopaedia.
Step 2 — load it into the chatbot
Paste the text, or point the tool at your URLs. This is the knowledge base the AI answers from. Keep each entry short and factual — "returns within 30 days, unworn, buyer pays return postage" is more useful than three paragraphs of reassuring prose.
Step 3 — connect your order data
If you want the AI agent to answer "where is my order", it needs to look the order up. Insist on verification: an order number alone is guessable, so pair it with something only the customer knows, such as the postcode on the delivery address. Customer data deserves the same care in a chat window as anywhere else.
Step 4 — set the boundaries
Decide what the assistant must never do on its own — promise refunds, quote a delivery date it cannot verify, discuss another customer's order. Then write the escalation rule: unsure, complaint, or an explicit request for a person means a human takes over.
Step 5 — embed and watch
With Veyra it is one line of script in your site's head, and the widget is live. Getting the code onto Shopify, WooCommerce or a custom build is covered in adding an AI chatbot to your website.
Then read the transcripts for a week. Every wrong or missing answer is a gap in your content, and fixing it takes two minutes. Shops that do this reading for the first fortnight end up with an assistant that resolves most routine conversations; shops that switch it on and walk away do not.
How much does a customer service chatbot cost?
A customer service chatbot costs anywhere from nothing to several thousand a month, and for a small webshop the honest answer is that you should not be paying much at all to begin with. The market splits roughly three ways.
- Free plans — a capped number of conversations per month, enough for a shop with modest traffic to run its whole support on. See free chatbot options for a website for what the caps typically look like.
- Per-seat or per-conversation subscriptions — the common middle, usually tens of euros a month at webshop volumes.
- Per-resolution pricing — you pay for each conversation the AI closes without a human. Attractive on paper; check carefully what counts as resolved.
Two costs are easy to miss. The first is the AI usage itself, which some tools bill through and others include. The second is your own time: a platform that needs a week of configuration has a real price even when the licence is cheap.
What it changes for your support team
The reason to automate customer conversations in a small shop is not headcount. It is that the first reply becomes instant and the repetitive work stops landing on a person. Conversational AI absorbs the routine customer inquiries, and your customer support hours go to the conversations that need judgement — the complaint, the awkward exchange, the customer with one last question before a large order.
It also changes what happens out of hours. Most orders in a webshop are placed in the evening, which is exactly when nobody is at the desk. An AI agent that can respond to customer questions at eleven at night removes a whole category of abandoned baskets, without anyone working late.
And it makes support across channels consistent. The same content that answers a customer interaction in chat answers the same question by email, so nobody gets two different return windows depending on where they asked. Customer issues that used to depend on which colleague picked them up now get the same answer every time.
Measuring whether it works
Three numbers tell you almost everything. How many conversations the AI closed without a handover. How long a customer waits for a first reply, day and night. And whether customer satisfaction moved after the widget went live.
Resist the temptation to optimise the first number alone. A chatbot that never escalates is not necessarily good — it may simply be guessing. The healthier target is a high resolution rate with a low rate of customers repeating themselves or asking again the next day. Chat is also one of the cheapest places to collect customer feedback, since the customer is already typing.
Four mistakes worth avoiding
- Hiding the exit. If a customer cannot reach a person, they leave. Make the handover obvious.
- Pretending the bot is human. Name it, be clear about what it is. People are fine talking to AI when they know that is what they are doing.
- Feeding it marketing copy. The assistant will answer in the register you give it. Feed it facts.
- Leaving it unread. The transcripts are the product feedback. Nobody else is going to tell you your sizing page is unclear.
Where this fits in the bigger picture
A chatbot is the visible part of a broader shift towards customer service automation — automatic order updates, routing, macros, and AI answers all pulling in the same direction. The chat widget is simply the piece your customers actually see, which is why it is worth getting right first.
If you want to test the idea rather than read about it, Veyra's free plan takes one line of code and no credit card. Load your returns policy, ask it three awkward questions, and judge it on how it behaves when it does not know the answer. That is the test that matters.