AI chatbot for customer service

Blai Lloveras Blai Lloveras
Updated:
chatbot de atención al cliente con IA para descargar al equipo

On a Monday morning at a small online store, the inbox usually looks the same. Weekend messages asking where an order is, others asking whether a size can be returned, and a few asking something that’s already spelled out on the shipping page. Someone on the team spends the first hour answering them one by one, with replies they’ve already written hundreds of times.

That’s the work an AI customer service chatbot can take off your plate.

AI chatbot vs. menu-based bot

You’ve probably already used one of the old-style bots. It greets you, offers four buttons, and if you type something that isn’t on its list, it replies “Sorry, I didn’t understand that” or sends you back to the menu. They run on rules: someone anticipated every question and wrote every answer, and the bot just looks for keywords in what you type to decide which branch of the tree to show you.

An AI chatbot works differently. Behind it is a language model, the same kind of technology as ChatGPT, Claude or Gemini, which understands the question even when customers phrase it their own way, with typos or two topics mixed into one sentence. The answer doesn’t come from a tree, it comes from the information you’ve given it to read: shipping terms, return policy, catalog, the FAQs on your website. If a customer writes “it arrived yesterday and it doesn’t fit, what do I do?”, the bot understands it’s a return without anyone having programmed that sentence.

That difference has a practical consequence for whoever sets it up. With a rule-based bot, the work is drawing the tree; with an AI bot, the work is preparing the texts it will draw on and setting its limits: what it can answer, what it can’t, and when it has to hand the conversation over to a person. Many current tools combine both, buttons for clear-cut cases and an AI model for everything else, and that’s the combination we recommend for an SME, because the clear-cut cases don’t depend on the model getting it right.

What questions it can handle

The fastest way to find out is to open your customer service email or your business WhatsApp and go through last month’s conversations. Two clear groups almost always show up, and both can be automated with different levels of effort.

reviewing the customer questions the business receives

Repeat questions: opening hours, shipping and returns

These are the questions whose answers already exist in writing somewhere on your website; the customer just hasn’t found them or would rather ask. Opening hours of the physical store, how long shipping to the Balearic Islands takes, whether shipping is free above a certain amount, how to process a return, which payment methods you accept, whether you issue business invoices. For this, the chatbot only needs to read those texts. There’s no need to connect it to any system, which is why it’s the cheapest place to start.

The groundwork is keeping those texts up to date and free of contradictions. If the shipping page says one thing and the terms of sale say another, the bot will repeat the contradiction with total confidence. Before setting anything up, it’s worth gathering the good answers, the ones your team actually gives, into a single document and using that document as the source.

Order status and store data

“Where’s my order?” is probably the most common question in any e-commerce business, and it can’t be answered by reading a text. The bot has to look up that specific customer’s order in Shopify, WooCommerce or PrestaShop, find out whether it has left the warehouse and return the carrier’s tracking number.

That requires an integration. Some chatbot platforms come with ready-made connectors for the most common stores and get it done in an afternoon; in other cases you need to build the connection to the store’s API, for example with an n8n workflow that receives the customer’s order number and email, queries the store and passes the result to the model so it can write the reply.

You also need to think about identification. The bot shouldn’t give order information to anyone who types in a number, so the usual approach is to ask for the order number together with the email or postal code used for the purchase. The same integration can handle similar questions: whether a size is still in stock, whether a product is coming back, what the lead time is for a made-to-order item.

What your business gains, and what it doesn’t

After-hours replies without adding staff

A customer asking on Sunday night whether a gift will arrive before Thursday doesn’t want to wait until Monday, and often doesn’t: they buy from another store. The chatbot answers at that hour with the same information your team would give, and any conversation it can’t resolve is logged so someone can pick it up first thing in the morning.

The other gain is inside the company. The hours someone spent copying and pasting the return policy can go to the questions that do need judgment: an issue with a carrier, an angry customer, a large business order. For an SME, this is often the difference between being able to grow order volume with the same team or having to hire someone just to answer messages.

Limits: made-up answers and handing off to a person

The best-known risk of language models is that when they don’t have the answer, they sometimes make one up in a perfectly convincing tone. A chatbot might promise a delivery time that doesn’t exist or accept a return outside the return window. This is greatly reduced by restricting the bot to the information you’ve given it, with clear instructions that if something isn’t in its documents it should say so and hand off, but it doesn’t disappear entirely, which is why you need to review real conversations during the first few weeks.

Complex queries aren’t its territory either. A complaint with several parts, a product that arrived broken and the customer wants a different model instead, a price negotiation for a large order. The bot can collect the details and get the case ready, but the decision has to be made by a person with the authority to make it.

And some customers simply don’t want to talk to a machine. If the bot puts obstacles in the way of reaching a person, the experience is worse than having no bot at all. The way out must always be in plain sight, with a clear “talk to someone on the team” option and an honest notice of when that person will reply if it’s after hours.

It’s also a good idea to make clear from the very first message that the replies come from an automated assistant. It’s what customers expect, it avoids misunderstandings when the bot gets something wrong, and it fits the transparency obligations that EU AI rules impose on systems that converse with people.

Text or voice: which channel to automate first

What decides the channel is how your customers write to you or call you right now. If most questions come in through WhatsApp, your website chat or email, the first step is a text chatbot: it’s cheaper, faster to test, and its mistakes are visible in writing, so they’re easier to fix. If, on the other hand, the phone never stops ringing with the same questions, which is typical for businesses with in-person service, clinics, repair shops or restaurants, the real savings are in voice.

An AI voice agent answers the call, understands what it’s told and replies out loud, using tools like Vapi or Retell AI that connect the phone line to a language model and a synthetic voice. It can give opening hours, take a booking or an appointment, or note the reason for the call and transfer it to the right person. It’s trickier to set up than a chat, because a slow reply or a voice that talks over the customer is far more noticeable than a badly worded message, so it’s worth testing it with your own calls for several days before opening it to the public.

We advise most SMEs to start with one channel, the one with the most repeat questions, measure it for a few weeks and only then decide whether to extend it to the other. The reference texts you prepare for the chatbot will later serve the voice agent too, so the work isn’t wasted.

No-code platform, helpdesk or n8n workflow

There are three routes, and the choice depends more on what you already have in place than on each tool’s features.

The first is a no-code chatbot platform such as Tidio, which has an AI customer service module designed for online stores. You install it on your website with a code snippet or a store extension, upload your FAQs or point it to the URLs to learn from, and within a few days it can be answering. It’s the option with the least technical work. The trade-off is that doing anything the platform didn’t plan for, such as querying your own system, can be complicated or impossible.

The second is turning on the AI in the helpdesk you already use. If your team works with Zendesk, Freshdesk, Intercom or Gorgias, they all have their own AI agent fed by your help center and ticket history, and the advantage is that any conversation the bot can’t resolve lands directly in the same inbox your team already works from. If you don’t use a helpdesk, we don’t recommend signing up for one just for the chatbot; it’s almost never the shortest path.

The third is your own workflow, built with n8n and the API of a model such as Claude, GPT or Gemini. That’s what we do when a business has systems no platform connects to out of the box, like an ERP, appointment scheduling software or an in-house database, or when it wants fine-grained control over what the bot does in each case. It gives you more freedom, and the cost per conversation is usually low because you pay for model usage directly, but it needs someone to build it and to maintain it when something changes in the store or the API.

Before deciding, check two things on each tool’s pricing page, because they change often: how the AI is billed, whether it’s included in the plan or charged per resolved conversation or by message volume, and which channels it works on, since not every platform covers WhatsApp on every plan.

Running the numbers to see if it pays off

The math is simple and you can do it with your own business data:

  1. Count the repeat questions you get in a month, only the ones the bot could answer without help.
  2. Multiply them by the minutes it takes your team to answer each one, including the time spent looking up the order or finding the right text.
  3. Convert that total to hours.
  4. Multiply the hours by what an hour of the person answering them costs you.

The result is what those questions cost you today.

Let’s take a fictional example to see how it reads. A store receives about four hundred questions a month about order status and shipping, each one takes its team about four minutes to read, look up and answer, and that adds up to just over twenty-six hours a month. Multiply that by the hourly cost of that person and you have the figure to compare against.

On the other side goes what the chatbot costs per month: the platform or helpdesk subscription, model usage if the tool charges for it separately, and a share of the initial setup spread over the months you expect to use it. Don’t forget maintenance time: someone has to review conversations, correct answers and update the texts when shipping or terms change.

The bot won’t resolve every question you answer today, either. Some will end up going to a person, so calculate the savings only on the ones you expect it to close on its own, and be cautious with that share until you’ve measured it with real conversations. If, with a conservative estimate, the cost of the bot comes out clearly lower, it pays off. If the numbers are tight, or you get few repeat questions a month, you’re probably better served by a good FAQ page and saved email replies than by a chatbot.

If you’d like to run these numbers with your own data and see which tool fits what you already have, get in touch and we’ll go through it with you.

About the author

Blai Lloveras

Articles by Blai

Let's talk

Do you have a project in mind?

Tell us about it and let's grow your business together.

Analyze my sales opportunities