AI chatbot for business: on your website and WhatsApp, with your real information
A business chatbot is only as good as its answers. Ours replies with what you have already published on your website, your shop and your policies, and when it is not sure of an answer, it does not improvise: it hands the conversation to a person. It works on WordPress, on WooCommerce and, if you want, on WhatsApp.
What a chatbot does with your business information
Think of the person on your team who knows the business best. They know how long a delivery to the Canary Islands takes, which size someone who is 1.80 m tall should order and why a personalised product cannot be returned. Now imagine that person being available at three in the morning. That is what we try to build.
The chatbot reads what already exists: pages, product pages, shipping and returns terms, privacy policy. It answers with that and only that. If the question goes beyond it, it says so and passes the conversation to you, with what the customer has already said, so they do not have to repeat it.
That is the difference from a generic widget: here you decide what it answers, what it never answers and when it stays quiet.
What a well-tuned chatbot sounds like
Three made-up conversations so you can see the tone and the limits. They are not real customers or real conversations.
- CustomerHi, I am 1.78 m tall and usually wear a medium. Does this jacket come up large?
- Assistant (AI)Hi. I am the shop's AI assistant. The product page says this jacket is a regular fit, so your medium should fit you well. If you like it looser, the large is the next size up.
- CustomerOK. And if it does not fit, can I return it?
- Assistant (AI)Yes, you have 30 days from when you receive it, with the label still on. I can explain how to do it if you like.
- CustomerAnd what if I want to complain because the seam arrived split?
- Assistant (AI)That is something a person on the team looks at. I will pass you over to them and leave them a summary so you do not have to repeat yourself.
Try it: we built the assistant on this website
The assistant that opens on this page is a real chatbot, built by us. It asks about your project, fills in a brief while you talk, accepts screenshots, PDFs and voice notes, and passes you to a person if you ask. You do not have to take our word for it: open it and ask it something odd.
Chatbot for WooCommerce shops
A shop always gets the same five or six questions: when will it arrive, which size should I order, is it in stock, how do I return it, can I change the address. A WooCommerce chatbot with your catalogue and your terms answers them at any hour and leaves your inbox for what really needs a person.
Connected to the shop, it can check a product's real availability instead of repeating what the product page said last month.
What it must never answer
Orders with a problem. A lost or damaged parcel is handled by a person.
Personal data. Without verifying who is writing, it does not say what anyone bought or where it was sent.
Complaints and late returns. These need a decision, and making decisions is not a bot's job.
Customer service chatbot: when it hands over to a person
The hard part of a customer service chatbot is not getting it to answer, but getting it to know when to stop. We agree the handover rules with you before building anything: which topics always go to a person, during which hours someone is on the other end and how we let you know (email, WhatsApp or a dashboard).
If it is the middle of the night and nobody can help, the chatbot says so honestly, takes down the reason and contact details, and the conversation is waiting for you in the morning. Nobody is left talking to a brick wall.
Which AI to use for your chatbot
There is no single best model for everyone. We use whichever your case needs, and this table shows the numbers behind that decision.
| Tier | Model | Price per million tokens (input / output, USD) | Quality index* | Cost per 1,000 conversations (USD) | Where data is processed | Open weights |
|---|---|---|---|---|---|---|
| Very cheap | GPT-6 Luna (no reasoning)OpenAI | 0.1 / 0.5 | 18 | ≈ 1.3 | US; EU option (+10%) | No |
| Very cheap | Claude Haiku 5.5 (no reasoning)Anthropic | 0.1 / 0.5 | n/a | ≈ 1.3 | US; no EU option | No |
| Very cheap | MiMo V2.6 Flash (no reasoning)Xiaomi | 0.14 / 0.28 | 38 | ≈ 1.2 | EU (Netherlands) | Yes |
| Very cheap | GLM-4.7 Flash (no reasoning)Z.ai | 0.061 / 0.4 | n/a | ≈ 1.6 | Singapore | Yes |
| Balanced | DeepSeek V4.1 Flash (no reasoning)DeepSeek | 0.15 / 0.6 | n/a | ≈ 1.7 | China | Yes |
| Balanced | DeepSeek V4.1 Flash on ScalewayScaleway (Paris) | 0.672 / 2.015 | n/a | ≈ 8.1 | EU (France) | Yes |
| Balanced | Mistral Small 4Mistral AI | 0.15 / 0.6 | 9 | ≈ 1.8 | EU (France) | Yes |
| Balanced | Gemini 3.5 Flash-Lite (minimal reasoning)Google | 0.3 / 2.5 | 22 | ≈ 5.2 | US; EU n/a | No |
| Higher quality | GLM-5.3 Flash (low reasoning)†Z.ai | 0.15 / 0.5 | 42 | ≈ 1.9 | Singapore | Yes |
| Higher quality | MiniMax M3.1 Flash (low reasoning)†MiniMax | 0.15 / 0.6 | 43 | ≈ 1.8 | US | n/a |
Fast, non-reasoning models only: models that "think" before answering (Opus, GPT-6.1 Sol, Gemini 4…) are slower and more expensive, and a chatbot does not need them. † GLM-5.3 Flash and MiniMax M3.1 Flash do not let you switch reasoning off; we use them on the minimum setting and the cost goes up a little depending on the question. Kimi is not listed because it currently has no non-reasoning model open to new accounts.
* Artificial Analysis Intelligence Index, in the mode it publishes for each model (not always the same one); "n/a" means no public figure in that mode. It is a general index: it does not measure quality in Spanish or in customer service.
Assumptions: 6 messages per conversation, around 3,000 input tokens per message (2,000 of them already cached) and 200 output tokens, with no extended reasoning. Per 1,000 conversations, the cost is six times the input price, plus twelve times the cached-input price, plus 1.2 times the output price. With reasoning switched on or more information per query, multiply it by two or three.
Prices checked on 24 October 2026. They update automatically every Monday with OpenRouter's prices; OpenRouter gives access to all these models with a single account. The Scaleway price is its own rate. "Where data is processed" is what each provider's official API states (on OpenRouter it depends on the provider serving each request); "n/a" means we have not been able to confirm it.
Chinese models and GDPR. DeepSeek's official API processes data in China, and Z.ai does so in Singapore: neither country has an EU adequacy decision, so sending personal data there requires standard contractual clauses and a transfer impact assessment. As many of these models have open weights, the alternative is to run them with a European provider (for example, DeepSeek on Scaleway, in Paris), with the data kept in the EU.
Shop FAQs
Shipping, sizes, returns: a very cheap model. The questions are simple and the content is well defined.
Pre-sales with a large catalogue
It has to compare products and understand what the customer wants: a balanced one.
Personal data or sensitive sectors
Health, finance or customer data: a model with the data in the EU (Mistral or an open model with a European provider) and more rules for handing over to a person.
And for Spanish, which one?
Even the most expensive model in the table costs less than a cent per conversation: the model is rarely the cost to worry about; the quality of the answers is. There is no reliable public benchmark comparing these models' quality in Spanish for customer service. That is why we do not choose from the table: we test the candidates with your business's real questions and keep the one that answers best.
WhatsApp chatbot, through the official channel
Yes, it can be done, and only one way: with the WhatsApp Business Platform, Meta's official API. A chatbot that helps with your own business (orders, bookings, questions) is allowed. What Meta restricts is general-purpose AI assistants.
There are tools that automate WhatsApp Web or use modified apps. They are cheaper and quicker to set up, and they work until WhatsApp blocks the number. Losing the number your customers know you by is a cost that is not worth it, so we do not use them.
Prior consentMessages only go to people who have given their number and agreed to hear from your business.
24-hour windowWhen someone writes to you, you can reply freely for 24 hours. Outside that window, only templates approved by Meta are allowed.
A way out to a personThere must always be a clear way to talk to someone.
Chatbot plugin or custom chatbot
A plugin can be installed in an afternoon, and it is enough if your questions are few, similar and do not depend on anything else. If that is your case, we will tell you and will not sell you anything else.
| Criterion | Chatbot plugin | Custom chatbot |
|---|---|---|
| Set-up | Quick | Slower: your information has to be loaded and tested |
| Information it uses | Whatever you paste in | Your website, your catalogue, your policies and, if needed, your live shop |
| Control over its answers | Whatever the plugin allows | Whatever we agree together, including off-limits topics |
| Your data | Usually stays on the vendor's service | Conversations are stored on your hosting or your cloud |
| Maintenance | Done by the vendor | Agreed with you |
The real difference is control. With a plugin, you accept the way it thinks; with a custom build, you decide what it says, what it does not and which data it touches.
Is a chatbot worth it for you?
Choose what you get asked most often and how long each answer takes you. With 5 enquiries a month it is not worth it: a chatbot starts to make sense when the same questions cost you hours, not minutes.
Each year, that task costs you
77 hours
Based on 46 working weeks a year and your own figures.
Data, GDPR and where conversations are stored
This is not legal advice: it is what we set up so that you can comply without nasty surprises. Please confirm the specific legal side of your case with your adviser.
Conversations are stored on your side: on your hosting or in your cloud, not in a third party's dashboard.
You set how long they are kept: after that period, they are deleted or anonymised.
Anyone who writes can ask you to delete their data: and we have a way to find and remove their conversations.
People are told from the very first message that they are talking to an AI: the EU AI Act (Article 50) requires it from 2 August 2026, and mentioning it in the legal notice is not enough.
Data processing agreement: with the model provider, and with the WhatsApp provider if there is one.
EU data residency: wherever the provider offers it. If you handle sensitive data, that is the first thing we choose.
The permissions and logs we apply to every AI connection are described in MCP server for WordPress.
How we build a chatbot
01. What keeps coming up
We go through your messages, emails and calls to see which questions keep coming up and which ones should not be touched.
02. Content and rules
We organise the information it needs to know, set which topics go to a person and choose the model using your real questions.
03. Testing
We test it on a staging copy, with tricky questions and traps, not on your live website.
04. Documented handover
What it does, what it does not do, where the data is stored and how to switch it off. If you like, we keep it up to date.
Who does the work
We are a studio from Oleiros (A Coruña, Spain) with 8 years of hands-on WordPress experience, working freelance since February 2018. You talk to the person who does the work, not to a go-between. Meetings are by video call and, if you are in the city of Barcelona, we can also meet in person. See also other ways to use AI in WordPress: an assistant such as Claude working inside your website through an MCP server, or automating what the chatbot collects (alerts, records, CRM).
Frequently asked questions about chatbots for WordPress
How much does an AI chatbot for a website cost?
It depends on how much information it has to handle, which systems it connects to and how many cases need testing. We do not publish a price list: we give you a fixed written quote before we start. Model usage is paid separately: between 1 and 8 US dollars per 1,000 conversations, depending on the model (prices checked on 24 October 2026). If you add WhatsApp, Meta's message charges apply too.
Can it answer using the information on my website and my products?
Yes, and only with that. It draws on your pages, your product pages and your policies. If it is not there, it does not make it up.
What happens if it does not know the answer?
It says so and hands the conversation to a person, along with what the customer has already said. If nobody is available, it takes down the reason and contact details so that someone can reply during office hours.
Chatbot plugin or custom chatbot?
A plugin is enough if your questions are few and simple. A custom chatbot pays off when it needs your catalogue, shop data, handover rules and control over where conversations are stored.
Does it work on WhatsApp?
Yes, only with the official WhatsApp Business Platform API and in line with Meta's policy. We do not use unofficial tools because they can get your number blocked. Messages carry their own Meta charges.
Is it GDPR compliant?
We set up the chatbot so that you can comply: conversations stored on your side, a retention period you choose, deletion on request, notice from the first message that people are talking to an AI and a data processing agreement with the model provider. Please confirm the specific legal side of your case with your adviser.
Which AI model do you use?
Whichever suits your case, from several providers. We decide by testing with your real questions, not from a price list.
Tell us what you get asked most
Tell us which questions keep coming up on your website or in your shop, and through which channels. We will reply within 48 working hours, after reviewing your case, with a fixed written quote.
Your first consultation is free. We look at what you tell us and reply within 48 business hours.