On this page
How well today's AI models handle Catalan
The big models understand and write Catalan. Google lists it among the languages every Gemini model understands, according to documentation updated on 30 October 2026, and OpenAI includes it in the list for its speech models. Neither list gives a quality score for each language.
Anthropic publishes a table of Claude's results in fourteen languages, measured against English. Catalan is not in it. Anthropic says Claude works in many more languages than the ones in the table and recommends testing it in the ones you use.
For a shop, hotel or town council in Catalonia, Catalan is supported but the quality is yours to check. If you would rather use an open model built for Catalan, the Aina project, led by the Government of Catalonia with the Barcelona Supercomputing Center, publishes them.
How the chatbot should switch language
The simplest rule is to reply in the customer's language and switch when they do. Claude infers the language from the conversation, but Anthropic recommends stating it in the system prompt so the rule holds for the whole chat. If customers pick a language with a button, that choice goes into the prompt too.
Settle the grey areas in advance. One Spanish word in a Catalan message is not a change of language. A visitor on the Costa Brava writing in English is. And the opening message should match the language of the page the customer is on.
Article 128-1 of the Catalan Consumer Code says consumers are entitled to be served, orally and in writing, in whichever official language they choose. According to ruling 88/2017 of Spain's Constitutional Court, this paragraph does not create a true consumer right or a language availability duty for businesses, because it defers to language legislation. Even so, someone who writes to you in Catalan expects a reply in Catalan.
Which AI model is best at Catalan?
The language pages from Google, OpenAI and Anthropic do not give a separate score for Catalan. The reliable way to choose is to run the same real questions through two or three models. Ask someone on your team who writes good Catalan to compare the answers and note the mistakes.
Voice or text. What changes when the chatbot speaks
Start with text, which is easier to test and correct. Voice adds two more pieces, one that turns what the customer says into text and one that reads the answer aloud. Either can stumble in Catalan even when the model writes it well.
OpenAI supports Catalan in its speech models but warns that its voices are optimised for English. A voice like that may sound foreign or mispronounce names such as Sant Feliu de Guíxols. The Aina project also produces resources so that voice assistants can work in Catalan.
If your customers send voice notes on WhatsApp, the system has to transcribe them. Test with older callers, background noise and regional accents, and always leave a way to carry on in writing. Ona, the assistant on the crida studio website, asks you to choose a language before it starts and, if the voice fails, lets you keep typing.
Can I make the chatbot answer only in Catalan?
Yes, one line in the instructions does it. But think about customers who do not understand it. Some will write to you in Spanish and, on the Costa Brava, in English too. If the bot only answers in Catalan, those customers get no answer. A good option is to open in Catalan and follow the customer's lead.
Want us to look at your case? Leave us your details and we’ll tell you where we would start.
Let's talkHow to test it before it goes live
Make a list of the real questions from a normal month, from WhatsApp, email and the phone. Write them the way people do, with typos and mixed languages. Then read the answers aloud with a Catalan speaker on your team. If a customer from Girona would never put it that way, fix it.
Also test what the chatbot does not know, such as a bank holiday, a price that is not in the sources or a booking. Look out above all for made-up answers, confirmations of things it has not done and slips into Spanish. The table gives a few examples to start with. Adapt them to your business.
| Example question, asked in Catalan | Answer to watch out for | What it should do |
|---|---|---|
| Are you open on Sant Esteve, 26 December? | Gives the normal opening hours. | Say what the sources say or hand over to a person. |
| A Catalan message with one Spanish word in it | Switches the whole conversation to Spanish. | Carry on in Catalan. |
| How do I get there from Cassà de la Selva? | Makes up a route or uses Gerona instead of Girona. | Use official place names and link to the map. |
| Have you got a table for six on Saturday? | Confirms a booking it never made. | Send the customer to the booking system or to a person. |
| A question about a price that is not in the sources | Makes up a price. | Say it does not know and offer a contact. |
| On a town council site, how do I register as a resident? | Describes another town's procedure. | Link to the procedure page on the council's e-office. |
| Any question | Writes “tenim que” instead of “hem de” or mixes formal and informal address. | Write the way your team would. |
Do I need a model made in Catalonia, like the Aina ones?
Not necessarily, because commercial models already understand Catalan. Aina's open models make sense if you want to run the model on your own servers or need a specific Catalan component, such as voice. Either way, testing with real questions from your customers is what settles it.
What data the chatbot keeps and where it goes
Every conversation can contain personal data. Customers may type their name, phone number or a health problem without anyone asking. Before you sign up for a tool, put the questions below to the provider and explain the answers in the privacy policy on your website.
Limit what the chatbot asks for, too. A clinic does not need the bot to collect symptoms to book an appointment, and a town council does not need an ID number to give the recycling centre's opening hours. If an outside company runs the tool, ask for a data processing agreement before launch.
- Where conversations are stored, and in which country.
- How long they are kept and how they can be deleted.
- Whether the provider uses them to train its models.
- How a conversation is handed to a person without losing what has already been said.
What should I do if the chatbot gets something wrong after launch?
First check where the answer came from. If it made up opening hours, maybe they were missing from the sources or only there in Spanish. Fix the source, add the question to your test list and run it again. And agree with the provider how often you will review the conversations.
It must say it is AI, in Catalan too
Our guide on AI chatbot disclosure explains when and how to tell people they are talking to an AI. Here we only add the language side. The notice should appear in the language of the conversation and is worth repeating if the customer switches language.
Voice works the same way. OpenAI's usage policies require a clear disclosure that the voice people hear is AI-generated. On the phone or in a voice note, the first sentence can say so in Catalan and offer to pass the customer to someone on the team.
