Multilingual apps with AI agents: serving customers in several languages without losing control (Mallorca and the Balearics)
Published on
In Mallorca, Ibiza and Menorca it is perfectly normal for one business to receive messages in Spanish, English, German, French and Italian in a single morning. So when a Balearic company starts looking at apps with AI agents, one of the first questions is: "Can the agent handle all those languages?" The short answer is that today's models write fluently in many languages. The useful answer is different: writing well is not the same as serving customers well, and multilingual support comes with its own traps that are worth solving before launch.
This post explains how to design an app with an AI agent that works in several languages without losing control of what it says, with concrete decisions and a hypothetical example.
The real problem isn't translation, it's consistency
An AI agent can reply in German even if all your documentation is in Spanish. The risk lives in the details:
- A cancellation policy that is clear in Spanish gets softened in English ("we usually refund...").
- Place names, products or services get translated when they shouldn't be (a cove, the name of a tour, a rate).
- The same question gets different answers depending on the language, because the agent fills gaps differently.
- A customer writes in French, the agent replies perfectly... and then the person who has to take over doesn't speak French.
None of this shows up in a demo. It shows up in August, across hundreds of conversations.
Step 1: decide which languages you really cover
"All of them" is not an answer. For each language, pick one of three levels:
- Full: the agent handles the conversation, manages bookings or orders, and your team can carry on in that language.
- Assisted: the agent answers common questions, but anything that needs a person is handed over with a summary in your team's language.
- Courtesy: the agent recognises the language, gives the basics and offers to continue in English or Spanish.
How to choose the level:
- Where do your customers come from? Look at your own bookings, orders or enquiries from the last year, not generic data.
- Can your team reply in that language when the agent escalates?
- Do you have reviewed legal and commercial texts in that language?
If the answer to the last two is "no", that language shouldn't be on the full level, however well the model writes it.
Step 2: one source of truth, in one language
The most common mistake is duplicated information: one sheet in Spanish, an older one in English, a German one nobody has touched for two seasons. The agent mixes them and contradicts itself.
A better setup:
- A single knowledge base in the language your team works in: opening hours, services, prices, terms, FAQs.
- The agent reads that base and replies in the customer's language.
- When a price or a schedule changes, you change it in one place.
The exception is text that must not be open to interpretation.
Step 3: fixed texts for what can't vary
Some content should never be paraphrased by the agent, in any language:
- Cancellation and refund terms.
- Prices, supplements and what each rate includes.
- Safety notices and requirements (minimum age, documents, being able to swim...).
- Privacy and consent wording.
For these, prepare fixed texts reviewed by a person in every full-level language, and have the agent insert them word for word. The agent decides when to show them, not how they are worded. For assisted or courtesy languages, the agent can show the English version and say that it is the valid one.
Step 4: a glossary of what doesn't get translated
Keep a short list of things the agent must keep identical in any language:
- Names of your services and products.
- Place names as they appear on signs (Sa Calobra, Cala Macarella, Es Vedrà).
- Meeting points and addresses.
- Names of rates or packages.
And the other way round: terms that must always be rendered the same way (for example, how you say "deposit" in German). Ten or twenty entries usually prevent most mix-ups.
Step 5: human handoff designed for several languages
When the agent escalates, the person picking it up needs:
- A summary in the team's language, not the customer's.
- The customer's language clearly visible, so the right colleague can take over.
- The customer's last question both translated and in the original.
- Anything the agent has already promised.
If nobody on the team speaks that language, the app should let staff write in their own language and have the agent translate the reply, while keeping a clear record of what was originally written.
Step 6: test per language, not just in Spanish
Before opening the agent to customers, collect a set of real questions (taken from your email or WhatsApp, with personal data removed) and run it in every full-level language. Check:
- Is the substance of the answer the same in every language?
- Does it insert the fixed texts when it should?
- Does it respect the glossary?
- Does it escalate the same cases in every language?
- What happens when the customer switches language mid-conversation?
Ideally someone fluent in each language reads a sample of replies. They don't need to read everything: checking a handful each week during the first month is usually enough to spot patterns.
Hypothetical example: bike rental in Alcúdia
Picture a bike rental company in Alcúdia whose customers are mostly German, British and Spanish, plus some French and Dutch visitors. It wants an app with an AI agent that answers questions, takes bookings and sends pickup reminders.
One way to set it up:
- Levels: German, English and Spanish at full level (the team speaks them); French and Dutch at assisted level.
- Source of truth: one base with bike models, pickup times, recommended routes and FAQs.
- Fixed texts: deposit, insurance, damage and cancellation terms, reviewed in German, English and Spanish.
- Glossary: route names, the pickup point and bike types, never translated.
- Handoff: if a Dutch customer complains about damage, the agent passes the case on with a summary for the team, shows the original complaint and drafts a translated reply that an employee reviews before sending.
The goal isn't for the agent to do everything. It's for every language to behave predictably, and for the team to know what can and can't happen.
Quick pre-launch checklist
- List of languages with their level (full, assisted, courtesy).
- One knowledge base with a named owner.
- Reviewed fixed texts for terms, prices and notices.
- Glossary of names and terms.
- Human handoff with a summary in the team's language.
- Test set run in every full-level language.
- Weekly review of a sample of conversations during the first month.
Next step
If you run a business in Mallorca, Ibiza, Menorca or Formentera and want an app with an AI agent that talks to customers in their own language without surprises, Derek Solutions designs the agent, its limits and its integration with your systems. See how we work on AI and automation and mobile app development.
Email derek@dereksolutions.com or call +34 680 283 973 and tell us which languages your customers speak and what you'd like the agent to handle.