RefugIA: designing a multilingual, anonymous AI assistant for refugees

When someone arrives in Spain fleeing their country, they often do not speak Spanish, have few resources and need to avoid drawing attention to themselves. The first thing they need is to know what to do: how to apply for protection, what rights they have, where to go for healthcare, housing or schooling for their children. The information exists, but it is scattered, written in administrative language that is hard going even for native speakers, and spread across official and charity websites.

A foundation that supports refugees and migrants brought this problem to us. That conversation led to RefugIA, a multilingual AI assistant that Bidatia has developed to bring access to that information into one place.

The problem, from the newcomer's point of view

The foundation described four recurring barriers:

Barrier	What it means in practice
Language	The information is almost always only in Spanish
Fragmentation	Each procedure is explained somewhere different, and not always consistently
Complexity	The content is technical and unclear, even for native speakers
Support capacity	Personal support has long waiting lists, there are too few volunteers and specialists, and public services are overstretched

The context makes things worse: in 2024 Spain received more than 160,000 applications for international protection, according to official figures, a level of demand that far outstrips the support capacity available.

The consequences

These barriers are not a mere inconvenience; they have real effects:

• Misinformation. Without a single, understandable source of information, contradictory versions circulate about how to start procedures or which rights apply to each person.
• No individual support. It is not possible to help each applicant according to their specific situation at one of the most critical moments of their life.
• Delays and avoidable mistakes. A missing document or a deadline missed through lack of knowledge can drag the process out for months.
• Rights left unused. Some people stop exercising basic rights simply because they do not know they have them or how to claim them.

The result is a sense of vulnerability and uncertainty precisely when the opposite is needed.

The solution: an assistant that speaks the user's language

RefugIA brings together in a single point of contact the information that is currently scattered. Each of its features responds to one of the barriers above:

• AI-based. A conversational assistant that understands questions asked in each person's own words, not just in administrative terms. Analysing queries, with no identifying data, helps to detect the most frequent questions and improve the answers.
• Multilingual. People can ask and receive answers in different languages, whatever their country of origin.
• Multi-channel. Available on the web, mobile and other channels, suited to people who often only have a phone.
• Practical. It explains the steps to follow, answers questions and provides legal and practical information on migration, asylum and official documentation.
• Anonymous. No registration is required and no personal information is stored.
• Up to date. The information is reviewed and verified continuously.
• Simple. An interface designed for anyone, including people with little digital experience.

Where the information comes from

An assistant like this is only as reliable as its sources. RefugIA combines four:

1. Vectorised regulatory documents. Relevant legislation and administrative procedures are turned into semantic representations so the assistant can find the passage that answers each question (the technique known as RAG).
2. Official sources. Information from the Spanish Official State Gazette (BOE), the Ministry of the Interior, UNHCR and other recognised bodies.
3. Structured extraction from official websites. Automated collection of information published on official sites, organised to make it easier to consult.
4. API connections. Integration with public and private databases to keep information current.

On top of all this sits a human layer: the sources feeding the system and the quality of the answers are overseen by specialists in immigration law and international protection.

Design decisions for a vulnerable group

Building an assistant for people in such a delicate situation forces decisions that would be optional in other contexts:

• Privacy by default. No registration, no identifying data and no conversations stored against an individual. For someone fleeing danger, anonymity is not a preference but a necessity.
• Plain language over exhaustiveness. An answer that is correct but incomprehensible does not help. The assistant explains in simple terms and points to the next concrete step.
• Answers grounded in sources. The assistant answers from retrieved official documents rather than from what the model "thinks it knows", and when it has no reliable information, it says so.
• Knowing when to refer. Some questions should not be resolved by a machine: risk situations, complex legal cases or emergencies. The assistant points people to the right organisation or professional.
• Continuous expert oversight. Regulations change and so do procedures; regular review of sources and answers is part of the service, not an extra.
• Genuine accessibility. Working well on a basic phone, with patchy connections and without having to install anything.

These decisions align with what the EU AI Act and the GDPR expect of systems dealing with vulnerable groups: transparency, human oversight and data minimisation.

What it aims to achieve

The benefits RefugIA pursues are concrete:

• For refugees: immediate help, in their own language and anonymously, with clear information on their rights and procedures, which strengthens their autonomy and confidence.
• For support organisations and public services: automatically handling a high volume of common questions, reducing unnecessary office visits and focusing face-to-face resources on priority or complex cases.

The goal is not to replace human support, but to make sure it reaches those who need it most.

How to approach a project like this

For any organisation or public body wanting to put an AI assistant at the service of a vulnerable group, the order matters:

1. Start from the users' problem, as described by the people who support them every day.
2. Select and maintain the sources with specialists in the field, before choosing the technology.
3. Design privacy and referral to professionals from the outset.
4. Test with real users and with support teams, and adjust the assistant's language and limits.
5. Review sources and answer quality continuously.

At Bidatia we developed RefugIA with this approach, and we apply it to other AI projects with social impact.