AI for Welsh-speaking businesses: why language matters

---

Wales is a bilingual nation. According to Census 2021 data published by the Office for National Statistics, 538,300 people aged three or over in Wales can speak Welsh — 17.8% of the population. The Welsh Government's Cymraeg 2050 strategy aims to reach one million Welsh speakers by 2050, with Welsh language skills actively expanding through education and public services.

Yet most AI tools available today are built with English as the primary — often the only — language. A chatbot trained exclusively on English data will struggle with Welsh sentence structures, mutate consonants incorrectly, and produce responses that feel foreign to native speakers.

This is not a niche problem. For a Welsh-speaking business serving a Welsh-speaking community, an English-only AI system introduces friction into every interaction. Customers who prefer to communicate in Welsh are forced into English. Staff training materials miss the mark. The technology feels imposed rather than integrated.

The technical reality is that Welsh poses specific challenges for large language models. Its mutation system — where the initial consonant of a word changes depending on grammatical context — is unlike anything in English or most other European languages. A model trained primarily on English text has no innate understanding of these patterns. As NLP research confirms, Welsh is classified as a low-resource language in the AI ecosystem, meaning there is significantly less training data available compared to major world languages, making it harder for general-purpose models to handle correctly.

Regional variation adds another layer. The Welsh spoken in Ceredigion differs from the Welsh spoken in Anglesey or the Valleys. Vocabulary choices, turn of phrase, and even preferred mutations vary. A single translation model trained on 'standard' Welsh will feel flat and unnatural to speakers in any particular region. This mirrors findings in computational linguistics research on dialectal variation — models trained on a single register fail to capture the full spectrum of a living language.

The solution is not to wait for the big AI vendors to solve Welsh. It is to build locally. By fine-tuning open-source models on region-specific Welsh language data — transcripts, written correspondence, and public sector Welsh language materials — we can create AI systems that speak Welsh the way your community actually speaks it. Research projects at Welsh universities, including work on Welsh-language sentiment analysis and text processing, demonstrate that fine-tuned models significantly outperform general-purpose LLMs on Welsh language tasks.

This is the approach we take at Hynt Digital. We do not bolt Welsh onto an English-first system. We build bilingual from the ground up, because anything less would be a disservice to the communities we serve.

If your business operates through the medium of Welsh and you would like to explore AI systems that respect and reflect your language, get in touch or book a Discovery Audit.

---

Sources:

• ONS: "Welsh language skills, Census 2021" — 17.8% of Wales population (538,300 speakers)
• Welsh Government: Cymraeg 2050 — Welsh Language Strategy (million speakers target by 2050)
• Research: "Does Welsh media need a review? Detecting bias in Nation.Cymru's political reporting" — arXiv Welsh NLP research