Voice AI and Welsh Language Standards: A Compliance Guide

Most enterprises buying a voice agent scope it for English, test it in English, and sign it off in English. If that agent will answer a public helpline in Wales, that sign-off is a compliance gap before the first call lands. The Welsh Language (Wales) Measure 2011 gave Welsh official status, and the Welsh Language Standards that flow from it require a great many organisations to treat Welsh no less favourably than English. A voice agent that greets only in English, or drops to a lower quality of service the moment a caller switches to Welsh, breaches that duty as clearly as a printed form would.

This is not a small audience. The 2021 Census recorded 538,000 people in Wales aged three and over who could speak Welsh, and the Welsh Government's Cymraeg 2050 strategy commits to reaching a million Welsh speakers by 2050. At the same time, adoption of AI is nearly universal: McKinsey's 2025 State of AI survey found 88% of organisations now use AI in at least one function, yet only 6% capture material earnings impact from it. The organisations pulling ahead are the ones that treat compliance as a design input, not a retrofit.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. For a scoped review before you commit, see the AI placement diagnostic inside DATS, our five-stage AI consulting system.

What do the Welsh Language Standards require from a voice agent?

The Welsh Language Standards require that Welsh is treated no less favourably than English across the services an organisation delivers. In practice that means a voice agent must offer Welsh from the opening greeting, let the caller continue in Welsh without penalty, and deliver the same standard of service in both languages. Section 1 of the Welsh Language (Wales) Measure 2011 states plainly that "the Welsh language has official status in Wales", and the Standards give that status operational teeth.

The critical phrase is "no less favourably". It is not enough to bolt a Welsh option onto an English system if that option is slower, shorter, routed to a voicemail, or only available at limited hours. The Standards were written to end exactly that pattern, where Welsh service technically existed but was quietly second class. For a voice agent the test is direct: could a Welsh-speaking caller complete the same task, at the same speed, with the same quality, as an English speaker? If the honest answer is no, the deployment is not compliant, however good the English experience is.

Who is bound by the Welsh Language Standards?

The Standards bind a wide set of organisations, not only the Welsh Government. The Welsh Language Commissioner imposes Standards on individual bodies through a Compliance Notice, and the list runs across local councils, NHS Wales health boards, national park authorities, fire and rescue services, Welsh universities, and many other public bodies. Certain bodies that receive substantial public funding, and some privatised utilities and transport operators, also fall in scope, pulling a surprising number of enterprises into the regime.

If you sell a voice platform into any of those buyers, their Welsh language duty becomes your procurement requirement. A council that adopts an English-only agent inherits a compliance failure, so its procurement team will screen for Welsh capability before signing. That is why bilingual support is increasingly a gating question in Welsh public sector tenders rather than a nice-to-have. Universities in particular sit across both consumer expectation and statutory duty, a point we develop in our guide to voice AI for higher education admissions, where Welsh-medium applicants expect a Welsh-medium service.

Even where an organisation is not directly bound, serving Welsh-speaking customers well is fast becoming a reputational baseline, and the same governance discipline that underpins our AI operating model consulting applies: decide the language duty before you build, not after a complaint.

Does using AI remove your Welsh language duties?

No. Adopting AI does not exempt an organisation from its Welsh language duties. The Welsh Language Commissioner made this explicit in an AI policy statement published in August 2025, which set out how bodies subject to the Standards must handle AI as it spreads into translation and public-facing services. The Commissioner was direct that the duty travels with the service, whoever or whatever delivers it.

"These standards ensure that the Welsh language is treated no less favourably than English and continues to be a living language in our everyday lives."

Welsh Language Commissioner, AI policy statement, August 2025

Importantly, the Commissioner did not frame AI as a threat to be blocked. The same statement noted that "the need to comply with language duties does not mean that technology or AI cannot be used to deliver services", and pointed to real opportunities in real-time translation, transcription, and Welsh speech tools. The message for enterprises is balanced but firm: AI is welcome, and the accountability for treating Welsh fairly stays with the organisation that deploys it. If you want to understand how that accountability maps onto data handling as well as language, our explainer on the privacy notice and transparency duties for voice AI covers the parallel obligations under UK GDPR.

What does a compliant bilingual voice agent actually do?

A compliant bilingual voice agent offers Welsh from the first second, lets the caller choose Welsh without asking twice, and then delivers the entire interaction in Welsh at the same quality as English. It does not route Welsh callers to a callback queue, shorten the Welsh script, or hand off to an English-only fallback when the conversation gets complex. The behaviour has to be equal end to end, because the Standards judge the whole journey, not the greeting.

This is a design specification, not a translation task. It shapes how you configure routing, how you script fallbacks, and how you measure success, because a containment or resolution metric that only counts English calls will hide a Welsh service failure. Getting this right before launch is far cheaper than unpicking it later, which is the whole logic behind our fixed-fee AI placement diagnostic: find the compliance and design gaps before they reach production. You can also see the underlying model on our approach page.

Can voice AI handle Welsh to the same quality as English?

Increasingly yes, but not automatically, which is where most deployments fail. Welsh is a lower-resource language than English, so speech recognition, understanding, and text-to-speech that sound flawless in English can be measurably worse in Welsh unless the models were built and tuned for it. The Welsh Language Commissioner's own AI statement noted that many modern systems already process Welsh to a high standard, but that capability is uneven across vendors, especially for the harder task of live telephone speech.

The honest engineering position is that Welsh voice quality has to be tested, not assumed. That means measuring Welsh speech-recognition accuracy on real Welsh callers, checking that the synthesised Welsh voice is natural rather than an English voice reading Welsh phonetically, and handling code-switching, where a caller moves between Welsh and English mid-sentence. Dilr Voice treats Welsh as a first-class language in this respect, and we are candid that closing the last gap in Welsh accuracy is genuine engineering work rather than a configuration toggle. Buyers should ask any vendor, including us, for Welsh-specific accuracy evidence, in the same spirit as our guidance on voice AI accuracy evaluation.

What happens if you get Welsh language compliance wrong?

Getting it wrong exposes the organisation to investigation, enforcement, and a civil penalty. Where a body fails to comply with the Standards, the Welsh Language Commissioner can investigate, and the 2011 Measure provides that "a civil penalty must not exceed £5,000". The cap is modest, but the penalty is rarely the real cost. The reputational damage of being found to have treated Welsh speakers as second class is far larger for a public body serving its community.

There is a route of challenge too: organisations can appeal Standards and certain decisions to the Welsh Language Tribunal, so the process is not arbitrary. But the practical lesson for an enterprise is to design for compliance rather than litigate after a complaint. Complaints are also cheap for the public to make and visible when upheld, and their volume tends to rise as public awareness grows, which is why we treat language duty as a launch-blocking requirement in the same tier as data protection. If Welsh call recordings and transcripts are involved, the data minimisation and redaction obligations under UK GDPR apply on top of the language duty, the lawful basis question is covered in our note on the GDPR legitimate interest balancing test, and whether the vendor is your processor or a controller is set out in our Article 28 controller and processor guide.

The scale of the challenge is best seen in the gap Cymraeg 2050 is trying to close.

The trajectory matters commercially: a shrinking base makes every point of contact more important, and public bodies are under pressure to make services genuinely usable in Welsh, not merely available. That pressure flows straight into procurement. See how we sequence this kind of work in our AI execution office engagement, and the wider DATS five-stage AI methodology.

What is the best voice AI for Welsh language compliance in 2026?

The best voice AI for Welsh language compliance in 2026 proves equal Welsh and English quality on your real call types, not the longest feature list. Developer-first platforms such as Vapi, Retell AI, and Bland AI give you full control but put the entire Welsh-quality burden on your team. Managed platforms such as Dilr Voice and PolyAI, and synthesis specialists such as ElevenLabs, aim to carry more of it, though Welsh depth varies and must be tested.

Here is the honest concession: if your Welsh call volume is low, your use case is a single simple task, and you have in-house Welsh speakers who can build and quality-assure the agent, a do-it-yourself stack on one of the developer platforms can be perfectly compliant and cheaper. The managed route wins when Welsh is a genuine, recurring part of the service, when the duty is statutory rather than optional, and when you need documented evidence of equal treatment for a procurement panel or the Welsh Language Commissioner. Whichever way you go, insist on Welsh-specific accuracy numbers and an equal-treatment test plan before signing, exactly as you would for our enterprise voice AI deployments.

Do private companies have to offer a Welsh voice service?

Not all of them, but more than assume they are exempt. The Welsh Language Standards bind public bodies and certain organisations that receive substantial public funding, and some privatised utilities and transport operators are also in scope. A purely private company with no public funding is usually outside the statutory duty, but if it serves Welsh customers or bids for public contracts, offering a Welsh voice service is fast becoming a competitive and reputational expectation rather than a legal one.

Is an English-only agent with a Welsh callback option compliant?

Almost never. Routing Welsh speakers to a callback queue while English speakers get instant service is close to the definition of treating Welsh less favourably. The Standards test the whole journey for equal treatment, so a slower, deferred, or narrower Welsh path fails even if a Welsh option technically exists. A compliant design lets the Welsh-speaking caller complete the same task, in the same call, at the same quality as an English speaker.

Does the caller have to ask for Welsh, or must you offer it?

You must offer it. A compliant bilingual agent presents Welsh proactively from the opening greeting rather than hiding it behind a request or a menu that only English speakers can navigate. Making the caller ask, especially in English, is itself a form of less favourable treatment. Offer both languages up front, honour the choice immediately, and keep the caller in Welsh for the rest of the interaction without repeated prompting.

Want to see this in production? Try Dilr Voice live, explore our AI execution office, read the voice AI compliance guides, or learn about Dilr.ai and how we place AI inside regulated systems.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.