Multilingual AI chatbot development is a feasible alternative here. It is a software that talks naturally to people in many languages. It is more accurate and consistent than other options.
The opportunity is not just global. The 2021 Census for England and Wales estimated that about 1 in 11 residents had a main language other than English. These languages include Polish, Romanian, Punjabi, and Urdu. It is a question of customer experience for businesses looking to serve multiple local markets and to reach beyond their domestic market.
Large language models have made it possible to create software that understands multiple languages. The challenges of code-switching, cultural tone, and managing multilingual content are significant, but they are downstream from the fundamental choice of architecture.
If you are a CXO, CTO, product owner, founder, or work in expansion teams in US or UK-headquartered companies, this article is for you. It explains the technical options to select an appropriate launch profile in adequate detail.
Here are some of the most common reasons why multilingual customer support is required:
● Customer preference: The majority of customers prefer to shop and be assisted in their own language. They do not like to engage with a business in a language they are unfamiliar with.
● Market expansion: Launching a product or service in a new country often requires delivering local customer support. A multilingual chatbot can act as a tested assistant before local staff is hired.
● Cost-efficiency: Managing multiple support teams that work around the clock is expensive. Automation helps reduce this cost and frees up human teams to handle more complex issues.
● Consistency: Digital assistants maintain consistent answers using a centralised knowledge base across multiple markets.
● Inclusivity: Many public services, from health and social care to utilities and housing associations, deliver services to populations who may not be fully comfortable with English.
There are three main approaches to building a multilingual chatbot.
The first approach uses a bot that communicates in one language, with a translation layer at the front and back ends. A customer can ask the bot in Spanish, the message will be translated into English. The bot will reply in English and it will be translated back into Spanish.
The second approach uses Large Language Models (LLMs) that can understand and reply to questions in many languages. There is no need to use multiple bots.
The third option is to use a hybrid of the two approaches. A native multilingual model that can handle most of the conversation. It has access to approved localised content for certain responses.
| Approach | Strengths | Weaknesses | Best for |
|---|---|---|---|
| Translation layer | Fast, reuses existing bot | Loses nuance, double translation errors | Low-risk FAQs, early market tests |
| Native multilingual LLM | Natural conversation, handles mixed languages | Needs careful grounding and testing per language | Customer service, sales conversations |
| Hybrid with approved local content | Natural tone plus controlled accuracy | More content management effort | Regulated sectors, pricing and policy questions |
A chatbot will not be relevant if it delivers the right words but in the wrong currency or date format. Localisation includes address formats, public holidays, units of measure, and regulatory requirements, as well as cultural norms around formality.
The Challenge: Vodafone sells communication products and services to consumers, enterprises, and governments in Europe, Africa, and beyond. There are significant differences in languages, products, regulations, and customer expectations across these markets. At the same time, customers expect a consistently high level of service, including the ability to get fast and personalised help in their own language.
The Solution: In 2017, Vodafone announced the launch of its virtual assistant TOBi, which was rolled out initially across 13 countries and 15 languages. In 2024, as part of a ten-year partnership with Microsoft, Vodafone began to launch SuperTOBi. It is an enhanced version of TOBi that combines generative AI from Microsoft Azure OpenAI to allow it to understand and respond to more complex questions.
TOBi is now being launched in Italy and Portugal, with plans to roll it out across Germany, Turkey, and other countries, translating questions into different languages and routing them to the appropriate responders.
The Business Impact: In Portugal, where TOBi was used to help schedule appointments, first-time resolution improved from 15 to 60%, with an increase in online NPS from 50 to 64.
The main lesson is that Vodafone was able to target several languages and local markets with a shared global platform. SuperTOBi was rolled out on a country-by-country basis after initial successes.
Testing is often considered a pivotal point in technical projects involving multiple languages.
There is value in adopting a mixed approach:
● Hiring native reviewers: Natives have language fluency and can give valuable feedback on a sample of questions.
● Golden test sets: Build a set of real customer questions with approved answers, for each language, and regularly test the chatbot’s performance against these.
● Automated scoring: Use automated tools to score responses against approved content for each language at a sentence level, with manual checks for low-scoring answers.
● Per language metrics: Resolution and satisfaction rates may well differ from one language to the next. An overall average may disguise weaknesses in one or more languages.
● Customer feedback: Make it easy for customers to report poor responses and then make sure those questions are reviewed by human moderators.
Avoid the trap of thinking that you need to build expertise in every language a model can use. Begin with the languages that carry the most customers and revenue. You can choose the languages to prioritise from your website. It might include the languages in which customers are viewing the site (as indicated by browser settings), placing orders, and contacting the company.
Local customers may have different needs than overseas customers. A nuanced understanding of the local community is particularly important for public services, such as social care or housing associations and utilities. It may be more valuable to have a high-quality service in three or four languages than to provide lower-quality responses in twenty.
The practical consideration of Large Language Models (LLMs) not being as accurate in low-resource languages is always critical. It is therefore worth doing a cost-benefit analysis for each language.
Here are some of the most commonly faced pitfalls when building or implementing chatbots for multiple languages.
● Translating instead of localising: Using the correct words, but failing to take account of cultural differences can cost you in the end. This also includes using the wrong currency or date format.
● Letting the model translate sensitive or legally binding language: Pricing and legal language should come from approved content. It shouldn’t be made up.
● Testing all languages at launch: Begin with the biggest markets and then move on to secondary ones. It is a gradual process.
● Not prioritising low-resource languages: Such languages are important too. If the model does not understand or represent these languages well, it can lead to customer frustration.
● Not providing a human alternative: For complex questions, customers always want the option to speak with a customer service agent. Include appropriate interpretation if necessary.
● Data residency: Data protection laws in different countries and regions may mean that data cannot be stored in certain locations.
The step-by-step process for developing a multilingual chatbot can be as follows:
● Prioritize languages: Select the languages to begin with based on customer data, commercial considerations, and contact center insights.
● Audit content: Identify shared content, variations by market, and gaps in the knowledge base for each language.
● Select architecture/models: Perform proof-of-concept testing for each language option with potential partners.
● Develop localized knowledge: Translate and approve content, and make sure it is easily accessible to the model.
● Systems integration: Integrate the bot with relevant systems, using local variants where necessary.
● Test the bot: Conduct testing with native reviewers, golden test sets, and automated evaluations.
● Launch: Begin with one or two languages and/or markets and use insights to inform the next phase of development.
IIH Global’s AI chatbot development services and natural language processing services can help your business develop an effective multilingual chatbot for your customers.
Multilingual AI assistants can be used in a variety of areas beyond the website to help deliver a better customer experience:
WhatsApp and other messaging channels, which are growing in popularity as a channel for customer communication.
● Voice assistants: With the addition of speech recognition and synthesis, multilingual assistants can also be used as voice assistants.
● Internal helpdesks: Many of the concepts that apply to public or consumer-facing chatbots can also be applied to internal helpdesks for multinational organizations.
● Agent-assist technologies: Multilingual chatbots can also be used to help agents communicate with customers.
● Email automation: Emails can be written in the customer’s language and automatically classified and routed.
Models continue to improve, particularly in lower-level languages, and there is likely to be increased use of speech assistants that can understand and translate multiple languages.
A greater attention to cultural differences and nuances in local languages can be expected. It will lead to more regulation of multilingual AI systems, particularly in different markets and jurisdictions. Companies that recognise the importance of language in customer experience will be able to navigate this environment successfully.
If you are looking to get insights into languages spoken in the UK, the Office for National Statistics has information from the latest census that may be helpful to you.
Multilingual AI chatbot development allows organisations to offer customer assistance in their chosen language at any time and from a single platform. This requires choosing the best architecture and approach, taking local variations into account, managing critical content, and conducting extensive localisation testing in each language. Vodafone’s TOBi provides valuable insight into the potential for such an assistant, provided that the technical and commercial fundamentals are in place.
As a full-stack AI development company, IIH Global can help businesses build multilingual AI chatbots tailored to their customer support needs.
To learn more about how you can launch a multilingual assistance service for your customers, contact IIH Global.
A multilingual AI chatbot is a conversational assistant that understands and can reply in multiple languages. It can detect the language that the customer is using and reply in that language, using natural-sounding utterances.
Modern large language models understand dozens of languages. The choice of how many languages to offer depends on a number of factors, including the quality desired for each language and the availability of content and testing resources for each language.
For factual queries, machine translation can yield satisfactory results. For other queries, such as those relating to prices and legal terms, it is best to use locally approved content and human moderation.
A combination of native reviewers, golden test sets including questions with known answers, and automated tools can be used to score responses in different languages. It is also important to analyse resolution and satisfaction rates for each language before launch.
Translation involves converting text from one language to another. While localisation goes beyond translation and considers other attributes as well. It includes cultural norms, such as dates, currencies, and public holidays. Differences in tone and other aspects of the user experience are also accounted.