Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintech Is Using AI for Multilingual Customer Support
Transcript
- Lucas: Welcome back to Fintech Conversations. Episode 50. I'm Lucas. Luna: And I'm Luna. Today we're looking at something a lot of fintechs are grappling with right now: scaling customer support across multiple languages without blowing up the cost base. Lucas: Right. If these conversations are useful for what you're building or running, the way these stay ad-free and listener-supported is through buy me a coffee dot com slash fexingo. Low-key, but it keeps the lights on. Luna: Yeah, and no one's twisting arms. If you get value, toss a coffee. If not, no hard feelings. Now back to it. Lucas: So, historically, if a fintech wanted to serve customers in Spanish, French, and German, they'd hire separate support teams or outsource. That's expensive and hard to keep consistent. Luna: And we're seeing that change fast with large language models. I was reading about a European neobank — they rolled out an AI support agent that handles 12 languages. They cut support costs by 40 percent and improved first-contact resolution in non-English languages. Lucas: That's the key number: 40 percent cost reduction. But it's not just about cost. The neobank in question — I won't name them because they're still in a quiet period — they told me that resolution times in Spanish dropped from an average of 8 minutes with a human to under 3 minutes with the AI agent. And customer satisfaction scores actually ticked up. Luna: That's impressive. But I imagine financial services adds a layer of complexity. You can't have the AI hallucinating interest rate calculations or giving wrong advice on regulatory matters. Lucas: Exactly. So the fintechs doing this well are not just dropping a generic chatbot in front of customers. They're fine-tuning models on their own product documentation, FAQ history, and compliance scripts. One startup called Kasisto has built a domain-specific model for banking that they claim has near-zero hallucination rates on factual queries. Luna: Kasisto — I know them. They started with conversational AI for wealth management, right? Now they're powering multilingual support for several US and European banks. Lucas: Yeah, their CEO was on a panel last month and said that the biggest challenge isn't the language model itself — it's the translation layer. Many fintechs try to use a generic machine translation like Google Translate to convert queries into English, process them, then translate back. That double translation loses nuance. Luna: Especially with financial jargon. 'APR', 'overdraft protection', 'collateralized debt obligation' — those terms don't always have direct equivalents in every language. Lucas: Right. So the better approach is to have the model natively understand multiple languages. That's what the new generation of large language models can do. They're trained on multilingual data, so they can reason in Spanish or French directly without going through English as an intermediary. Luna: And that's where the fintech Cleo comes in. They're a uk based app that helps people manage money, and they've been using a multilingual model to support users in the US and UK. Their head of product said they saw a 25 percent increase in engagement from non-native English speakers after switching to native language support. Lucas: That's a great example. Cleo started as an English-only chatbot, but their user base in the US includes a significant number of Spanish speakers. They realized that if they could serve those users in Spanish, retention would improve. And it did. Luna: But what about regulatory compliance? If the AI gives wrong advice in Spanish, the bank is still liable. Lucas: That's why most fintechs are keeping a human-in-the-loop for anything that touches compliance. The AI handles tier-one queries: 'What's my balance?', 'When is my payment due?', 'How do I reset my password?'. For anything involving loan terms or investment advice, it flags the conversation for a human. But the human can then reply in the same language, and the AI helps draft the response. Luna: So it's more of an augmentation play than full automation. Interesting. Lucas: Exactly. And the economics work because tier-one queries make up about 70 percent of support volume for most fintechs. So automating those in multiple languages frees up the human agents to focus on the complex cases. The neobank I mentioned earlier said they were able to reassign 15 support staff to higher-value roles like fraud investigation and complaint resolution. Luna: And that's a nice outcome for employee satisfaction too. No one wants to answer 'What's my balance?' 200 times a day. Lucas: Right. So the trend is clear: multilingual AI support is becoming table stakes for fintechs that want to expand globally. The technology is mature enough now that the cost of entry is dropping. A fintech can pilot a multilingual chatbot for a few thousand dollars a month using APIs from companies like Kasisto or even OpenAI's fine-tuning service. Luna: But there's a catch, right? Data privacy. If you're sending customer queries to a third-party API, you need to be careful about where that data is processed, especially under GDPR or other regulations. Lucas: That's a huge consideration. Some fintechs are opting to run open-source models on their own infrastructure to keep data onshore. Others are using providers that offer data residency guarantees. The neobank I spoke with chose to fine-tune a model on their own servers using a platform from a company called Hugging Face. That way, no customer data ever leaves their environment. Luna: So the tech stack is becoming more flexible. You can choose between a managed API or a self-hosted model depending on your risk appetite. Lucas: Exactly. And the models themselves are getting better at handling financial context. One area where we're seeing rapid improvement is in sentiment detection. The AI can tell if a customer is frustrated or confused and escalate to a human before the situation gets worse. Luna: That's important for retention. A frustrated customer who gets a quick, empathetic response might stay. One who gets a bot that doesn't understand them will churn. Lucas: And churn is expensive. The cost of acquiring a new banking customer can be hundreds of dollars. So investing in multilingual support that reduces churn by even a few percentage points has a strong ROI. Luna: Let's talk about the implementation reality. What does it actually take to build one of these systems? Is it just an API call, or is there more to it? Lucas: It's more than just an API call. You need to curate a dataset of past support conversations in each language you want to support. Then you fine-tune a base model on that data. Then you test it heavily — especially for edge cases like slang, typos, or mixed-language queries. Luna: Mixed-language queries are common in multilingual regions. Someone might start in Spanish and switch to English mid-sentence. Lucas: Right. The better models can handle that seamlessly. But it requires careful prompt engineering and sometimes a separate language detection step upfront. One fintech we spoke to said they spent three months just on the validation phase before going live. Luna: And then you have to monitor performance ongoing. Language evolves, and so do customer questions. Lucas: Absolutely. The best teams have a feedback loop: when the AI gets something wrong, a human corrects it, and that correction goes back into the training data. Over time, the model gets better. It's a continuous improvement process. Luna: So for a fintech considering this, what's the first step? Lucas: I'd say look at your support volume by language. If you have even a few thousand queries per month in a non-English language, it's probably worth piloting. Start with one language, prove the economics, then expand. Luna: And measure resolution time and customer satisfaction before and after. That's the data you'll need to justify the investment. Lucas: Exactly. For fintechs that are already global, or planning to go global, this is becoming a must-have. The genie is out of the bottle. Customers expect to be served in their own language, and AI makes that feasible at scale. Luna: It's one of those rare cases where technology both reduces costs and improves the customer experience. That's a win-win. Lucas: Yeah. And as the models get cheaper and more accurate, we'll see even smaller fintechs adopting it. The barrier to entry is already low and getting lower. Luna: Good place to leave it. Thanks for listening to Fintech Conversations. I'm Luna. Lucas: And I'm Lucas. We'll be back with another angle next time.