Latest / The CMO Podcast with Fexingo: Marketing Leadership, Budgets, and Executive Strategy / How CMOs Are Using Generative AI for Content Localization
Transcript
- Lucas: It's July 2026, and probably half the CMOs listening right now are trying to figure out how to take one campaign and make it work in Tokyo, São Paulo, and Berlin without spending a fortune on agency localization fees. Luna: Right — and the old way was basically throw bodies at it. You'd have a separate agency partner in each market, or you'd run everything through a central translation team that bottlenecked for weeks. Lucas: Exactly. And what we're seeing in the past eighteen months is that generative AI has changed that equation radically. I've been tracking a specific case — a global cosmetics brand, household name, that used generative AI to localize a single product launch into twenty-seven languages simultaneously. Luna: Twenty-seven languages — that's not just French and Spanish. That's Vietnamese, Arabic, Swahili, right? Lucas: Yes. And they did it in under three weeks, start to finish. The old process would have taken four to five months and cost something like eight hundred thousand dollars. They spent under three hundred thousand. And the brand team told me the quality was actually better in some markets because the AI didn't just translate words — it adapted cultural references. Luna: That's the part that's hard for a lot of marketers to trust — can an LLM really know that a certain color means mourning in one culture and celebration in another? Lucas: It can, if you train it right. The team built what they called a 'brand voice pack' — a custom GPT fine-tuned on their global brand guidelines, past campaign assets, and a cultural sensitivity handbook. So before the model ever saw a product name, it had ingested decades of local nuance. Luna: If today was actually useful to you, the way these stay ad-free is listener support. You can buy me a coffee dot com slash fexingo. It's how we keep the podcast independent and focused on real strategy, not sponsor messages. Lucas: Yeah, and we mean that — every episode is built around one concrete case you can use. If you've gotten value from that, the link is buy me a coffee dot com slash fexingo. Now, back to the localization stack. The interesting part is what the model actually did differently from a standard translation tool. Luna: So give us the technical breakdown. What was in that custom model? Lucas: They started with a base LLM — I believe it was GPT-4 at the time — but they layered in a retrieval-augmented generation, or RAG, pipeline that pulled from a vector database of approved copy from every market. So when the model needed to write a tagline for the Japanese launch, it could look at all previous Japanese taglines and match the tone and length conventions. Luna: That's smarter than just asking ChatGPT to translate 'radiant glow' into Japanese. Because 'radiant glow' might not be a phrase that resonates in that market at all. Lucas: Exactly. In fact, for the Korean market, the model suggested a completely different emotional angle — it moved from 'brightness' to 'evenness,' because the brand's research showed Korean consumers cared more about skin tone uniformity than luminosity. The AI surfaced that from the cultural handbook they'd fed it. Luna: So the AI wasn't just translating — it was adapting the creative strategy. That's a huge leap from where we were even two years ago. Luna: That's the kind of thing that keeps CMOs up at night. One tone-deaf idiom and you've got a PR crisis. Lucas: Right. So they built a two-step human review process. First, a native-speaking linguist reviewed every output for cultural appropriateness. Second, a local brand manager signed off on strategic fit. That added maybe a week to the timeline but caught dozens of issues. The model's accuracy rate on first pass was about eighty-five percent, but that fifteen percent gap is where the reputational risk lives. Luna: Eighty-five percent is actually quite good for a first pass. What were the common errors — literal translations, clichés? Lucas: Mostly clichés and overused idioms. The model would default to safe, bland phrases because it was trained to avoid risk. So you'd get copy that was technically correct but had no personality. The human reviewers had to inject brand voice back in. The lesson is that AI is great for the skeleton, but you still need human storytellers for the soul. Luna: I've heard similar from CMOs in the B2B space. One told me they use AI for first-draft localization of their white papers, but they always have a regional sales leader rewrite the opening paragraph. That's where the market-specific hook lives. Lucas: Exactly. And the cosmetics brand eventually developed a scoring system for cultural resonance. They'd run every localized asset through a panel of local consumers — not just a focus group, but a quantitative survey that measured emotional response. They found that ai adapted copy scored within ten percent of fully human-crafted copy on emotional engagement, but at a fraction of the cost. Luna: That ten percent gap — is that shrinking as the models improve? Because we're only in July 2026, and the pace of improvement is staggering. Lucas: It is shrinking. The same brand told me that six months later, when they did a second launch, the gap was down to six percent. The model had ingested the feedback from the first campaign and improved its cultural calibration. So there's a learning curve, but it's steep. Luna: What tools are actually leading in this space? I know DeepL has always been strong for European languages, but for something like Swahili or Vietnamese, are there specialized platforms? Lucas: DeepL is still excellent for European and some Asian languages. But for the long tail of languages, a lot of teams are using custom GPTs or open-source models fine-tuned on domain-specific data. There's also a startup called Lokai that focuses specifically on marketing localization — they have pre-trained cultural modules for different regions. I've seen demos and they're impressive. Luna: What's the one piece of advice you'd give a CMO who's considering this today? Lucas: Start with a single market that you already know well. Run a pilot where you localize one campaign with AI and one with your traditional agency. Measure everything — cost, time, engagement, brand sentiment. Use the data to build trust with your team and your board. Then scale. Don't try to go from zero to twenty-seven languages overnight. Luna: That's smart. And it also gives you the ammunition to prove that the human reviewers are still essential — you can show exactly where they added value. Lucas: Exactly. The CMO of that cosmetics brand told me that after the pilot, she actually increased her localization headcount — but she shifted their role from translators to cultural strategists. They weren't writing copy from scratch anymore; they were guiding the AI and polishing its output. The team felt more valued, not less. Luna: That's the optimistic version of the AI story — not replacement, but elevation. And it's a good note to end on. Lucas: It is. Because the goal isn't to cut costs until quality suffers. It's to free up your best people to do the work that actually requires human judgment — and let the machine handle the heavy lifting at scale. Luna: And for listeners who want to dig deeper, the brand's case study is public. We'll link it in the show notes. For now, thanks for being here.