They have closed a door on Luzia on WhatsApp, but her bet was already going the other way
The news was an expected splash of cold water: Meta will close general chatbot access to its WhatsApp business API starting January 15, 2026. ChatGPT, Perplexity, Poke or Luzia will be left out. Only Meta AI, the company’s own assistant, will remain. For Luzia, the Spanish chatbot that reached one million users faster than Instagram thanks to its viral function of transcribing WhatsApp audiosthe measure is a setback, they admit, but minor. “It is not the news that makes us most amused,” admits Álvaro Higes, CEO of the company. “It is not our main channel, but it is one of the secondary channels” where they find users who, due to having modest phones or data problems, depended on WhatsApp as a comfortable and convenient access point. Luzia already has more than 70 million downloads on its mobile applications, its main channel, and maintains “very healthy” organic growth on WhatsApp as well. The platform continues without charging its end users, financing itself with the 30 million euros raised in investment rounds. But that is changing. Contextual ads and in-chat purchases The startup has begun to monetize through contextual ads inserted into conversations“marking very clearly that it is an advertisement,” according to Higes. The logic they apply is that of commercial intent: “If you go to Amazon, you go looking for a specific product. But if you go to Leroy Merlin, you go with a problem and you come out with a product to solve it.” Luzia is in that whole part of the commercial funnel. Brazil is the laboratory. There they have launched a shopping tool integrated into the chat that allows you to browse and buy products from Amazon, Mercado Libre and other stores. When the AI detects that it can recommend a product, it opens a catalog with which you can chat until the purchase is completed. This is where they want to move: pure transactionality. “Eventually we will also release a premium plan,” adds Higes. Better models, better generated images, without limits. But the CEO is skeptical about the paywall as a main model: “It is going to be very difficult to monetize via paywallthe differentiation between AI products is complicated,” he says in line with something we have commented on more than one occasion: the commoditization. Álvaro sees it more as a tool to close distribution agreements in bundle with other services, in the style of Perplexity with Telefónica just a year ago. The two routes are complementary: Contextual ads and transactions on the one hand. Distribution agreements by another. “It is still not a priority, we prefer to prioritize growth,” says Higes, “but it is something we have already started to do.” Forty people, more than half in engineering Luzia’s team is around forty people, distributed mainly between Spain and an emerging presence in Brazil. 60% work in engineering and product. The rest focuses mainly on branding, a department of five people that Higes considers essential: “Unless you are a super technical runner, people usually buy Nike because they think it’s pretty and because they know the brand. In AI it’s a bit similar,” he says as a simile. The analogy makes sense in a market where models are becoming trivialized. Higes invests a lot of effort in what they call the classifier, a system that identifies in real time the topic and the user’s intention to direct them to the most appropriate model. “If you say ‘hello’ to me, I’ll send you a very basic model. If you ask me about mathematics, I’ll send you a different model and give you different tools,” explains the CEO. That optimization is relevant because cost per user is a delicate balance. On the one hand, the price per token of a model equivalent to GPT-4 (more than suitable for the most basic queries) has fallen 90% in two years. On the other hand, the market has also been launching more expensive products that require more inference. “One force counteracts the other,” he summarizes. Higes recently wrote about this idea: There comes a time when the models are good enough for most use cases and there is no longer a need to always play with the most powerful one. “If someone says ‘hello’ to me, we are not going to use GPT-5 to answer ‘hello.’” Instead, the flows of e-commerce that they are building have a higher inference cost, but also direct monetization potential. AI for the 70% who “do not spend the day on a computer” The competition is evident: ChatGPT and Gemini They dominate the market. But Higes sees them as tangential rivals. “We make AI for the 70% of people who are not in front of a computer all day.” The bet is on presenting AI in a more accessible, more intuitive way. They have a specific tool for solving mathematics that generates four times more use than if the user had to ask the question in a blank chat. “We see that people solve more math if you are super clear and say: ‘I can help you solve math.’” It is the lesson they have learned by observing OpenAI data on how people use ChatGPT: Percentages by topic have barely changed from GPT-3.5 to GPT-5. “The level of use is still very limited,” says Higes. “There is a lot of work to present the product in the most intuitive way to the user and remove that cognitive load.” The biggest initial challenge was to change the mental model of users who arrived thinking that the AI was Alexa or Siri. “They asked us the time and asked us to set timers. Many people said ‘this is rubbish, why doesn’t it tell me the time?’, when it can tell you many more things.” Presenting what AI can do is more relevant than the jump from GPT-5 to GPT-6. Maybe people don’t care about that as much as they care about having their problems solved. 2026: transactionality as focus The plan for next year is clear. … Read more
