Magie: redefining the conversational banking experience

Magie: redefining the conversational banking experience

Magie: redefining the conversational banking experience

On April 23, 2024, a financial assistant launched in Brazil with no app to download. You saved a phone number, opened WhatsApp, said hello, and a few messages later you had an account.


From then on you paid boletos by sending a photo, made Pix transfers by voice note, asked for your balance the way you would ask a friend what time it is. It was invite-only, aimed at a few hundred people who made five or ten payments a day and were tired of copying codes between apps. In the closed pilot before launch, that small group had already moved about R$13 million.


Two years later the same technology runs behind the brands of banks, telecoms and retailers, in two languages and two countries, for more than a million people. This is the story of how it got there, and what design had to do at each step.

The thesis

Magie was founded by Luiz Ramalho at the end of 2023 on an observation about timing. Four things had matured in Brazil at the same moment:

  • Pix had made instant payments universal and free.

  • Open Finance had made it legal for a person to let one institution act on their accounts at another.

  • WhatsApp was on 200 million Brazilian phones and had just gained native features, like password entry, that made transactions inside a chat viable.

  • Language models had become good enough to understand intention in natural language, including audio and images.


Separately these were four trends. Together they formed a new layer between people and their money: agents that converse, understand and act. The bet was that conversation is the natural interface for money, and that the hardest problems of a challenger bank, getting someone to download an app, learn it and trust a name they have never heard, could be sidestepped by living where people already are.


I joined as part of the founding team and the sole designer. Every step below is a chapter of the same problem seen from a different angle.

2024: banking inside a conversation

The first design problem was that a chat has no screens. There is no hierarchy to lean on, no place to hide a complex flow, no layout to organize information. Every confirmation, every balance, every error has to read clearly in a grey bubble sitting between a meme and a family group.


The early work was about establishing the grammar of a financial conversation: what a payment confirmation must always contain, in what order (amount, recipient, date, then the password request), how money is written, how the agent asks for one thing at a time, how it says no. Much of what later became the platform's voice rules was decided here, message by message.


The second problem was trust in an untrusted medium. A password typed into a chat keyboard is a bad idea, so the moments that carry risk moved into WhatsApp Flows, the native forms that open on top of the conversation. Authentication, confirmation, anything a regulator would want to audit, happens in a Flow. Everything a person just needs to say happens in the chat. That rule was set in the first year and never changed.


Investors read the same timing. Canary funded the pre-seed before launch. In August 2024, Lux Capital made Magie its first investment in Brazil.

2024: banking inside a conversation

The first design problem was that a chat has no screens. There is no hierarchy to lean on, no place to hide a complex flow, no layout to organize information. Every confirmation, every balance, every error has to read clearly in a grey bubble sitting between a meme and a family group.


The early work was about establishing the grammar of a financial conversation: what a payment confirmation must always contain, in what order (amount, recipient, date, then the password request), how money is written, how the agent asks for one thing at a time, how it says no. Much of what later became the platform's voice rules was decided here, message by message.


The second problem was trust in an untrusted medium. A password typed into a chat keyboard is a bad idea, so the moments that carry risk moved into WhatsApp Flows, the native forms that open on top of the conversation. Authentication, confirmation, anything a regulator would want to audit, happens in a Flow. Everything a person just needs to say happens in the chat. That rule was set in the first year and never changed.


Investors read the same timing. Canary funded the pre-seed before launch. In August 2024, Lux Capital made Magie its first investment in Brazil.

2025: from a nice trick to a daily habit

The question for the second year was whether a WhatsApp bank was a novelty or a habit. The answer arrived in June 2025: R$1 billion transacted in the first twelve months, driven by a base that kept growing on word of mouth. Two product moves made that possible, and both were design problems as much as business ones.


The first was Open Finance built as a product. Until then the market had used Open Finance to read people's data and score them for credit. Magie made it something you use: connect the accounts you already have, then pay boletos, bills and Pix with the balance from any of them, see all your balances in one place, schedule payments, put recurring bills on automatic debit from the bank of your choice. Designing it meant making the mandatory friction of the regulation, the consent that must happen inside the other bank, read as a security signal instead of a dead end, and keeping the agent from ever printing another bank's balance on a lock screen. It was the first time the connection itself was the experience.

2025: from a nice trick to a daily habit

The question for the second year was whether a WhatsApp bank was a novelty or a habit. The answer arrived in June 2025: R$1 billion transacted in the first twelve months, driven by a base that kept growing on word of mouth. Two product moves made that possible, and both were design problems as much as business ones.


The first was Open Finance built as a product. Until then the market had used Open Finance to read people's data and score them for credit. Magie made it something you use: connect the accounts you already have, then pay boletos, bills and Pix with the balance from any of them, see all your balances in one place, schedule payments, put recurring bills on automatic debit from the bank of your choice. Designing it meant making the mandatory friction of the regulation, the consent that must happen inside the other bank, read as a security signal instead of a dead end, and keeping the agent from ever printing another bank's balance on a lock screen. It was the first time the connection itself was the experience.

The second was widening the door. Magie+: a loyalty program built entirely in WhatsApp Flows, an app as a complement to the conversation for the things a chat does badly, like charts and reports, and a rebuilt onboarding. The onboarding work is where the design job gets least glamorous and most important: writing every refusal and error a person can hit when opening a regulated account, from a failed security check to a fraud-risk block, so that being told no still feels like being treated as an adult. By the end of the year the base had passed 400,000 users and R$2 billion moved.

2026: MagieOS, the platform behind other brands

By early 2026 the company had learned something the numbers only hinted at. Every organization with customers on WhatsApp wanted what Magie had built, and none of them wanted to build it.


In January, Magie raised US$5 million, passing US$10 million since founding, and announced the move: the same technology offered whitelabel, running under the client's brand, with the consumer product kept alive as the laboratory where the platform keeps evolving. Banks, a telecom, a retailer's fintech, a bank in another country, each deploying its own agent on Magie's rails.

2026: MagieOS, the platform behind other brands

By early 2026 the company had learned something the numbers only hinted at. Every organization with customers on WhatsApp wanted what Magie had built, and none of them wanted to build it.


In January, Magie raised US$5 million, passing US$10 million since founding, and announced the move: the same technology offered whitelabel, running under the client's brand, with the consumer product kept alive as the laboratory where the platform keeps evolving. Banks, a telecom, a retailer's fintech, a bank in another country, each deploying its own agent on Magie's rails.

This changed what design meant at the company.

And three problems appeared at once:


  • Every agent sounded the same. A telecom's agent and a bank's agent answered their customers in the same neutral register, because there was no structured way to define a brand's voice and configure it. I built Lizzie, an internal agent that takes any brand material a client has, a guide, social posts, screenshots, and turns it into a Brand Voice Profile with an engineering handoff: a system prompt in the brand's voice, formatting rules, response templates, error messages, configuration values. Built in five days, without engineering cycles, and used in production before it had a proper interface.

  • The experience diverged across verticals. Onboarding, Pix, boletos, scheduled payments and Open Finance had been designed at different moments by different hands, and an audit of about ninety screens found a de facto design system nobody had written down: six phrasings for the same confirmation, money formatted three ways in one journey, a receipt in English next to one in Portuguese. The answer was the Magie Conversation System: voice rules, a catalog of the moments every financial conversation goes through, and blueprints per vertical, mapped onto how engineering actually assembles prompts, so the rules live where the agent reads them.

  • The product had to cross a border. Taking the platform to a Spanish-speaking bank meant localizing the entire conversational experience, from vocabulary to the assumptions behind each flow, and running QA on a first-payment journey with a client whose expectations about how a bot should behave differed from ours.

This changed what design meant at the company. And three problems appeared at once:


  • Every agent sounded the same. A telecom's agent and a bank's agent answered their customers in the same neutral register, because there was no structured way to define a brand's voice and configure it. I built Lizzie, an internal agent that takes any brand material a client has, a guide, social posts, screenshots, and turns it into a Brand Voice Profile with an engineering handoff: a system prompt in the brand's voice, formatting rules, response templates, error messages, configuration values. Built in five days, without engineering cycles, and used in production before it had a proper interface.

  • The experience diverged across verticals. Onboarding, Pix, boletos, scheduled payments and Open Finance had been designed at different moments by different hands, and an audit of about ninety screens found a de facto design system nobody had written down: six phrasings for the same confirmation, money formatted three ways in one journey, a receipt in English next to one in Portuguese. The answer was the Magie Conversation System: voice rules, a catalog of the moments every financial conversation goes through, and blueprints per vertical, mapped onto how engineering actually assembles prompts, so the rules live where the agent reads them.

  • The product had to cross a border. Taking the platform to a Spanish-speaking bank meant localizing the entire conversational experience, from vocabulary to the assumptions behind each flow, and running QA on a first-payment journey with a client whose expectations about how a bot should behave differed from ours.

Underneath all three sat a change in how I work. With one product, the designer decides and engineering implements. With a platform, the designer's job is to make the decisions transmissible: written as rules, structured as data, testable on every prompt version. Design stopped being a set of screens and became part of the infrastructure.

My role across the arc

I was the founding designer and remained the only one until 2026. In 2024 that meant the conversational grammar, the brand and the first journeys. In 2025 it meant Open Finance, onboarding, error language, surveys, the flows nobody sees until something goes wrong. In 2026 it meant systems: a voice tool, a conversation system, localization, and the go-to-market material that explains to a bank why a partner model beats building from scratch. Throughout, I worked directly with engineering, with the team that tunes what the agents say, and with GTM, in a company small enough that the same person writes the error message and the customer story.


Where it stands

Magie today describes itself as the first financial assistant in Brazil on WhatsApp, with more than R$2.5 billion transacted and half a million people using it every day. The user base crossed one million in 2026. The consumer product that started as an invite-only tool for heavy payers is now the proving ground for a platform that lets other organizations offer banking inside a conversation, in their own voice, without writing a line of it.

My role across the arc

I was the Founding Designer, and in 2024 that meant the conversational grammar, the brand and the first journeys.


In 2025 it meant Open Finance, onboarding, error language, surveys, the flows nobody sees until something goes wrong.


In 2026 it meant systems: a voice tool, a conversation system, localization, and the go-to-market material that explains to a bank why a partner model beats building from scratch.


Throughout, I worked directly with engineering, with the team that tunes what the agents say, and with GTM, in a company small enough that the same person writes the error message and the customer story.


Where it stands

Magie today describes itself as the first financial assistant in Brazil on WhatsApp, with more than R$2.5 billion transacted and half a million people using it every day. The user base crossed one million in 2026. The consumer product that started as an invite-only tool for heavy payers is now the proving ground for a platform that lets other organizations offer banking inside a conversation, in their own voice, without writing a line of it.

Clients

0M+

Using MagieOS modules across B2C and B2B products.

Processed

USD0B+

In transactioned volume across B2C and B2B products.

The bet from 2023 has held. The channel was never the point. The intelligence behind it was, and the two years since launch were spent making that intelligence sound like a bank people trust, then like several banks at once.