Cognition Expands into Latin America: Devin Already Runs at Itaú and Nubank

Cognition, the creator of the autonomous AI engineer Devin, has announced its expansion into Latin America. The announcement took place last week at the MASP museum in São Paulo, in front of executives from large companies, government agencies and tech firms. The company's main argument is its client case studies. At the bank Itaú, more than 75% of tech teams use Devin, while at Nubank a migration of a monolith with several million lines of code took weeks instead of years and cost more than 20 times less. All figures come from Cognition itself and have not been independently verified.

Why This Region

Cognition describes Latin America as one of the most mobile-first markets in the world. WhatsApp, digital banks and new financial products spread there very quickly, while the systems underneath them lag behind. A customer in Brazil pays for lunch with a QR code, yet the bank processing the payment still runs on COBOL.

The developer market is also huge: 15.8 million GitHub accounts, almost five times more than in 2020. That makes it the third-largest community after the US and India. But by Cognition's logic, hiring alone cannot solve the problem. What's needed is a different way to write, modernize and maintain code, and Devin is positioned as exactly that kind of tool.

What the Client Case Studies Show

The company names Itaú, Nubank, Santander, Natura and EBANX among Devin's customers in the region. The most detailed figures cover two banks:

• Itaú: more than 75% of tech teams use Devin;
• over 300,000 repositories have been documented;
• migration of legacy services from .NET to Java runs 6 times faster;
• about 70% of security vulnerabilities are fixed automatically;
• across the bank as a whole, engineering throughput has grown by 20–30%;
• Nubank (about 140 million customers): migrating a monolith of several million lines of code dropped from years to weeks, and the cost fell by more than 20 times.

All of these tasks — documentation, migrations, vulnerability fixes — are limited in scope and verifiable: there are tests and there are completion criteria. Autonomous agents perform best on exactly this kind of work. The methodology for measuring "throughput" has not been disclosed, nor has the baseline it is compared against. The 20–30% figure for the bank looks plausible, but it is more accurate to treat it as an estimate from the client and the vendor.

Cognition's Plans in the Region

The São Paulo team will work with customers across Latin America and invest in the local developer ecosystem. The company is hiring across the region, including deployed engineers: engineers embedded in client teams. For deploying an agent in a regulated environment, this is a key element. Without people next to the codebase of a bank or government agency, a pilot often gets stuck on access, review processes and security requirements.

Cognition's message is aimed at local banks, telecom operators, retailers and governments: why import software written somewhere else when the region has developers, a market and now the tools.

Where Teams Considering Agents Should Start

If you are responsible for engineering at a company with a lot of legacy code, what's worth copying is not the slogans but the choice of tasks. Start with migration, documentation or vulnerability fixes, where the result can be verified automatically. Before the pilot, record baseline metrics: time per task, share of accepted changes, and cost per merged pull request.

Check test coverage: without it, the agent has no signal that it hasn't broken anything. Agree in advance with the security team on access, data storage and code review rules. And calculate the economics over the full cycle, including the time of the people who review the output: that is usually where the gap between the claimed and the real effect hides.