In Brazil, 83% of banking transactions are carried out digitally, 78% of those on mobile devices, and mobile banking volume has grown 169% over the past five years. Instant payments are standard, open finance is live, and the sector remains the country’s largest technology investor. On paper, this is exactly the kind of digital maturity that should make it easy to scale AI. In practice, it hasn’t been enough.
That gap was the central topic when the CEOs of seven of Brazil’s largest banks appeared together in São Paulo on August 24th, at the opening session of Febraban Tech 2026. The conversation wasn’t about roadmaps or model selection. It was about three things that consistently separate a successful AI pilot from an AI capability that actually runs at scale: modernizing the systems AI depends on, redesigning the customer journeys AI touches, and building trust into the architecture from the start, not bolting it on at the end.
The Distance Between Investment and Impact
Budget isn’t the constraint. In 2026, Brazilian banks planned to spend 50.4 billion reais on technology , up 8% year-over-year and 58% over five years, with roughly 3 billion Reais earmarked specifically for AI, analytics, and big data. 84% of institutions already treat generative AI as a strategic priority.
The real bottleneck is conversion, not investment: turning pilots that work in a controlled environment into capabilities that hold up in production. Three factors determine whether that conversion succeeds.
1. Modernize the Core Before You Scale the Intelligence
An AI agent cannot outperform the systems it’s connected to. When business logic is buried inside legacy applications with limited APIs, thin documentation, and unreliable data, a pilot can look impressive in a demo and still fail the moment it touches real operations.
What’s changed is the economics of fixing that. AI now accelerates the modernization work itself, reading legacy code, generating tests, producing documentation, refactoring applications, which makes projects that were once too slow or too expensive to justify newly viable. That’s the case for treating modernization as the invisible engine behind agentic AI: organizations need systems that can support intelligence before they can scale it.
2. Redesign the Journey, Don’t Just Automate the Old One
Even with the right technical foundation, scope is still a trap. Layering a chatbot onto a poorly designed service process doesn’t fix the process; it just makes a bad experience faster.
The 2026 Global AI in Financial Services Report from the Cambridge Centre for Alternative Finance shows exactly where most institutions are stuck, still optimizing internal operations rather than rethinking how work gets done. Here’s the share of most frequent AI applications that have reached the pilot stage or further, by function:
- Process automation: 79%
- Data visualization: 75%
- Software engineering: 75%
- Data and knowledge management: 69%
On the customer-facing side, AI-powered support leads adoption at 74%, fintechs are ahead at 82%, versus 67% for traditional institutions. In risk and compliance, fraud detection (57%) and credit risk modeling (54%) remain the most common applications.
The pattern is consistent: AI is being used to optimize existing workflows, not to rethink them. The real opportunity opens up when organizations design customer journeys around what AI now makes possible, instead of automating the process they already had.
3. Trust Is Infrastructure, Not a Final Checkpoint
This was the point of greatest agreement among the CEOs on stage. Operational collaboration across fraud, cybersecurity, and AML functions is already well established: 88% of banks have integrated these functions, with teams collaborating directly on incident response and information sharing.
One example raised at Febraban Tech 2026: VU’s work with Volkswagen Financial Services Brazil, a digital financing platform combining digital onboarding, KYC, electronic signatures, and biometrics to validate identity and contracts while strengthening fraud prevention.
The lesson isn’t the biometrics, it’s the sequencing. Trust was designed into the journey from day one, not added as a compliance gate at the end. As AI systems take on more autonomy, that principle becomes non-negotiable: trust has to be architecture, not a validation step.
What Comes Next
AI adoption is moving from assistants that help with single tasks to agents that execute multi-step work across systems, under human supervision. That shift changes the real question from what can this system do to can we safely govern what it does, who oversees it, where its autonomy ends, and how its decisions get audited. That’s why Febraban Tech 2026 treated human leadership as a precondition, not a supporting theme: it’s what determines whether modernization and journey redesign actually translate into results.
If your organization has AI pilots that work but haven’t reached production, the cause is rarely the AI itself. It’s usually one of the three gaps above, legacy systems that can’t support it, a journey that was automated instead of redesigned, or governance that wasn’t built in from the start. That’s the model behind Globant’s AI Pods: task-specific units run by AI agents and supervised by human experts, built on the same principle Febraban Tech 2026 kept circling back to, autonomy that stays accountable.
