Nubank Ships AI Agents 20× Faster. The 12% Who Do Return 171% ROI.

The obvious story about Nubank's AI strategy is that it has 135 million customers and more transaction data than most banks in Brazil have ever seen. That part is correct. But the moat isn't in the data. It's in what Nubank does with simulated data nobody else can generate.

In August 2026, at KDD — one of the leading annual conferences on data mining and machine learning, held this year in Seoul — Nubank's engineering team published a paper on what they call SimulationMaxxing: a method for building synthetic evaluation environments for AI agents before those agents touch a single real customer. The headline result is 20× faster agent shipping, with release cycles that once took weeks now taking hours. But the figure that matters for understanding the competitive position is smaller: zero production regressions since the framework went live.

Most financial institutions testing AI agents rely on a process that goes: deploy to a small pilot group, watch what breaks, fix it, redeploy. That process is slow because you need real customer interactions to generate evaluation data — and real interactions gone wrong carry real costs. Simulation-based evaluation bypasses this by building a synthetic version of the customer interaction graph, running the agent against it offline, and catching edge cases before they reach anyone. The catch: a realistic simulation requires a large, diverse dataset of actual interactions to build from in the first place.

Nubank has that dataset. Its five production agents — handling card delivery, debt management, credit-limit support, card management, and product explanations — have generated enough behavioral data across 135 million customers that Nubank can build simulations covering edge cases it would take years to encounter in live traffic. Documented outcomes include a 2× gain in transactional Net Promoter Score for some agents, a 4% increase in self-service rates, and more than ten experiments per quarter per agent without breaking production. The evaluation infrastructure is now the reason Nubank iterates at a pace no other Brazilian bank matches.

For every fintech building AI agents in Brazil today, the implication is precise. Simulation quality doesn't improve with compute — it improves with interaction data. The institution with the largest, most varied dataset can build the most realistic synthetic environment, which means more experiments, faster iteration, and agents that reach production in better shape. Nubank's advantage isn't a better model; it's a feedback loop that compresses the distance between idea and production. That loop runs faster every quarter.

What KDD 2026 actually published is a blueprint. The fintechs best positioned in agentic finance are not the ones with the best engineering teams, but the ones with the densest behavioral datasets. In a market where most Brazilian banks are still running early AI pilots, the institutions that built production systems first aren't just a few product cycles ahead. They are building the evaluation infrastructure that makes every subsequent improvement faster. The moat is the clock speed.