Best AI banks in LatAm ranked

Nubank leads all Latin American banks on AI maturity, edging out in-country rival Itaú Unibanco, the new Evident AI Index for Banks - LATAM shows. 

The Index, Evident’s second regional ranking (following last month’s Middle East and Africa Index), assesses 20 Latin American banks on the AI talent they’ve amassed, their innovation efforts, the tech leadership of their top executives and the guardrails they’ve set up to govern this technology effectively.

Without further ado, here’s the region’s top 10:

Latin American banks have, by necessity, a flair for improvisation. People’s financial lives move through fintechs and traditional banks alike. Business and payments increasingly run through phones. And after decades of economic troubles, trust is low. So banks here started from a very different place in the AI race than peers in North America and Europe. 

Only 39% of people across eight Latin American countries said they trusted financial institutions, research from Peru’s Credicorp – ranked fifth – showed. AI, meanwhile, gets a warmer reception than you see in the U.S. or across the Atlantic these days: Brazil, Peru and Colombia were among the countries most excited about the technology, according to polling firm Ipsos. In the U.S., banks are trying to introduce AI tools without burning the trust they already have. In Latin America, they’re betting the technology can help them earn some. 

Nowhere is that more apparent than in Brazil, home to half the Index’s top 10 and its four leading banks. Historically, the country’s incumbents had left a chunk of the population poorly served, a gap that gave Nubank the opportunity to lean into AI and build a cheaper alternative that could grow fast (see: “Nu world order,” The Brief, June 25). Its success revealed just how up for grabs this market is: The average Brazilian has relationships with roughly six financial institutions, up from 3.5 in 2020. Customers shift their activity toward whichever one is cheapest, safest or most useful, meaning a good AI tool can have an immediate impact on a bank’s business. 

Incumbents didn’t take Nubank’s rise lying down. Leaning on Pix, the instant payments network the central bank rolled out in 2020, big banks used transaction data to power new models and design new AI tools: Itaú uses AI to warn when a transfer looks unusual or is headed toward an account linked to scams. It, along with Banco Bradesco, built new credit models with AI, which are helping grow their business (see: “Extra Credit,” The Brief, July 16). Banco do Brasil has a generative AI tool that lets customers send a payment via a voice note or a picture on WhatsApp.

The conditions supporting Brazil’s lead aren’t evenly spread across the region. In Mexico, the financial system is modernizing fast, but the Spanish banks – namely BBVA and Santander – are snapping up AI talent in droves (see: “The Spanish are coming (back),” The Brief, July 2). With it, the Spanish banks are scaling AI tools into the market faster than the regional banks can build homegrown tools from scratch.

Argentina’s banks, meanwhile, have a steeper deficit than 1-0 after the 106th minute to erase. After so much sky-high inflation and economic mismanagement, Argentinians have pushed hundreds of billions of dollars outside of the country’s formal financial system. It leaves the country’s banks with a more basic task than proving AI can improve banking: proving it makes sense to keep money in a bank at all.

Looking ahead: The difference between Brazil and Argentina shows how uneven banking is across the region. But also how dynamic it is. As trust in banks and uptake of AI shifts, the tech could narrow those gaps – or widen them further.

🚀 Builder: David Vélez
Founder, chairman and global CEO at Nubank

Vélez built Nubank into a 135 million-customer giant by attacking high fees and financial exclusion in Latin American banking. AI powered that growth, particularly in credit, where the bank’s foundation model, NuFormer, was used to improve underwriting and let the bank serve a wider customer base than the incumbents (see: “Nu World Order,” The Brief, June 25). Now, Vélez has his sights set on a U.S. expansion. Cracking that market will be a test of how the AI that’s let it grow in Latin America can scale to other businesses. “Our goal is to be the leading in the world in using AI for financial services,” he said in February. 

🏗️ Architect: Cíntia Scovine Barcelos
CTO at Bradesco

Barcelos championed Bradesco’s rollout of a platform that lets the bank build and scale AI tools. The bank says it has 600 of them now in production around the firm.  Among them is Pix Inteligente, a tool which uses AI to validate balances and perform security checks before transactions. But the key isn’t one tool, or even hundreds of them; it’s the “shared foundation that enables reuse, standardization and governance,” she wrote last month.

💸 Scaler: Raimundo Morales
CEO at Yape (a Credicorp company)

Morales turned a digital wallet into a digital payments infrastructure that 16 million Peruvians – roughly 80% of the economically active population – use to send money from one place to another. His ambition isn’t to compete with the banks, but to “beat cash” by making digital payments the default way money changes hands in the country. With that level of use, his firm is compiling a huge amount of transaction data that will be critical as parent company Credicorp rolls out new AI tools. “The Peruvian economy depends on us,” Morales said.

👔 Governor: Roberto Campos Neto
Global head of public policy at Nubank

Campos Neto was the head of Brazil’s central bank when it rolled out Pix, the payment rails that allowed the country’s banks to process payments instantly and Open Finance, which lets consumers securely share financial data with fintechs and banks. Both gave banks a way to harvest data that became critical to the AI buildout. Now, out of government, he’s helping Nubank expand through Latin America and the U.S. “What began as workflow automation is rapidly becoming something else, AI as a generator of ideas, not merely an executor of tasks,” he wrote.

🧪 Scientist: Roberto Frossard
Head of emerging technologies at Itaú Unibanco

Frossard is responsible for turning frontier AI into things bank customers can actually use at Itaú. His “research-to-product” model lets the bank test tech against real customer problems without it dying in the lab. One recent tool lets customers pay by sending screenshots to WhatsApp, part of a bigger roadmap of multimodal AI, he wrote. His goal is to use the scientific method to anticipate the impacts of the bank’s AI builds: “When we learn faster, impact follows,” he wrote.

Joseph McCarthy would smile knowing the most pressing question in business today is, “Are you now or have you ever used a Chinese model?” He’d be less impressed that plenty of banks would like their answer to be yes.

China’s Moonshot AI last week dropped Kimi K3, a new open weights model on par with Anthropic and OpenAI’s best. OpenAI strategist Dean Ball warned that it could usher in “full AI communism.” Users were less worried: Demand got so heavy Moonshot paused new subscriptions. 

For banks, Kimi is a big deal, though not because they should rush to move every workload onto it when the full details arrive next week. Rather, it’s a dress rehearsal for what happens when the best model for a valuable piece of bank work is Chinese, but the bank’s systems, risk team or the government won’t let it through the door. 

Kimi hasn’t forced that conversation just yet. It ranks near the frontier and even scored better than GPT-5.6 on an office-work benchmark test that asks the model to search thousands of company files and produce spreadsheets, presentations and digital prototypes. But just because it’s a Chinese open model, doesn’t mean it’s bargain-basement. K3 used so many extra steps to solve the test that it ran three times slower than GPT-5.6 and ended up costing twice as much. In other words, it’s cheap by the token and expensive by the job.

On top of that, to get the best performance from K3, Moonshot recommends running it across 64 advanced processors – which is about $4 million worth of Nvidia chips, plus the power, networking and specialist staff required to set it up. It adds up to paying a lot for the third-best option and creates what Alexandre Dos Santos, BNP Paribas’s chief AI architect, calls the “open-weight facade,” where the models of Kimi’s ilk look more attractive on paper than they actually are because of the real cost of running them.

Still, banks can’t afford to brush off Kimi K3’s raw performance. The model might not be right for the enterprise today, but the next open weights model from China may well be. Picking a model is “starting to look a lot more like hiring for a role,” wrote the Wharton School’s Ethan Mollick last week.

This week, OpenAI revealed that models it was testing broke containment and hacked open source tool provider Hugging Face. In this week’s “Corner,” we look at a Capital One tool, open-sourced this past week, which can hunt down the kind of vulnerability that lets a system get breached and direct the bank on how to fix it. 

Use Case: VulnHunter
Vendor: Anthropic
Bank: Capital One

Why it’s interesting: AI security tools are good at finding potential vulnerabilities. The problem is they find so many that developers spend a good chunk of their time just proving that most of them are not. Capital One designed its tool around that bottleneck, building in multiple layers that stress test every alert before actually sending the threat back to the person that needs to check it out. At that point though, the AI has done enough poking around that it generates what it thinks the fix should be, which saves extra time.

How it works: When a potential threat gets uncovered, VulnHunter puts itself in the mind of an attacker and looks at places a bad actor could’ve gotten in. If those doors are locked, it can determine it’s a false alarm. If it can’t rule that out, it passes it through a “falsification engine” which tries to disprove the tool’s conclusion by looking for faulty assumptions in its logs. If the falsification engine can’t disprove that it’s a real threat, it gets passed to a developer, with suggested code for how to patch it. 

How they did it: Capital One used three Claude Code skills: One that hunts for vulnerabilities, one that develops and tests a fix and one that works like a cleanup crew to see if the repair worked. Rather than asking developers to learn a new platform, it embedded it inside of Claude and existing coding workflows.

By the numbers: The bank has used the tool to fix vulnerabilities across “thousands of repositories, spanning tens of business areas.”

Bigger picture: Security tools are often kept close to the vest at banks and very rarely open sourced, the way Capital One has here. But with OpenAI’s own model breaking containment this week as a preview of the next wave of cyber threats, more banks might follow suit. “No single organization can solve this challenge alone,” wrote Chris Nims, Capital One’s CISO. “Stakes are only rising for security teams to fight AI-enabled threats with equally capable AI-driven defenses to protect our digital environments.”

That was the hypothetical payoff Santander used to test whether four back-office tasks should be fully automated, in new research out this week. For three of them, even halving the cost wasn't enough to justify taking humans out of the loop. To reach that conclusion, the bank looked at four jobs – triaging IT incidents, reviewing architecture, coaching employees and processing invoices – and asked whether doing each at half the cost was worth the risk of taking humans out of the loop. Only processing invoices cleared that bar. For triaging incidents, people were too important a last line of defense. In architecture reviews, 80% of the work involved exceptions that needed human judgment. And in coaching, they determined handing the role over to AI wouldn’t fly with employees. 

Bigger picture: The conversation around AI ROI is advancing beyond how much faster it makes people to whether the process it gets applied to is worth automating. OpenAI CFO Sarah Friar took a stab at the quantitative side with a new scorecard this week. She suggests that businesses measure “useful intelligence per dollar,” a metric that combines how much work models can do with how dependable they are and compares it to the cost. Both point to the conversation getting more nuanced: “I don't care anymore about saving 10 minutes from a task,” one bank executive told us. “I’m now interested in where we can use this technology to do work that you never could before.”

Financial services firms are showing off their AI returns this earnings season. Some highlights from the past week’s results:

• Santander’s AI use has generated $96 million through the first half of the year, the Spanish bank reported Wednesday. That’s up from $40 million through the first quarter. Santander remains the only bank breaking out AI’s value – measured here by new revenue, cost savings and how much less it has to set aside to cover loan losses – on a quarterly basis.
  
• State Street says it’ll get $1 billion worth of benefits by 2029 thanks to AI and its operating model transformation. Productivity gains will make up $750 million of that, while the remaining $250 million will come from new revenue, the bank said. With the announcement, State Street became the 16th of the 50 lenders in the Evident AI Index for Banks to report an overall realized or projected ROI figure.
  
• In insurance, Travelers saw a 0.5-point improvement in its loss ratio, a measure of how much it pays out in claims for every dollar of premiums it collects,  thanks to AI, CEO Alan Schnitzer said. The firm recently rolled out TravelersLLM, its in-house foundation model trained on insurance data, which now powers some underwriting and claims decisions. Using it for insurance-related tasks is “more efficient and more cost-effective than relying on frontier models alone,” said Mojgan Lefebvre, the firm’s chief technology and operations officer, in an interview this week.

Microsoft is deepening its partnership with Mistral to “expand AI infrastructure in Europe.” Microsoft will spend billions to use some of the computing power that becomes available as the French AI lab builds more data centers across the Continent. Mistral, in turn, gets better distribution: Microsoft will make more Mistral models available in its products, the announcement said. For banks, the arrangement creates more choice over where AI tools run and where data gets stored – an increasing concern as regulators turn up the temperature on so-called AI sovereignty. Keeping data closer to home can make it easier for banks to meet ownership rules and reduce reliance on U.S. infrastructure.

Wells Fargo’s wealth and investment management unit launched an “AI teammate,” a copilot advisors can use to navigate the firm’s tools and systems. It sits on top of Advisor Gateway, the platform the firm rolled out in May, which brought each of the firm’s tools under the same roof (see: “Superadvisors,” The Brief, May 14). The teammate follows deterministic paths, meaning what data it’s allowed to access and what actions it's allowed to take follow predetermined rules rather than allowing the AI to decide on its own, Andre Mansour, head of AI for the wealth and investment management unit, told Evident. As part of the build, Wells Fargo used AI to generate first-draft UI components, “taking what would have been weeks of prototype work down to hours,” Mansour said.

“The market misunderstood the impact of DeepSeek the first time. It’s misunderstood the impact of Kimi again this time…Great models lead to great use which leads to great growth, and that’s just the starting point. That’s what happened with DeepSeek. It’s going to happen with Kimi.”

– Jensen Huang, CEO at Nvidia, in an interview, July 22

OpenAI appointed BNY CEO Robin Vince and Nubank CEO David Vélez to its board of directors. The pair of bank bosses brings “complementary perspectives on how technology can reshape industries,” a statement about the move said. Both BNY and Nubank have existing relationships with the ChatGPT-maker. Nubank worked with OpenAI on a call center copilot that cut the response time on two million monthly chats by 70%, the firm reported. BNY built a tool with the firm that reduced the time it takes lawyers to review each of the more than 3,000 vendor agreements it assesses per year from four hours down to one.

Bank of America appointed Kevin Milsom to be head of platforms AI transformation. He was most recently head of platform development and AI products and will report to Ashok Krishnan, the bank’s head of platforms for global markets, in the new post. Amy Avery, who leads the bank's Analytics, Modeling and Insights unit, is bringing her team into the Global Markets team as well.

Deutsche Bank hired Dara L. Sosulski as head of AI for operations and controls across the corporate and investment banking unit. She was previously MD and head of AI and model management at HSBC.

Hiran Ganegedara was hired by Westpac as its head of AI, consumer. It’s a return to banking for Ganegedara, who spent more than a decade in data science roles at CommBank before taking over as head of data science and engineering at LMG in 2023.

Leila Bekkouche Merdassi is now head of AI specialists at UBS, leading a team that will “partner across the organization to accelerate AI adoption, develop reusable AI capabilities, enable transformational solutions, and drive innovation at scale,” she wrote on LinkedIn.

Ashima Bhalla is now head of risk and investment oversight engineering and Aladdin product engineering at BlackRock. She was previously head of research, risk and analytics technology at Goldman Sachs.

JPMorganChase is hiring for an executive director of AI transformation within its payments unit.