Generative AI in banking: what Itaú’s acquisition of NeoSpace teaches us

Summary

• Itaú's purchase of NeoSpace is the headline, but the lesson is in what supports generative AI inside a bank.
• Large banks already use the technology beyond customer service, with caveats about scale and return.
• Results depend on structured data, integration with legacy systems and governance with human review.
• For a midsize bank, the path is to start with one process, not with a purchase.

Itaú bought a generative AI startup, and the lesson is not the purchase

On October 26, 2026, it was reported that Itaú bought 100% of NeoSpace, a startup from Uberlândia (MG), for R$ 2.92 billion. The company develops generative AI models, technology that produces text, summaries and answers from large volumes of information. NeoSpace continues to operate independently. According to Neofeed, cited by the Baguete website, the bank's AI platform is expected to reach all individual customers in the app by the end of 2026. How Itaú will use the startup has not been disclosed in detail. What matters to decision-makers at other banks is the question the purchase raises: what needs to be in place for generative AI to work in the operation of a financial institution?

Large banks already use generative AI beyond customer service

At Febraban Tech 2025, executives from large banks described how they use the technology, in a Finsiders Brasil report from June 16, 2025. The overall reading: generative AI has moved from chatbots to areas such as auditing, development, testing and business onboarding.

Area	Example reported by the bank itself
Customer service	Bradesco's BIA, in its generative version, serves 3 million customers, with a resolution rate of up to 90%
Back office	At Santander, agents cut business onboarding from up to 40 hours to about seven minutes, in some cases
Process analysis	At Caixa, analysts went from one or two analyses per day to up to 20
Software testing	Generating test cases and synthetic data cuts scenario design from weeks to minutes

  These are figures reported by the banks themselves in 2025, and they cannot be compared with each other. The same report notes that scaling safely and with governance is the main obstacle and that doubts remain about real return, because computing costs and errors at scale are expensive.

Under generative AI: organized data, systems that talk to each other, and a trail

The articles we read point to three prerequisites:

• Structured data. At Febraban Tech, Luiz Humberto Sucupira, of Banco do Nordeste, said that without structured data there is no robust generative AI, according to Diário do Comércio (06/14/2025).
• Integration with legacy systems. According to Let's Money (03/14/2026), banks combine mainframes, cloud and mobile apps, and connecting generative AI to these legacy systems requires investment in modernization. The analysis comes from Hector Arias, of Red Hat.
• A trail of decisions. In an opinion piece in ABF Journal (04/24/2026), Kris Kowal, a SAP executive, states that AI agents that depend on fragmented systems perform poorly and that working with data now includes compliance and auditability. This is the view of someone who sells software, and it is worth reading that way.

AI is the top layer. If data is scattered and systems do not talk to each other, it inherits the mess.

Governance: someone needs to check what the AI answers

At an EY panel at Febraban Tech (06/11/2025), reported by Diário do Comércio in content from Agência EY, executives called for continuous vigilance. Giuliane Paulista, of Banco do Brasil, said that post-deployment verification gets less attention than development and that the technology can produce hallucinations - answers that sound right and are not. Telma Luchetta, of EY, explained red teaming: AI agents, supervised by humans, test the model for vulnerabilities. Marco Aurélio Pereira, of BV, warned that many companies use AI without governance and without everyone knowing. Three points from the panel that any bank can define before starting:

• One owner per business area, responsible for AI use there.
• A level of review proportional to how critical the use is.
• A clear rule on which data can or cannot be exposed to an AI tool.

What changes for a midsize bank that will not buy an AI startup

A midsize bank has less room for error. The ABF Journal piece, written by a vendor executive, says banks of this size need solid foundations before scaling AI, and we agree with the direction. Three questions help get started:

1. Which process to test first? One with low risk and gains that are easy to measure.
2. Where is the data for that process? And which of it cannot be exposed to an AI tool.
3. Who owns the result? And who reviews before the answer reaches the customer or the system.

If you would like to talk about the systems foundation these questions require, get in touch.

What Espresso understands about banking systems, and how we help

The question Itaú leaves hanging, what needs to be in place underneath, runs almost entirely through systems: organized data, integrations, an audit trail and operations that cannot stop. It is a field we know well. Espresso works with fintechs and financial-sector systems, especially in outsourcing and support of critical systems. This applies to any bank that wants its foundation in order before any new initiative.

• Staffing and Outsourcing: developers or a full team integrated with the bank's team, starting in up to 30 days. Learn about the service.
• Maintenance & Hosting: support for critical systems in production, today with more than 57 systems under continuous maintenance. See the service.
• Software development: integrations and APIs between the systems the bank already has. See the service.
• Discovery & Design: understanding the process and the data before writing code. Learn about the service.

If the challenge is having people to run the project, also see how to choose an IT outsourcing vendor for fintechs and how to modernize legacy systems without stopping operations.

Frequently asked questions

What is generative AI in banking? It is the use of models that generate text, summaries and answers in bank tasks, such as customer service, process analysis, software testing and business onboarding. Unlike models that only classify or predict, it produces new content, which is why it requires review and rules on where it can be used. Does a midsize bank need to buy an AI startup like Itaú did? There is no rule. The purchase was Itaú's choice, and the report does not say it is the path for others. For a midsize bank, the safest approach is usually to start with a low-risk process, with known data, a defined owner and human review, and only then decide whether to build, hire or buy. What are the main risks? The panels and reports cited point to hallucination, meaning wrong answers delivered with confidence, AI use without governance, and exposure of personal and strategic data in unsuitable tools. There is also the cost of errors at scale and doubt about the real return on investment. Where to start? With three questions: which process to test first, where its data is, and who answers for the result. With that, you can choose the first case and define how to measure the gain before hiring any larger solution. This text provides general information and does not replace legal, regulatory or information security advice. Reference date for the news and data cited: 10/27/2026.