It is no longer news for banks to integrate generative AI into their customer service channels. The real challenge lies in ensuring that this technology remains consistent and controllable when used for bill viewing, credit card inquiries, insurance processing, and customer transfers to human representatives, while also keeping costs under control. On October 16th, OpenAI released a case study of Itaú Bank in Brazil, detailing the phased progress of its conversational service ia.i and its internal governance platform IARA. In the first phase, the service reached approximately 300,000 customers, and the bank aims to cover all its customers by the end of 2026. These two initiatives should not be confused as the same thing. What makes this case study particularly noteworthy is not so much the fact that the bank used a large model, but rather how a major financial institution managed to integrate the model, customer data, business processes, and human services together.

Starting from 2024, Itaú will incorporate intelligent capabilities to meet specific financial needs, with ia.i serving as the unified entry point for the next phase. Customers can ask questions about accounts, credit, bank cards, and insurance via text or voice, and if necessary, they will be referred to specialized personnel. The bank also aims to integrate clicks, images, and conversations into a cohesive experience, so that users do not have to first figure out which department within the organizational structure a particular service belongs to before searching for the corresponding menu. The "financial situation of approximately 70 million customers" mentioned in the OpenAI case refers to the Customer DNA background layer used by the bank to understand customer needs and habits, which does not mean that 70 million people are already using ia.i. For large financial products, separating potential coverage, existing pilots, and actual active usage is the first step in measuring the degree of implementation.

In the first phase involving 300,000 people, what is being verified is the process, not a fully automated advisor.

Banking services are not about general Q&A. When a customer asks, “Where does this deduction come from?”, the system must first identify the transaction, confirm authority, and then explain it in language that humans can understand. If the question relates to investment advice, the applicable risk warnings, suitability assessments, and manual reviews are completely different. A chat box can serve as an entry point, but it cannot replace the underlying authorization processes, record-keeping, or responsibility allocation. The Itaú case combines text, voice, and operations, while also maintaining a pathway for handing over to professionals. This indicates that its goal is to shorten the distance for customers to access the correct services, rather than claiming that AI can independently make all financial decisions.

The first batch of about 300,000 customers also provides banks with an opportunity to observe real user behavior. Are users willing to entrust complex issues to these dialogue interfaces? Can a single question lead to an actionable next step within the authorized scope? When incorrect answers are given, can the system quickly identify the source and direct the issue to human personnel? These questions cannot be answered through demonstration videos. Especially in credit and investment scenarios, fluent language can easily create the illusion that "the system has understood the entire context," but the actual decisions still depend on whether the account data is up-to-date, whether the rules are effective, and whether the customer agrees to use those rules. Pilots are more suitable for accumulating evidence regarding these boundaries, rather than treating coverage targets as a achieved scale.

Itaú didn't just choose one model. According to the OpenAI case, the bank selected different models based on latency, cost, and reasoning requirements, with the GPT model being one of them. Daily balance inquiries require speed and low cost, while complex tasks that involve compliance checks need more robust reasoning and basis. If all requests are sent to the most expensive model, operational costs could exceed customer value; if they are all handled by lighter models, difficulties may arise in terms of security and accuracy. Hierarchical routing, permission verification, and manual upgrades are more important here than the stunning performance of a single response.

What IARA really needs to address is governance accounting on a scaled-up scale.

The internal IARA platform of the bank is responsible for model integration, orchestration, monitoring, and policy enforcement. According to case disclosures, its legal affairs department handles a database of over 90 million documents and processes more than 2 million tasks per month; efficiency improvements in certain process steps have been reported to range from 20% to 80%. This range applies only to specific departmental processes and does not represent a similar improvement in the efficiency of all bank employees, nor is it an independent industry benchmark for auditing. Once legal documents involve customer rights and contract interpretation, preserving the source, version, and human approval is even more crucial than saving a few minutes.

The governance platform also determines whether banks can clearly see how much money is spent on each model call, what data is used, and whether any authorized limits are exceeded. Privacy and responsibility in the financial industry cannot be addressed simply by including a warning phrase like "Please proceed with caution." It is necessary to control access to sensitive information before it enters the models, to integrate the output with business validation processes, and to include exceptions and deviations in continuous monitoring. As a provider, OpenAI emphasizes the use of multiple models and responsibilities in these cases AI, but since these are still customer cases published by partners, without comparable error rates, complaint rates, and retention data, it cannot be concluded that this approach has been validated across all financial scenarios.

Future plans include broader customer coverage, product simulation, and periodic financial analysis. With each step forward, the system moves from providing "information" to offering "recommendations and execution," and with that comes an increased level of risk. Explaining spending trends to customers requires a different factual basis and regulatory considerations than recommending changes to their financial products; automating application processes, on the other hand, involves handling consent, cancellations, and appeals. The practical value of the Itaú case lies in bringing these stages together on a single product roadmap, while also conveying to the market that the real competition is not about which bank installs a chat feature first, but about who can still clearly explain the source of each decision and who is responsible for the next steps even after widespread use.

Source: OpenAI, October 16, 2026 Customer Case " How Ita ú is building generative AI for its entire customer base ": https :// openai.com / index / itau-generative-ai /