OpenAI published a case study on generative AI at Itaú Unibanco on October 16. It focuses on the conversational interface ia.i and IARA, a governance platform providing controlled access to different models. Reaching the entire customer base is described as a goal for the end of 2026, rather than a completed rollout. This is an account of how a real product is organized, not a recommendation to use financial services.
What the dates establish
Itaú’s press section dates the original ia.i announcement to July 27. That release described approximately 300,000 people as an initial group and expansion as a future plan. The October case study also uses the number for the first phase. We have no basis to present it as a newly measured active-customer count today or to assume that every intervening target was achieved.
This distinction matters when evaluating AI deployments. The date of an account about a project can differ from its launch date. Invited customers, people granted access, and actual users are also different measures. Without an explicit definition, a number can sound more impressive than the available evidence supports. Historical context should therefore accompany a current case study rather than be silently converted into a fresh launch.
Conversation does not remove the rest of the interface
The bank’s own October15 explanation calls ia.i a beta. AI-generated responses are labelled and may contain inaccuracies; customers can request human support. It also describes conversation coexisting with conventional navigation: when clicking is more suitable, the interface directs the customer to the relevant app function.
Our editorial interpretation is that this raises a more useful question than whether every button will disappear. Someone asking about an expense needs an understandable answer. Someone taking an action needs to see its precise contents and outcome. A consistent conversational style does not give an informational response and an actual operation identical consequences or prove that the latter was successfully completed.
Governance has a role of its own
OpenAI describes IARA as a layer for observability, cost, and policies. That is a supplier-and-customer account, not an independent audit. It does not automatically demonstrate accuracy, absence of bias, or compliance with every regulatory regime. Such conclusions require their own verifiable evidence rather than inference from the presence of a governance platform.
As an evaluation criterion, we would seek a clear trail from the question to the data used and the answer returned. Handoffs to a person and treatment of incorrect or ambiguous requests should be examined separately. An engineering team would benefit from measuring the cost of successfully completed service, including retries, rather than only the price of an individual AI response. These are editorial criteria, not tests SIExpo has performed at the bank.
The Brazilian case is interesting because it separates customer context, the conversational product, and shared model governance. The verifiable October development is a more detailed description of that architecture. It is not a new ia.i launch, proof of universal availability, or grounds to promise the same outcome to another bank without its own evaluation and evidence.
Editorial verification: 2026-10-18T16:11:34.881843+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.
