Major financial institution Itaú Unibanco has officially transitioned from exploratory pilot projects to a large-scale enterprise rollout of its generative artificial intelligence infrastructure. Through its mobile banking application, the bank has launched an integrated voice and text interface named ia.i. In the initial phase, 300,000 customers have been granted access to the tool. The institution aims to expand this deployment gradually across its entire user base of roughly 70 million customers by the end of 2026.
Operationally, ia.i handles far more than standard scripted customer support tasks. The interface manages interactive financial advisory dialogues and connects directly to Brazil's real-time payment network, Pix. By integrating directly into payment processing, the system enables users to review account context and initiate live financial transactions conversationally, eliminating the need to navigate through nested banking menus.
Because financial regulators demand rigorous audit trails and strict data governance, ia.i does not communicate directly with third-party model providers. Instead, Itaú Unibanco designed and deployed a dedicated central orchestration and governance platform called IARA. This internal middleware shields core banking infrastructure and ensures all model interactions adhere to regulatory compliance requirements.
Architecturally, IARA operates as an intelligent multi-model router that directs individual prompts to the most appropriate backend engine. The platform integrates OpenAI models for complex reasoning while routing latency-critical workloads, such as real-time Pix payment confirmations, to lightweight open-source large language models. Simultaneously, IARA enforces dynamic token cost controls, data residency constraints and continuous bias monitoring across all incoming and outgoing requests.
The internal impact of the IARA framework is already measurable beyond consumer-facing services. In Itaú's corporate legal department, the platform actively processes over 2 million operational tasks and approximately 90 million documents each month. According to the institution, integrating automated document review and classification through IARA has generated internal efficiency gains of up to 80 percent.
The deployment offers a clear reference blueprint for how heavily regulated financial institutions can safely move generative technologies into production. By inserting an abstraction layer between user interfaces and raw model endpoints, the bank mitigates primary enterprise risks regarding data exfiltration, regulatory non-compliance and unpredictable computing overhead. This architectural approach demonstrates how autonomous and conversational tools can be embedded responsibly into sovereign financial rails.