How Itaú Deploys Multi-Model Generative AI to Serve 70 Million Customers
Brazilian banking giant Itaú is scaling its generative AI architecture across 70 million customers using model orchestration, token management, and internal governance platforms.
Itaú, one of the largest financial institutions in Brazil, is expanding its generative AI initiatives across its entire operations to serve its massive user base. According to research conducted by Itaú in partnership with Grupo Consumoteca, 80% of Brazilians want clearer support to organize their financial lives, and 65% desire automated technologies that function as financial copilots. To address this demand, the bank is focusing on making complex backend models unnoticeable to the end user by delivering direct, practical tools. Carlos Eduardo Mazzei, Chief Technology Officer at Itaú, emphasizes that the bank's strategy relies on solving real everyday problems rather than highlighting the underlying technology itself.
Conversational banking through ia.i and Customer DNA
To bring natural language interactions to daily financial tasks, Itaú introduced ia.i, a unified conversational system activated by text or voice. The platform consolidates services ranging from checking account inquiries, credit cards, and insurance to investment advice and WhatsApp transactions, while seamlessly routing complex requests to human specialists. During its initial deployment phase, ia.i reached approximately 300,000 customers, with plans to expand access across the bank's full customer base by the end of 2026. Upcoming capabilities will incorporate product simulations, automated sign-ups, and ongoing analysis of income and spending habits.
Central to this personalized experience is Customer DNA, an internal financial context layer designed for roughly 70 million customers. Powered by OpenAI's GPT models under strict governance controls, Customer DNA tracks user habits, needs, and life stages. By combining multimodal capabilities—text, speech, images, and interface clicks—the bank ensures that AI responses remain grounded in individualized context.
Model orchestration with the IARA governance platform
Operating generative AI at scale within a heavily regulated banking environment required Itaú to create an internal governance and orchestration platform named IARA. IARA provides bank development teams with controlled access to multiple AI models, ensuring strict compliance with responsible AI policies, token-cost controls, response accuracy checks, and risk mitigation against model bias.
The architecture allows Itaú to route tasks according to specific needs. Simpler interactions use lightweight, low-latency models to minimize costs, whereas complex financial inquiries leverage frontier reasoning models. Beyond customer-facing tools, IARA powers enterprise tasks across the organization. For example, Itaú's legal department relies on the platform to search, analyze, and interpret a repository containing more than 90 million documents. Executing over 2 million automated tasks monthly, the legal team has achieved efficiency gains between 20% and 80% across various workflow stages.
Internally, the AI-first approach assists managers and customer specialists with meeting preparations, document summaries, and tailored action plans. Itaú projects these internal tools will double service speed, make customer interactions three times more effective, and free up three times more time for relationship-building.
What it means for developers
Itaú's enterprise deployment offers several practical lessons for software engineers and architects building large-scale AI applications. First, relying on a single AI model is rarely optimal for complex enterprise products. High-volume, low-latency workflows require lightweight models to remain economically viable, while nuanced multi-step logic requires high-reasoning frontier models.
Second, centralized governance and token management are essential before scaling. As demonstrated by Itaú's IARA platform, monitoring token usage, controlling API costs, and maintaining context privacy are core engineering requirements rather than afterthoughts. For developers seeking to integrate multiple foundation models into their applications without managing complex individual vendor relationships, unified access platforms like Apixo (https://apixoai.online) provide pay-per-token API access to leading models like Claude, GPT, Gemini, Grok, and DeepSeek through a single API key.
Finally, Itaú’s approach highlights that user experience takes precedence over model sophistication. Success in production relies on solving concrete problems—such as statement clarification or routine transactions—while abstracting away model mechanics from the end user.
Source: How Itaú is building generative AI for its entire customer base — OpenAI News. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
