OpenAI published a customer story on Itaú, the Brazilian bank, and how it is building generative AI for customers and staff. The customer-facing product is ia.i, a conversational layer you can use by text or voice. It pulls checking accounts, credit, cards, insurance and other services into one place, and hands over to human specialists when needed. The bank says it reached roughly 300,000 customers in a first phase and aims to cover its whole customer base by the end of 2026.
Underneath is IARA, a governance and orchestration platform. It gives Itaú's teams controlled access to several models, OpenAI's among them, in line with internal policies. Itaú says it handles cost per token, answer accuracy and responsible-AI rules together. Its legal department's tool runs on it. The tool reads and analyses documents across a repository of more than 90 million, runs over 2 million tasks a month, and shows efficiency gains of 20% to 80% in some process stages.
The bank also uses the same layer internally, for example to prepare relationship managers for meetings. It expects three times more time for customer relationships and service up to twice as fast. Those are expectations, not measured results. The source says GPT-5.x models help maintain Customer DNA, a financial-context layer covering about 70 million customers.
My take
The part I'd pay attention to is IARA, not the chatbot. Itaú's own advice is to build the governance layer before you scale. That means one place for model access, observability, and token-cost tracking from day one. If you're building anything with LLMs for real users, that's a practical template. Don't let every team call a model API directly and then wonder why the bill and the quality are all over the place.
I also like that they use a range of models. Cheap, fast ones handle simple requests, and heavier reasoning models handle harder ones. Putting everything through the biggest model is the mistake that hurts your latency and your budget. Another tip I liked: not every action should go through natural language, so mix conversation with normal navigation and buttons.
I'd stay a bit sceptical, though. This is a vendor's customer story, so it's the good-news version. The 20% to 80% gain is a wide range with no baseline. The internal targets are promises. We also don't know how ia.i performs once 70 million people use it, or how often it gets money questions wrong. A first phase of 300,000 is a small slice, and the full rollout is where problems will show up.
For builders in Africa, the useful lesson is that a regulated market like Brazil's pushed the bank toward careful controls first. Fintech teams here face similar pressure. I'd copy the order of operations: governance, then a small launch group, then expansion.
Sources
This post was written with AI assistance from the sources above.
Comments
What do you think? Drop your thoughts below.
Loading comments…