Itaú's conversational banking assistant reached about 300,000 customers in its initial phase, according to the bank. A new OpenAI case study, published October 16, describes the less visible system behind that launch: IARA, a platform through which bank teams access AI models under internal controls. The same account gives workload figures for a legal department tool. These are three different measures of Itaú's AI program, not evidence that its assistant already serves every customer or that the legal results apply across the bank.

The big change

  • What changed: Itaú is pairing a limited customer launch with a shared internal layer for model access. The bank says IARA lets teams use different providers under its policies, so the customer assistant and legal work are parts of one broader program without being the same product.
  • Why it matters: In a bank, expanding a conversational service means controlling how staff applications reach models as well as making the interface available to customers. Itaú reports substantial legal workload, but its published efficiency figures describe stages of that department's processes. They do not establish a bank-wide productivity gain.
  • What to watch: Itaú says it intends to extend ia.i to its full customer base by the end of 2026. The key evidence will be how far access actually expands and whether the bank publishes measured customer and operational outcomes beyond its initial figures.

The customer rollout started in July

Itaú announced ia.i in July as a text and voice interface within its banking app. The bank said customers could ask about their finances, get information about products and services, and use some transaction flows such as Pix. It described a combination of conversation, conventional app navigation and a route to human support when needed.

The bank said the first phase was available to roughly 300,000 customers. OpenAI's October account repeats that figure; it does not provide a newer customer count. Itaú's goal of availability across its customer base by the end of 2026 is still a stated target in that account. Product simulations, sign-ups and recurring analysis of income and spending are described as planned next-stage capabilities.

The case study also describes Customer DNA, a layer of financial context for approximately 70 million customers. It says GPT-5.x models are among the components used to maintain and enrich that representation. The 70 million figure refers to the context layer's stated scope. It does not mean 70 million people currently use ia.i or that every interaction with those customers runs through a GPT model.

IARA controls access to models inside the bank

IARA is Itaú's governance and orchestration platform, according to the OpenAI account. It gives bank teams controlled access to different models in line with internal policies; OpenAI is one of the providers available. The case study describes observability, token cost control and responsible AI policies as aims of the platform.

Those descriptions explain the role Itaú assigns to IARA, but they are not an independent test of its security, accuracy or compliance. The account does not publish the platform's full architecture or the evaluation method behind those controls. It also does not say that all bank applications use the same model or that IARA itself is the customer-facing assistant.

The distinction matters when reading the bank's scale claims. A common access layer can support multiple internal applications while each has its own users, data and measure of work. The legal application is the clearest workload example disclosed in the case study.

The legal figures describe tasks and process stages

Itaú says its legal solution operates on a repository of more than 90 million documents and executes more than 2 million tasks a month. The reported efficiency gains range from 20% to 80% in stages of the legal department's processes. A repository size, a monthly task count and a process-stage efficiency range measure different things. None gives the number of lawyers using the tool, the share of all legal work it handles or a total department time saving.

The figures come from the bank and OpenAI's customer case study. It does not disclose a comparison method, observation period or independent audit for the efficiency range. The figures show a reported operational deployment at substantial workload, with limits on what can be concluded about its effect.

Itaú also describes AI assistance for specialists, managers and service teams. Its figures of three times more time for customer relationships, service up to twice as fast and conversations up to three times more effective are presented as expectations for that implementation, not measured results. The company has not supplied results in the case study that would verify those targets.

The practical reading of Itaú's announcement is narrower than its ambition: a limited customer release, a described model governance layer, and a legal application with company-reported workload and stage-specific efficiency figures. The next customer reach and the effects on banking service remain to be measured publicly.

Sources & further reading