Itaú Puts a Conversational Layer in Front of 70 Million Customers — Starting With 300,000

Itaú, Brazil's largest private-sector bank, says it has begun rolling out a conversational assistant called ia.i to customers, reaching roughly 300,000 people in a first phase with a stated goal of covering its entire customer base by the end of 2026. In an OpenAI customer story published October 16, 2026, Itaú CTO Carlos Eduardo Mazzei describes a two-part architecture: a customer-facing assistant and an internal governance and orchestration platform called IARA that brokers access to multiple models. The internal platform, rather than the chatbot, is the part other regulated enterprises should study closely.

What was actually shipped

According to Itaú, ia.i is a conversational layer activated by text or voice that pulls checking accounts, credit, cards, insurance and other services into a single environment, with handoff to human specialists when needed. It builds on capabilities the bank introduced in 2024 under the Inteligência Itaú banner, which targeted specific financial tasks rather than offering a general interface.

The bank describes three design principles: interactivity, multimodality across text, voice, image and clicks, and individualization. That last point leans on something Itaú calls Customer DNA, a context layer the bank says represents the needs, habits and life stages of approximately 70 million customers. Itaú says GPT models are among the components that help maintain and enrich that representation, under governance and responsible AI controls. The company does not break down which parts of Customer DNA are model-driven versus conventional data infrastructure, and the distinction matters — a profile layer for 70 million people is a data engineering problem long before it is a model problem.

The scale gap is worth stating plainly. Roughly 300,000 customers in phase one is under half a percent of the 70 million figure the bank cites for Customer DNA coverage. Itaú frames this as deliberate: a learning period to observe real usage before expanding. That is sound practice, but it also means the end-of-2026 target implies a very steep ramp, and the source material does not describe what gates the bank has set between phases or what would cause it to slow down.

The governance layer is the interesting part

IARA is described as a governance and orchestration platform that gives internal teams controlled access to different models in line with bank policy, with OpenAI among the providers available. Itaú says it solves three problems simultaneously: per-token cost control, response accuracy, and enforcement of responsible AI policy.

That framing reflects a pattern now common at large enterprises. Rather than wiring applications directly to a single vendor API, the bank inserts a brokerage layer that handles routing, observability, policy enforcement and spend tracking. Mazzei's stated requirement is that information stay within the bank's context and that models be accessed in a secure, governed way with an understanding of token-use efficiency so the economics work.

Mazzei also makes an argument against single-model standardization. Itaú says using a range of models made it possible to deliver the full experience in one product — cheap, low-latency models for simple interactions, and stronger reasoning models where grounding is required. The source says GPT-5.x models are part of Customer DNA. No routing logic, model-selection criteria, or cost figures were disclosed, so there is no way to evaluate whether the multi-model approach is delivering meaningful savings versus adding operational complexity.

The legal department numbers, read carefully

The most concrete production figures in the announcement come from Itaú's legal department, where the platform automates document reading, analysis and interpretation. Itaú says the solution operates on a repository of more than 90 million documents, executes more than 2 million tasks per month, and has produced efficiency gains of 20% to 80% in stages of the department's processes.

Two million tasks per month is a substantial production workload, and it is a better signal of real deployment than most enterprise AI case studies offer. The efficiency range, however, is wide enough to be hard to interpret. A 20%-to-80% band across unspecified process stages tells you some steps improved a lot and others improved modestly, but not which steps, what share of total departmental work they represent, or how the baseline was measured. Itaú also does not disclose accuracy rates, error rates, or how often a human reviews or overrides model output in a legal context — the metric that matters most when documents carry regulatory weight.

Internal rollout: expectations, not results

Itaú says the same intelligence layer supports specialists, managers and service teams, generating pre-meeting summaries, meeting preparation and plan development. The bank attaches specific targets: three times more time available for customer relationships, service up to twice as fast, and conversations up to three times more effective.

These are framed in the source as expectations, not measured outcomes. "Conversations up to three times more effective" in particular has no stated definition. Editors and buyers should treat these as goals the bank has set, not results it has reported.

Regulation, bias and the Brazilian context

Itaú operates in one of Brazil's most regulated sectors, and the bank positions responsible AI as a product requirement rather than a compliance afterthought. Mazzei says responses must align with the institution's values and culture, and explicitly cites reducing bias risk that may arise in frontier models. The bank says responsible AI principles guide the full lifecycle from development through continuous monitoring, aiming at safety, transparency, ethics and regulatory compliance.

What is not described is the mechanism — no bias testing methodology, audit cadence, red-team process, or regulator engagement is detailed. The announcement also does not name which Brazilian regulations or supervisory expectations shaped the architecture.

On demand, Itaú cites a study it conducted with Grupo Consumoteca finding that 80% of Brazilians want to better organize their financial lives and 65% would like technologies acting as a financial copilot. This is bank-commissioned research supporting a bank product, and sample size and methodology were not disclosed. Stated interest in a copilot and sustained use of one are different things, which is arguably why the 300,000-customer learning phase exists.

Why this matters beyond Brazil

The transferable lesson is sequencing. Itaú built the governance and cost-control layer before scaling the customer-facing product, treated per-token economics as a design constraint from the start, and kept navigation alongside conversation on the view that not every action belongs in natural language. For product teams at regulated institutions, that ordering is more useful than any individual metric in the announcement.

The open question is whether a conversational front end materially changes customer behavior at scale, or whether it becomes another channel layered on an app most people already know how to use. Phase one is too small, and too new, to answer that.

What to watch

• Whether Itaú hits the end-of-2026 full-rollout target, and whether it publishes usage and retention data rather than reach numbers.
• Any disclosure of accuracy, error and human-override rates for the legal document solution, which is the deployment with the clearest regulatory exposure.
• Whether the internal productivity targets — 3x relationship time, 2x faster service, 3x more effective conversations — are later reported as measured results with stated definitions.
• Promised next-stage capabilities, including product simulations, sign-ups and recurring income and expense analysis, which move the assistant from answering questions to initiating transactions.
• Whether Brazilian financial regulators issue guidance that constrains customer-facing generative AI, and how IARA's observability tooling holds up under that scrutiny.