BusinessMarkets, money & strategy
Itaú launches ia.i banking assistant and says its IARA multi-model platform cut generative-AI costs by 65%
Itaú's in-app assistant ia.i reached 300,000 customers in July 2026, and the bank aims to reach every customer by year-end. It relies on IARA, an internal platform that chooses among more than 80 open and closed models to manage cost and latency.
From the OpenAI News article of October 16, 2026 and cited sources: IARA is an orchestration layer that picks a model for each application. [2] [4] [9] [10]
SOURCES[2] Mobile Time (alt. URL) mobiletime.com.br[4] Mobile Time: Itaú custo IA generativa mobiletime.com.br[9] Finsiders Brasil: IARA finsidersbrasil.com.br
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- IARA is an orchestration layer that picks a model for each application. [2] [4] [9] [10]
- Original source
- OpenAI News article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 16, 2026 article published
- Covered
- Oct 17 edition
- Evidence
- 11 sources across 9 sites
- 1OpenAI: How Itaú is building generative AI for its entire…openai.com
- 2Mobile Time (alt. URL)mobiletime.com.br
- 3Convergência Digital: Itaú incorpora assistente IA ao appconvergenciadigital.com.br
- 4Mobile Time: Itaú custo IA generativamobiletime.com.br
- 5Finsiders Brasil: Pix no WhatsApp com IAfinsidersbrasil.com.br
- 6Canaltech: Nova IA do Itaúcanaltech.com.br
Reach: not measured yet
IARA is an orchestration layer that picks a model for each application. It weighs cost, latency, regulatory limits and how sensitive the data is. It also uses a prompt cache to reuse answers to similar requests. For Pix payments on WhatsApp, small open-source models reportedly responded faster than large proprietary ones. [2] [4] [9] [10]
Itaú itself reports the 65% cost reduction, and its scope is unclear. One outlet dates it from the start of the bank's generative-AI work in 2022. Another says it covers unit token cost over the past year. [2] [9]
a large regulated bank runs customer-facing generative AI through central routing across many models.
Read the full assessment
Implications: choosing the cheapest model that is good enough may be the main way to control cost, and payments made by chat raise new authorisation and liability questions.
Executive brief
Itaú, Brazil's largest private bank, says it has cut its generative-AI costs by 65%. It credits a single internal platform, IARA, which picks among more than 80 off-the-shelf models for each task, including small open-weight ones (Mobile Time). That cost control underpins ia.i, an assistant inside the banking app that launched to 300,000 customers on 27 July 2026, with every customer targeted by year-end (Convergência Digital). OpenAI's 16 October 2026 case study presents this as Itaú's "AI First" strategy in a heavily regulated sector (OpenAI). Every performance number comes from Itaú itself.
What changed and event timeline
Generative AI programme begins
Itaú dates its generative-AI work, and the starting point for its 65% cost reduction, to 2022 ().
Pix on WhatsApp in pilot
A generative-AI Pix (Brazil's instant-payment system) service on WhatsApp was being tested with 200,000 customers, with about 2 million planned. A separate LLM investment-advice tool reached 10,000 customers ().
Ia.i launches
A conversational assistant in the app went to 300,000 customers. The September target is reported as 3 million by but 1 million by.
Cost and governance figures disclosed
CTO Carlos Eduardo Mazzei described IARA, more than 80 models in use, and the 65% cost cut ().
OpenAI expands in Brazil
OpenAI formally announced a São Paulo office and cited Itaú as an example of corporate ChatGPT adoption ().
OpenAI publishes case study
The case study frames ia.i and IARA as AI First for both customers and employees ().
Capabilities and access
- Where it lives: ia.i is a floating button on the app's home screen.
- What it does: checks balances, breaks down spending by category, suggests savings, explains products and makes Pix transfers (Convergência Digital).
- Personalisation: it opens conversations based on what it knows about each customer (Canaltech).
Read the full section
- Where it lives: ia.i is a floating button on the app's home screen. Customers can type, speak or send images, and answers come back as text, charts or interactive cards (Canaltech).
- What it does: checks balances, breaks down spending by category, suggests savings, explains products and makes Pix transfers (Convergência Digital).
- Personalisation: it opens conversations based on what it knows about each customer (Canaltech).
- Which model: ia.i's underlying model is not disclosed. Itaú says it uses both open and closed models (IT Forum). Models in its portfolio are reported as GPT 5.6, Sonnet 5 and Qwen 32B (Mobile Time).
Technical analysis for researchers and developers
- Architecture: IARA is an orchestration layer between commercial frontier models, Itaú's own engines and internal components.
- Caching: a prompt cache reuses answers already generated for similar content, so they are not reprocessed (Finsiders Brasil).
- Routing example: for Pix on WhatsApp, small open-source models reportedly beat large proprietary ones on latency (Fast Company).
Read the full section
- Architecture: IARA is an orchestration layer between commercial frontier models, Itaú's own engines and internal components. For each application it weighs cost, latency, regulatory constraints and data sensitivity to choose a model (Mobile Time summary via search).
- Caching: a prompt cache reuses answers already generated for similar content, so they are not reprocessed (Finsiders Brasil).
- Routing example: for Pix on WhatsApp, small open-source models reportedly beat large proprietary ones on latency (Fast Company).
- Data foundation: 1,500 data curators, 95% of data with a named owner, and continuous quality monitoring on 70% of internal tables (Mobile Time).
- Not published: benchmarks, routing logic and safety evaluations.
Claims and evidence
All of these figures are Itaú's own; no independent audit or evaluation of them was found.
Read the full section
- 65% cost reduction: one outlet dates it to 2022 onward (Mobile Time); another says it applies to unit token cost over the past year (Finsiders Brasil). The baseline is unclear.
- Credit results: 30% better credit-risk accuracy and R$1 billion in net financial margin from AI-supported lending in the first year (Mobile Time).
- Legal automation: 43 million documents and 300,000 calls processed (Mobile Time).
- Rollout targets: conflicting September figures (see timeline).
All of these figures are Itaú's own; no independent audit or evaluation of them was found.
Context and prior work
- Earlier pilots: Pix on WhatsApp and the LLM investment-advice tool came before ia.i (Finsiders Brasil).
- Competitor: Bradesco says its BIA assistant resolves 85% of requests after a generative-AI upgrade (Finsiders Brasil).
- OpenAI in Brazil: Brazilians send ChatGPT 215 million messages a day, and OpenAI reports five times more corporate users than a year earlier (Forbes Brasil).
Read the full section
- Earlier pilots: Pix on WhatsApp and the LLM investment-advice tool came before ia.i (Finsiders Brasil).
- Competitor: Bradesco says its BIA assistant resolves 85% of requests after a generative-AI upgrade (Finsiders Brasil).
- OpenAI in Brazil: Brazilians send ChatGPT 215 million messages a day, and OpenAI reports five times more corporate users than a year earlier (Forbes Brasil).
- Name clash: Brazil's tax-appeals council CARF has an unrelated AI tool also called IARA (TAGD Law).
Limitations, safety and contested findings
- Safeguards stated in general terms: Itaú cites ethical standards, data security and responsible use of data.
- No public detail on ia.i's model, error rates or how it handles hallucinations (Canaltech).
- Contested figures: the September rollout target (1 million vs 3 million) and the scope of the 65% cost cut.
Read the full section
- Safeguards stated in general terms: Itaú cites ethical standards, data security and responsible use of data. Mazzei says ia.i supports and guides customers without replacing their own decisions (Convergência Digital).
- No public detail on ia.i's model, error rates or how it handles hallucinations (Canaltech).
- Contested figures: the September rollout target (1 million vs 3 million) and the scope of the 65% cost cut.
- OpenAI's role is thin: OpenAI's case study is the main link to the company. Independent outlets name no exclusive vendor.
Business and practitioner implications
- Model routing saves money: sending each task to the cheapest model that is good enough, often a small open-weight one for low-latency work, is the main cost lever Itaú describes (Mobile Time).
- Governance comes first: central model governance and data ownership came before a full customer rollout, a sequence regulated firms can copy.
- Agentic payments: Pix transfers move from menus into conversation, raising new questions about authorisation and liability.
Read the full section
- Model routing saves money: sending each task to the cheapest model that is good enough, often a small open-weight one for low-latency work, is the main cost lever Itaú describes (Mobile Time).
- Governance comes first: central model governance and data ownership came before a full customer rollout, a sequence regulated firms can copy.
- Agentic payments: Pix transfers move from menus into conversation, raising new questions about authorisation and liability.
- No single vendor: OpenAI's showcase sits alongside a multi-model setup, so suppliers should expect to compete for routed traffic.