Banco do Brasil Phases In Autonomy for AI Agents, Saying Governance Must Be Built In
Banco do Brasil is expanding its use of agentic artificial intelligence by granting AI agents autonomy in stages, keeping people in charge of supervision and building controls into each solution from the start, TI Inside reported on October 19. Giuliane Paulista, the bank’s AI and data executive, told the outlet’s Innovation Forum 2026 that the first autonomous agent the bank disclosed to the market, in the first half of 2026, reached roughly 90 percent efficiency.
With more than 80 million customers, she said, any rollout at Banco do Brasil quickly reaches large scale, which is why phased autonomy became one of its rules for building and using agents. The bank has extended its ethical and responsible AI guidelines, first written for generative AI, to cover agentic systems, and Paulista said it was the first bank in the world to publish such a set, on its investor-relations site. Decision-making power is meant to grow only as test results, interaction safety and architecture reliability allow, including in multi-agent setups where some agents orchestrate and others check the system.
Paulista argued that governance can no longer be a compliance gate at the end of development. Security, monitoring, metrics and protective mechanisms have to sit inside the architecture, and developers, product owners, data owners and business teams all share responsibility. Accountability, she stressed, cannot be handed to software, remarking that AI has no CPF, the Brazilian taxpayer number.
What it means in practice
The approach contrasts with competitive showcases such as Itaú and Google’s Batalha de Agentes and speaks to the gap described in the MIT SMR–TOTVS study showing only 7 percent of Brazilian firms put AI at their core. For a state-controlled lender, moving slowly but visibly may be the point.
Paulista also flagged cost. A customer-service agent that resolves more cases can drive up token consumption, and she said some U.S. financial institutions are weighing hybrid setups and their own infrastructure to make AI spending predictable. Data quality, semantic context and consistent versions of the same information across hybrid and multicloud environments are, in her view, as important as the models themselves.
What to watch: whether Banco do Brasil publishes measurable results for its next agents, how regulators such as the Central Bank respond to phased-autonomy frameworks, and whether other Brazilian banks adopt similar public guidelines.