From Las Vegas to São Paulo: what Oracle brought to Brazilian companies

The Oracle AI World conference ran from October 25 to 28 in Las Vegas. On November 4, exactly one week after it ended, the Brazilian agenda included an AI World Las Vegas Download session with Adrián Durán and Marcelo Christianini. The São Paulo edition paired that update with conversations among customers, partners and technical teams. According to Paulo Ribeiro, Oracle’s marketing director for Latin America, the two-day program was designed for more than 2,000 attendees, with around 80 sessions.

Oracle chose Beyond The Club in São Paulo. The club began operating in 2025 and announced the opening of Terraço The View, its new event space, in October 2026. Oracle AI Experience Brazil was the terrace’s first corporate event. Its wave pool near Marginal Pinheiros sets the scene: a surf club in the middle of the city alongside labs, panels and customers discussing how to put AI to work.

A local take on the global announcements

In Las Vegas, the conversation was about where the platform was heading. In São Paulo, the program put that direction to the test against operations with rules of their own. November 4 covered finance and healthcare, with labs on agents connected to Oracle AI Database, NVIDIA infrastructure on Oracle Cloud and engineering with Codex. Governance, security and integration ran through those tracks.

An agent in a financial operation, for instance, needs to know what data it can access, what it can do on its own and when it must hand a decision back to a person. If the team cannot reconstruct how a decision was reached, it is hard to fix errors or understand why an exception was treated as routine. This is the work that surfaces when a demonstration has to fit into a real operation.

Data: a query can run and still answer the wrong question

The São Paulo lab on agents with Oracle AI Database 26ai is one example of how the local agenda connects to Las Vegas. Version 26.10 of Private Agent Factory, released on October 26, added Deep NL2SQL, which helps turn natural-language questions into SQL queries using business definitions. The tool also lets users inspect the generated query.

Imagine a retail manager asking which stores lost margin. The answer changes depending on whether returns count on the sale date or the refund date, and whether shipping and discounts are assigned to the store or the online channel. The database says where the numbers are. Someone from operations needs to define how they are counted. Before rolling the agent out, test questions with known answers, including cases where a different definition changes the result.

Oracle also announced a semantic layer for Autonomous AI Lakehouse to share metric definitions and context across data sources, including Apache Iceberg tables. The timing matters: Oracle said this capability would become available within 12 months of the announcement. It describes the direction of the platform, and customers should verify availability in their own environments.

Applications: authorization matters when agents take action

At the other end, the Oracle Fusion announcements show a push to put agents inside processes such as finance, HR and supply chain. Fusion Claw is the runtime used in these applications. Fusion Companion and Agentic Home were introduced as ways to describe an objective, delegate work and track progress. This is the connection to the Brazilian enterprise applications track: the software starts coordinating steps in a process as well as answering questions.

One way to evaluate the proposal is to follow an exception. In accounts payable, an agent can gather information about a disputed invoice and route it to the person responsible. Releasing the payment is a separate action, subject to different approval limits and controls. A pilot needs to show which step was authorized, how the system records what happened and what the team does if partial execution leaves the process unfinished.

This helps explain Oracle’s strategy: it is connecting where data lives, the tools that interpret it and the systems that record the work. Customers could benefit from reusing existing context and permissions. The question is how much of that context stays accessible when a workflow crosses into tools from other vendors, and how much work it takes to maintain those integrations.

Startups gain speed. Scale depends on access to operations

Marcelle Paiva, vice president of Oracle’s AI Hub for Latin America, pointed to an advantage for AI-native startups: they can design a process from the start without reorganizing decades of systems and routines. She also highlighted the challenge of winning customers and gaining access to the data and problems that make a solution relevant.

A partnership makes a difference when it provides access to a task worth solving and someone who can follow the test. A startup can move quickly; a large enterprise knows the exceptions and can show what happens when volumes rise. A pilot becomes more useful when both sides agree on a criterion: what result needs to improve against the current process to justify deployment?

In robotics, the environment is part of the product

The discussion also reached physical AI. Victor Rysovas, a director at AutorIA, described proofs of concept with the Cruiser S2 humanoid. The focus is on identifying which tasks make sense and what needs to be adapted and trained for each environment.

An object-handling demonstration shows what the hardware can do. Deciding whether to put it in a factory or warehouse raises other questions: does the task repeat often enough, can the robot handle variations at the site, and can the team operate and maintain the system? The result depends as much on deployment as on the technology itself.

What to watch on the second day

For November 5, the Oracle agenda lists a keynote with OpenAI, a data sovereignty discussion with Dataprev, Serpro and Caixa, and tracks on the public sector, retail, telecommunications and enterprise applications.

A useful case shows the process before and after AI. In a triage pilot, for example, reducing the time to the first handoff tells us little if the human review queue grows at the same rate. A convincing case shows time to resolution, the exceptions that needed intervention and the cost of keeping the workflow running. That is the evidence that helps distinguish a demonstration from a change in daily operations.