During AI World (formerly CloudWorld), Oracle can’t stop talking about AI. The striking thing is that the most impressive additions to the Oracle portfolio use as little AI as possible. That could well prove to be a formula for success.
A cynic would say that what Oracle is announcing at AI World in Las Vegas is exactly what the competition has already presented. Here, too, agentic AI plays a leading role and the importance of business data being available for AI usage is reiterated. Even the cliché that “data is the new oil” gets several mentions. However, this obscures the fact that Oracle has a fairly unique story to tell, with its traditionel strengths in data management coming to the fore in a new context. These strengths start with the renamed Oracle AI Database 26ai (yes, two AIs in one title), but extend to every cloud, the integration of best practices, and a smart interpretation of familiar third-party technologies.
We hear from co-CEO Mike Sicilia that Oracle is the only hyperscaler that is truly verticalized. According to the company, it’s the only one able to provide the data, infrastructure, applications, and enough trust (read: security) to build AI. We start with the changes within AI Database 26ai and elaborate on them to clarify Oracle’s vision. But the bottom line is this: AI usage must fit within known patterns and be only a means to achieve the answers that have always been available in Oracle databases. They’ve only become more accessible than ever.
We’ve already highlighted how Oracle AI Database 26ai is an evolution of Database 23ai. According to Oracle, that release from early 2024 was the foundation on which 26ai builds. In fact, the new version is a fully compatible replacement for 23ai that can be installed with the October patch. In addition, new features such as the integration of the Model Context Protocol even work on the older 19c.
26ai introduces a new data type: AI Vectors. For generative and agentic AI to function, data must be housed in a vector database. Oracle makes it as easy as possible for users to make this conversion from objects. Whether they are JSONs, PDFs, XML files, etc., Oracle displays their properties in vector data. Similar files are given a similar dataset without users having to specify this. Anyone searching for specific information via a query can then see which data matches best.
Finding relevant business data is easier than before thanks to the addition of AI Vectors. Oracle gives the example of a business leader wondering which manager is performing best. A semantic AI system translates this freely interpretable question into an SQL query that can actually be answered by the database. Someone without technical knowledge can therefore ask subjective questions that can be answered with objective data. For example, the query may request a table of managers who generate the most revenue per customer. Nevertheless, the AI translation is meaningful, because not every business manager knows exactly what the table looks like and what SQL code is needed to answer the question.
The end result of the SQL query then appears in understandable language and/or in the form of a table. Here, too, AI plays a significant role, but the underlying activity is classic database work.
Autonomous AI Lakehouse
All this facilitates the step from raw data to insights. Nevertheless, it is also important to break down data silos. The same business manager without database expertise does not know if and why some data is only available on-premises, within OCI or another public cloud. That is why the new Oracle Autonomous AI Lakehouse must bring this data together. Oracle embraces the open table format Apache Iceberg, used by Snowflake and Databricks, among others. The company promises that it will maintain high performance, something that, according to Oracle, sometimes lags behind Iceberg. Thanks to an Exadata cache, latency remains acceptable.
Data stores outside Oracle are also accessible for AI queries. Retrieval-Augmented Generation picks up data from all versions of Oracle Database (in any cloud, on-premises) and beyond. A federated catalog of catalogs keeps track of which business data is available and makes searching this data plug-and-play through ready-made integrations.
Eliminating risks
Oracle naturally pays close attention to all the risks associated with AI use. Consider the danger to data privacy when you unleash an AI model on your business data. Or what if your AI outputs are hallucinatory and therefore share unusable information (which you can’t even verify)? Oracle circumvents all this by minimizing the use of AI. A question in plain language translates directly into a query, while the answer is only formulated by an LLM in the very last step. LLMs can still be useful here, because the strongest AI models can reason quite well on this precise data. Oracle supports everything from GPT-5 to Claude Sonnet 4.5, Gemini 2.5 Pro, and Grok, which it claims is the widest choice among all hyperscalers.
How much does the choice of the exact LLM matter? For some privacy-sensitive on-premises customers, there is simply no option to choose the best AI models via API calls. Juan Loaiza, EVP Oracle Database Technologies, tells us that there are still many options for Oracle users, wherever they are located. More importantly, and clearly Oracle’s focus, is that the LLMs behave as securely as possible and actually have as little agency as possible.
This actually goes against the grain. With agentic AI, the promise is that AI systems will do everything for you. But the LLMs are only the first point of contact and the ultimate wordsmith. The real data, the real insights, must come from the answer to the SQL query.
Conclusion: AI within limits
Again, it is fairly easy to say that Oracle is very limited and cautious in its use of AI. However, it cannot behave any differently without sending the wrong message about AI. That message should be that LLMs are still in their infancy, as are the applications surrounding them. Loaiza does not mince words; during his presentation, he devotes a lot of attention to the dangers of AI and how Oracle circumvents them. It does this by largely restricting AI as we know it in 2025 and redesigning its own systems where necessary.
Read also: Oracle AI Database 26ai: how a database becomes “AI-native”