Oracle AI Footnotes
What Oracle's AI features do with your data, as Oracle publishes it. Checked October 31, 2026.
Agent access
Fusion agents can be called from other applications through the /invokeAsync REST API, and published agents are exposed through an A2A API with agent card discovery, registry search and invocation endpoints. Oracle's Service Descriptions allow outside AI agents and MCP implementations only through its authorized APIs. source, source 2, source 3, source 4, source 5 · checked October 31, 2026
Generative AI features in Fusion Applications
Oracle Fusion Cloud Applications (Oracle AI for Fusion Applications)
| On by default? | Not published as of October 31, 2026 |
|---|---|
| How an admin turns it off | Setup and Maintenance work area: enable or disable each generative AI feature. Predictive AI features also need configuration steps, because their models train on the customer's data. source · checked October 31, 2026 |
| Trains on customer data? | No: prompt data does not train the LLMs, and one organization's data is not used to train models or generate responses for others. Predictive models are trained on each customer's own data and kept isolated. source, source 2 · checked October 31, 2026 |
| Retention | Generative AI data is deleted after the period set in the Retention period for Generative AI Data profile option (ORA_FAI_DATA_RETENTION_DAYS): 90 to 365 days, 90 by default. No customer data is retained in the LLM after a prompt is processed. source, source 2 · checked October 31, 2026 |
| Where data is processed | Each customer gets its own partition of the Oracle AI Platform in its local OCI region. Input to and output from the generative models is processed in an OCI environment that might be outside the local region. source · checked October 31, 2026 |
| Model providers | Oracle's Service Descriptions list two Included models for generative AI functionality that is not an agent: Cohere Command-R-08-2024 and Meta Llama 3.3. They run on the Oracle AI Platform for Fusion Applications, which is built on OCI Generative AI. source, source 2, source 3, source 4 · checked October 31, 2026 |
| Price | Included: use of an Included model does not draw down AI Units. General actions such as Q&A, summarization and approvals on Oracle-provided, OCI-hosted models do not draw down AI Units either. source, source 2 · checked October 31, 2026 |
AI agents and Fusion Agentic Applications (Oracle-built)
Oracle Fusion Cloud Applications (HCM, ERP, SCM, CX)
| On by default? | Not published as of October 31, 2026 |
|---|---|
| How an admin turns it off | Security Console: users get an agent through a custom role with the Access Intelligent Agent Chat (HRC_ACCESS_AI_AGENT_CHAT_PRIV) privilege and the Fai Genai Agent Runtime Duty (ORA_DR_FAI_GENERATIVE_AI_AGENT_RUNTIME_DUTY) duty role. The role is then added on the Security tab of each workflow and in the App Settings of each agentic application in AI Agent Studio. source · checked October 31, 2026 |
| Trains on customer data? | No: Oracle states the foundation models behind its agentic features are not fine-tuned with customer data, and no prompt or response data is retained for model training. source, source 2 · checked October 31, 2026 |
| Retention | Generative AI data is deleted after the period set in the Retention period for Generative AI Data profile option: 90 to 365 days, 90 by default. source · checked October 31, 2026 |
| Where data is processed | The OpenAI endpoints used by Fusion agents are in Europe (EEA and Switzerland) for data centers in Europe, the UK, the Middle East, India and Singapore, and in the USA for data centers in the Americas, Australia and Japan. source · checked October 31, 2026 |
| Model providers | Oracle-provided models for agents: a Basic OpenAI GPT-OSS model hosted on OCI; Balanced models from OpenAI (GPT-5.6 Luna, the default in AI Agent Studio) and Google (Gemini 3.1 Flash Lite); and Frontier models OpenAI GPT-5.5 and Google Gemini 3.5 Flash. Admins can switch an agent's model in AI Agent Studio. source, source 2, source 3, source 4, source 5, source 6, source 7, source 8 · checked October 31, 2026 |
| Price | Oracle-built agents are included in the subscription and draw down AI Units by action type and model: general actions on the Basic OCI-hosted model use none, and Balanced and Frontier models use AI Units. Applicable Fusion services include 30,000 AI Units per production environment a month; more are sold at $1,000 per 100,000 pooled AI Units (list price, October 2026). AI Units start to decrement in the month a customer receives Release 26C. source, source 2, source 3, source 4, source 5, source 6 · checked October 31, 2026 |
AI Agent Studio (custom agents and workflows)
Oracle Fusion Cloud Applications
| On by default? | Not published as of October 31, 2026 |
|---|---|
| How an admin turns it off | Security Console: access to AI Agent Studio comes from custom roles; removing those roles from a user's account removes access. source · checked October 31, 2026 |
| Trains on customer data? | No for Oracle-provided models: no prompt or response data is retained for model training. With a bring-your-own LLM, the customer is responsible for the model's governance and compliance. source, source 2 · checked October 31, 2026 |
| Retention | Not published as of October 31, 2026 |
| Where data is processed | Not published as of October 31, 2026 |
| Model providers | Oracle-provided models, or the customer's own LLM credentials from OpenAI (direct API or through Microsoft Azure), Google Gemini (direct API), Gemini Enterprise Agent Platform (through Google Vertex AI), Anthropic (through Google Vertex AI) or Oracle Cloud Infrastructure. source, source 2, source 3, source 4 · checked October 31, 2026 |
| Price | AI Agent Studio is available at no additional cost (Oracle, March 2026). Agents built or modified in it draw down AI Units in development, testing and production. With a bring-your-own LLM, the customer pays the model provider directly. source, source 2, source 3 · checked October 31, 2026 |
Connections to outside MCP servers
Agents and workflow nodes built in AI Agent Studio can call outside MCP servers through an MCP tool, added under Resources > Tools with the server URL, the transport (SSE or Streamable HTTP) and a credential type. An agent that uses outside MCP servers counts as a Custom AI Agent. No release stage is published. source, source 2 · checked October 31, 2026
Compliance
EU AI Act
Oracle's AI questionnaire for its SaaS applications (AI-CAIQ, October 2026) says its SaaS compliance audit plans verify compliance with requirements including applicable AI regulations, such as the EU AI Act. source · checked October 31, 2026
Not published by Oracle as of October 31, 2026
- Whether generative AI features and Oracle-built agents are on or off when a customer receives them
- An official MCP server for Fusion Cloud Applications (Oracle publishes one for Fusion Data Intelligence, a separate analytics product)
- Release stage of the /invokeAsync and A2A APIs and of the AI Agent Studio MCP tool
- Processing region for models other than OpenAI's
- Which model powers each embedded generative AI feature
- Whether Fusion AI features are covered by Oracle's HIPAA or FedRAMP offerings
- A dedicated EU AI Act statement for Fusion Applications (Oracle is not on the European Commission's GPAI Code of Practice or AI Pact signatory lists on the check date)
Changes
No changes recorded since the first check on October 31, 2026.
How this page is made, and how to send a correction: method.