How to Set Up Oracle Analytics Cloud AI Agents

Oracle Analytics Cloud AI Agents are configurable assistants that answer business questions by combining your dataset with uploaded reference documents. To set one up, you need an instance running the January 2026 update or later, the correct platform roles, a connected dataset, and optionally up to 10 knowledge documents in PDF or TXT format. The full configuration takes under 30 minutes once your data source is ready.

What Are Oracle Analytics Cloud AI Agents

Oracle Analytics Cloud AI Agents are purpose-built assistants that combine structured data queries with retrieval-augmented generation (RAG). Unlike the standard AI Assistant, which operates at the platform level across all your content, an AI Agent is scoped to a specific dataset and trained on internal documentation you provide. The result is a conversational interface that can answer questions like "What was EMEA revenue last quarter?" while also referencing your internal KPI definitions or contract terms stored in uploaded PDF files.

Oracle introduced AI Agents in the January 2026 platform update and expanded their capabilities in the March 2026 release. A key addition in March 2026 was the ability to embed AI functions directly into custom calculations, allowing analysts to classify records or filter rows based on semantic meaning rather than exact keyword matches. According to Oracle's release documentation, this compresses the gap between analyst teams and business users who previously had to wait for custom report builds.

Prerequisites and Instance Requirements

Before creating an AI Agent, confirm the following:

• Your Oracle Analytics Cloud instance is on the January 2026 update or later. Earlier instances do not include the AI Agent feature.
• You hold either the BI Service Administrator or BI Author application role.
• At least one dataset, subject area, or local subject area already exists in your instance.
• Key fields in your dataset are at least 95 percent complete with formats standardized, particularly for any ERP-integrated deployments. Data gaps will reduce response accuracy in any deployment, not just ERP-connected ones.
• Optional: up to 10 files in PDF or TXT format containing internal policies, metric definitions, or reference documentation.

Step 1 — Verify Your Roles and Permissions

Log into the Oracle Analytics Cloud console as an administrator. Navigate to the Security section and confirm that your user account holds the BI Service Administrator or BI Author application role.

If the AI Agent option does not appear in the Create menu after confirming roles, check that your tenant instance version is current. Contact your Oracle tenancy admin to trigger the platform update if needed. The January 2026 update is a prerequisite that cannot be skipped.

Step 2 — Create the AI Agent Object

From the Oracle Analytics Cloud Home page:

1. Click Create in the top navigation bar.
2. Select AI Agent from the dropdown menu.
3. The AI Agent editor opens to a blank configuration panel.

If AI Agent is absent from the Create menu, your instance is not on the January 2026 release or later. You cannot manually unlock this feature without an instance update.

Step 3 — Connect a Dataset

In the AI Agent editor, click Add Data Source. You can connect to:

• Dataset: A standalone dataset already uploaded or connected to OAC.
• Subject Area: A semantic model defined in your RPD or semantic modeler.
• Local Subject Area: A dataset-specific semantic model.

Select your data source and confirm the connection. The agent's natural language engine uses the column names and relationships in this source to interpret user questions. If your column names are technical abbreviations, add synonyms in Step 6 to improve accuracy.

Step 4 — Write Supplemental Instructions

Supplemental instructions define the agent's behavior and support up to 6,000 characters of plain text. Cover four areas:

Role definition. What the agent is for. Example: "You are a sales analytics assistant for the EMEA region. Answer questions only about sales performance data."

Scope limits. What the agent should refuse. Example: "Do not answer questions about HR data, legal matters, or topics outside the connected dataset."

Fallback behavior. What to say when data is unavailable. Example: "If you cannot find the answer in the connected data, say 'I don't have enough information to answer that question' and do not guess."

Metric definitions. If your organization uses non-standard definitions, for example ARR calculated excluding trial accounts, document those definitions here so the agent applies them consistently across all users.

Well-written supplemental instructions are the single largest factor in output quality. Vague instructions produce vague, inconsistent answers. Invest time in this step before worrying about knowledge documents or synonyms.

Step 5 — Upload Knowledge Documents

Click Upload Knowledge Document to add PDF or TXT files. These files are processed using RAG: the agent retrieves relevant sections before generating a response. This is separate from the dataset, which the agent queries directly.

Good candidates for knowledge documents include:

• Internal glossaries of business terms
• Finance or HR policy documents the agent should reference
• Metric definition guides or calculation methodologies
• Compliance frameworks or contract term summaries

You can upload up to 10 files per agent. Each file should be well-structured and machine-readable. Scanned PDFs with poor OCR quality will reduce retrieval accuracy. If you are working with scanned documents, run them through an OCR tool first and export to clean text before uploading.

Step 6 — Define Synonyms

The Synonyms panel lets you map natural language terms to column names or values in your dataset. If your dataset uses a column named "acct_rev_usd," you can create a synonym so that a user asking "what is account revenue" receives the correct result without needing to know the technical column name.

To add a synonym:

1. Click Add Synonym in the Synonyms panel.
2. Enter the business term in the left field.
3. Map it to the technical column name or value on the right.

Adding 10 to 15 synonyms for your most-queried metrics significantly reduces errors from ambiguous terminology. This step is optional but recommended for any deployment with non-technical end users.

Step 7 — Add a Welcome Message and Save

The Welcome Message is the first text end users see when they open the agent in a workbook. It should introduce the agent's scope, provide two or three example questions the agent handles well, and set expectations on what the agent cannot answer. Keep it under 300 characters so it displays cleanly in the chat panel.

Once the welcome message is set, click Save. The agent appears in your OAC content library as an AI Agent object. Creating it here does not yet make it visible to end users.

Step 8 — Enable the Agent in a Workbook

You must attach the agent to a workbook and enable the Workbook Assistant before end users can interact with it:

1. Open the target workbook in OAC.
2. Switch to Present Mode configuration (not Edit Mode).
3. In the Workbook Assistant settings, toggle Enable Workbook Assistant to on.
4. Select your newly created AI Agent as the data source.
5. Save the workbook.

End users can access the AI Agent only in Present Mode. In Edit Mode, workbook authors see the standard AI Assistant instead. This separation means business users interact with the scoped, governed agent while authors retain access to broader AI features. If you need the agent accessible across multiple workbooks, repeat this step for each one.

Data Quality Is the Underestimated Factor

The March 2026 Oracle release documentation states that AI-ERP integrations require critical training fields to be more than 95 percent complete, with formats standardized across modules, before models produce trustworthy outputs. This threshold applies most strictly to Oracle Fusion ERP integrations where the agent cross-references transactional records with analytics data.

For standalone OAC deployments, the threshold is not technically enforced, but data gaps still degrade response quality. Before deploying an AI Agent to business users, audit your dataset for null values in key metric columns and standardize date formats across joined tables. Fixing data quality issues upstream is faster and more reliable than trying to work around them in supplemental instructions.

If building an AI Agent feels like too much setup for an immediate need, platforms like VSLZ AI let users upload a file and ask questions in plain English without any schema configuration or agent setup.

Summary

Oracle Analytics Cloud AI Agents are available to any tenant on the January 2026 update or later. The setup covers eight steps: verify roles, create the agent object, connect a dataset, write supplemental instructions, upload knowledge documents, define synonyms, set a welcome message, and enable the agent inside a workbook. Data quality, particularly field completeness and format standardization, is the most underestimated factor in deployment success. Start with a single well-governed dataset rather than connecting multiple sources at once, and iterate on supplemental instructions after gathering initial user feedback before expanding access more broadly.