Search 800 + Posts

Jan 29, 2026

From Natural Language to Insights: 10 Tests for Oracle Select AI (With 23ai and 26ai)

This blog is in continuation to our previous blog setting the Stage: A 5-Step Guide to Configuring OracleSelect AI

None of test explained in this blog possible without a secure, robust configuration. Select AI is much more these test , here we just capture some BASIC functionality of Select AI. In the world of enterprise data, there has always been a "Translation Tax." Business leaders ask questions in English; analysts translate them into SQL; and three days later, a report arrives.

With the release of Oracle Database 23ai and the recent enhancements in 26ai, that tax is being abolished. Select AI allows users to converse with their private database schemas using natural language. But does it actually work on complex, multi-table enterprise data?

At Bizinsight Consulting, we put Select AI to the test using a standard 7-table Order-to-Cash (O2C) schema. Here is how we stress-tested the engine—and the results that prove the future of data is conversational.

The Setup: An "AI-Ready" Schema

To make the test realistic, we built a relational foundation covering the entire lifecycle of a transaction:

  • CUSTOMERS & ITEMS: The master data.
  • SALES_ORDERS (Headers/Lines): The demand.
  • SALES_ORDER_SHIPMENT: The logistics.
  • INVOICES (Headers/Lines/Receipts): The finance.

10 Tests for Select AI

We categorized our tests by complexity, moving from simple lookups to "Agentic" business reasoning.

Phase 1: Linguistic Mapping

Can the AI understand business synonyms?

Test 1 (The Geography Test): "List all customers who are based in Columbus."

Result: Passed. The AI correctly mapped "based in" to the CITY column.


Test 2 (The Identity Test): "Find the email address for the customer named John Smith."

Result: Passed. It performed a precise filter on CUSTOMER_NAME and selected only the EMAIL column.

john.smith@example.com

Phase 2: Relational Navigation

Can the AI handle the "Join Tax" automatically?

Test 3 (The Sales Link): "Show me all sales orders placed by Emily Johnson in January 2026."

Result: Passed. The AI joined CUSTOMERS to SALES_ORDER_HEADER using the CUSTOMER_ID foreign key without any manual prompting.



Test 4 (The Item Detail): "What items and quantities were included in order ORD-43?

Result: Passed. It navigated from the Header to the Lines and pulled the human-readable ITEM_NAME from the ITEMS table.



Phase 3: Quantitative Intelligence

Can the AI perform reliable math?

Test 5 (The Revenue Summary): "What is the total revenue generated by each customer this year?

Result: Passed. The AI identified TOTAL_AMOUNT in the INVOICE_HEADER table as the correct source for revenue and performed a SUM with a GROUP BY.



Test 6 (The Pricing Audit): "Show me the average unit price of items sold to Emily Johnson."

Result: Passed. A complex three-table join resulted in a perfect AVG() calculation.


Test 6.1 (The Pricing Audit): select AI narrate Show me the average unit price of items sold to Emily Johnson;

Result: The average unit price of items sold to Emily Johnson is $495.

Phase 4: Operational Friction

Can the AI spot discrepancies in the supply chain?

Test 7 (The Shipment Gap): "List all order lines that have been shipped but the quantity shipped is less than the quantity ordered."

Result: Impressive. The AI compared SALES_ORDER_LINES.QUANTITY against SALES_ORDER_SHIPMENT.QUANTITY_SHIPPED, identifying "Partial Shipments" instantly.


Test 8 (The Logistics Lag): "Which orders are marked as 'PENDING' even though we have recorded shipments for them?"

Result: Passed. This proved the AI could be used for data quality auditing.

Phase 5: The CFO Challenge

Can the AI reason like an executive?

Test 9 (The Aging Report): "Find all invoices that are past their due date and haven't had a receipt recorded yet."

Result: Passed. By checking for the absence of records in INVOICE_RECEIPTS, the AI created a sophisticated NOT EXISTS or LEFT JOIN query.



Test 10 (The Customer Health Check): "Which customers have more than 2 unshipped orders and at least one overdue invoice?"

Result: The customer "Nordic Design Corp" has more than 2 confirmed but unshipped orders and at least one overdue invoice..

The AI successfully correlated shipping delays with financial risk—a query that would normally take an analyst 20 minutes to write.

Lessons Learn

Through these 10 tests, we identified three critical factors that make Select AI production-ready:

1.      The Power of Annotations (26ai): In Oracle 26ai, we used Data Annotations to tell the AI that "Revenue" specifically means "Paid Invoices." This eliminated the ambiguity that often plagues LLMs.

2.      Metadata is the New Code: Your SQL is only as good as your Comments. We found that adding descriptive comments to every column increased the AI’s accuracy from 70% to nearly 100%.

3.      Security by Inheritance: Because Select AI runs inside the database, it automatically respected our Virtual Private Database (VPD) policies. A regional manager asking Test #5 only saw revenue for their specific region.

Conclusion: The Era of the "Conversational ERP"

Select AI isn't just a "feature"; it’s a fundamental shift in how businesses interact with their data. By removing the technical barrier to entry, we empower everyone—from the warehouse floor to the C-suite—to make data-driven decisions in real-time.

At Bizinsight Consulting, we specialize in making your Oracle schemas "AI-Ready." The future is talking to your data. Are you ready to listen?