OpenAI Internal Data Agent
Suggested workflow: An internal data agent that combines data table usage patterns with organizational knowledge to support employee data queries. Task: Suggested implementation: Build an internal data agent that integrates six layers of context, including table usage metadata and institutional knowledge, to answer employee queries in natural language. The system retrieves relevant tables, generates SQL, and validates queries using the user's specific permissions. Human involvement includes approving heavy analysis plans and correcting the agent's memory to improve future accuracy.; tools: Tool described in the source. Human role: Not stated in the source; limitations: Prerequisites include a comprehensive knowledge layer with table descriptions, metric definitions, and example queries. The agent operates with read-only permissions by default, requiring human approval for heavy analysis or write actions. Source reliability is secondary, as the details are summarized from a public article; actual implementation costs and autonomy levels are not specified..
Tools
Tool described in the source
Published on Agent Work Atlas:
Overview
Suggested workflow: An internal data agent that combines data table usage patterns with organizational knowledge to support employee data queries.
Workflow
Suggested implementation: Build an internal data agent that integrates six layers of context, including table usage metadata and institutional knowledge, to answer employee queries in natural language. The system retrieves relevant tables, generates SQL, and validates queries using the user's specific permissions. Human involvement includes approving heavy analysis plans and correcting the agent's memory to improve future accuracy.
Reported result
The source reports that OpenAI's in-house data agent serves 3,500+ internal users over 600 PB of data. Specific performance metrics such as query accuracy rates or time savings are not provided in the text. To verify success, propose a measurable acceptance check: compare the percentage of correct SQL answers against manual lookup time for a sample of 50 common business questions.
Human role
Not stated in the source
Limitations
Prerequisites include a comprehensive knowledge layer with table descriptions, metric definitions, and example queries. The agent operates with read-only permissions by default, requiring human approval for heavy analysis or write actions. Source reliability is secondary, as the details are summarized from a public article; actual implementation costs and autonomy levels are not specified.
Try it
Act as a data engineering lead. I need to pilot an internal data agent for our analytics team. Inputs: - [replaceable: sample of 10 common business questions] - [replaceable: schema and metadata for 3 non-sensitive internal tables] - [replaceable: target output location, e.g., Slack channel or internal wiki] Steps: 1. Construct a knowledge layer using the provided table descriptions and metric definitions. 2. Configure the agent to retrieve relevant tables and generate SQL queries based on the sample questions. 3. Ensure queries run under simulated user permissions (read-only). 4. For any query estimated to take >5 minutes, generate a plan and pause for human confirmation. 5. Output the generated SQL and a natural language answer for each question. Deliverables: A report listing each question, the generated SQL, the answer, and a flag for 'requires human approval'. Quality Check: Verify that all SQL syntax is valid and that no data outside the provided schemas is accessed. Error Handling: If the agent cannot find a relevant table, return 'No data source found' instead of guessing. Safeguards: Do not execute any write operations. Require explicit human confirmation before sending any results to the [replaceable: target output location]. Ask clarifying questions only if the table schemas are missing.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.