A data agent needs context, permissions and memory.
Asking questions of data in natural language looks simple. Producing trustworthy answers requires internal definitions, respected access controls and a clear path to every conclusion.
A data agent makes an attractive promise: ask a normal question and receive an answer supported by company information. Between the question and the answer, however, sits a complex layer of definitions, permissions, queries and validation.
DATA DOES NOT SPEAK FOR ITSELF
Common words such as customer, revenue, growth or active user can mean different things across departments. If the agent does not know those definitions, a technically correct query may answer the wrong question.
Business context must be treated as part of the system: glossaries, approved examples, table relationships and visible assumptions.
A FAST ANSWER IS USELESS IF NO ONE CAN EXPLAIN WHERE IT CAME FROM.
PERMISSIONS MUST TRAVEL WITH THE QUESTION
An agent should not turn natural language into a way around controls. It must respect the same permissions the person would have when consulting each source.
Security also means minimizing exposed information, recording queries and preventing a sensitive result from appearing in the wrong context.
TRUST THROUGH TRACEABILITY
A good experience does not show only a number. It lets people review sources, filters, periods and assumptions. When the answer is uncertain, it should say so and request the right clarification.
That design turns the agent into an analytical collaborator: it accelerates exploration without hiding the reasoning or removing professional review.
WHAT A BUSINESS CAN DO NOW
- Create a glossary of approved metrics and definitions.
- Apply existing permissions to every query and result.
- Show sources, filters and assumptions with the answer.
- Evaluate the system with real questions before expanding access.
Engineering article published by OpenAI on January 29, 2026.
OpenAI — In-house data agent ↗