How to Use ChatGPT Data Agent: Setup, Connectors & Semantic Layers

01

Overview

OpenAI's Data agent is useful only when it has the right approved sources, business context and permissions. This guide focuses on the practical setup job: making the plugin available, connecting data systems, giving it reliable metric context and checking its answers before using them for decisions.

02

Before you start

OpenAI currently documents Data for ChatGPT Work and Codex. Availability can depend on the workspace and administrator configuration. The Data plugin itself must be available, and the underlying data-source plugins or apps you want it to use must also be available and connected.

OpenAI recommends having at least a data warehouse source, a semantic/business-context source and the appropriate sharing or publishing tools for the workflow you want. Not every workspace will expose the same connectors, and a connector being listed in documentation does not mean every user automatically has access to it.

03

Step 1: Ask the workspace admin to enable the required plugins

For managed workspaces, the first dependency is administrative. OpenAI's Help Center directs admins to Workspace settings and the Plugins controls to make approved plugins available.

Before enabling a broad set of sources, decide which systems should be exposed to the agent, who is allowed to connect them and which roles should be able to create or publish reports. The Data agent is not a reason to weaken an existing data-governance policy.

04

Step 2: Connect the data source

OpenAI currently lists Amazon Redshift, ClickHouse, Databricks, Google BigQuery, MongoDB and Snowflake among supported structured sources. Google Drive and SharePoint can provide documents and files when those integrations are approved and connected.

Complete the source-specific connection flow using the user's real account and permissions. OpenAI says source restrictions continue to apply, including table-, row- and column-level access controls. If an analyst has access to only a subset of a warehouse, Data should be expected to work within that subset rather than across the entire system.

05

Step 3: Add semantic and business context

Connecting a warehouse is not enough for many business questions. A database may contain a column called revenue, but a company can still have several accepted definitions of revenue depending on refunds, tax, subscription timing or booked versus recognized value.

OpenAI therefore emphasizes semantic layers and trusted business context. Its documentation mentions metric definitions, custom calculations, data relationships and systems such as Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon and existing BI dashboards as possible context sources.

The practical goal is to make the authoritative definition easier for the agent to find than an outdated or informal one. If two sources disagree, resolve the conflict before relying on a generated analysis.

06

Step 4: Install and invoke Data

Once the plugin and required sources are available, install Data through the plugin directory or workspace-supported installation flow. OpenAI's current Help Center documents starting a conversation with the Data plugin and invoking it with @Data where that interaction is supported.

A good first task is narrow and verifiable. Instead of asking “why is growth bad?”, try a request that defines the metric, date range, comparison period and relevant segment.

For example: “Using our approved conversion definition, compare paid-social conversion rate this week with the previous four-week average, break the change down by campaign and landing page, and cite which source and filters you used.”

This gives you several things to check: whether the correct source was selected, whether the date range was interpreted correctly, whether the expected metric definition was applied and whether the output can be reconciled with an existing report.

07

Step 5: Validate the analysis

Natural-language analytics should still be auditable. OpenAI explicitly advises users to verify source, time period, filters and metric context when reviewing an analysis.

A useful validation checklist is:

1. Confirm the source system and table or dashboard used.
2. Confirm the metric definition.
3. Confirm the date range and timezone.
4. Confirm filters and excluded segments.
5. Compare at least one key number with an existing trusted report or direct query.
6. Ask the agent to explain any assumptions or joins that materially affect the result.

If the answer changes dramatically after a small clarification, investigate why before acting on it.

08

Step 6: Move from analysis to a dashboard only when the question is stable

Data can turn findings into interactive reports and dashboards, but it is usually better to validate the underlying question first. A polished dashboard can make a wrong definition look authoritative.

Once the numbers and definitions are stable, a reusable dashboard can be valuable for recurring questions such as weekly acquisition performance, customer retention, sales-pipeline health or operational KPIs. OpenAI also documents working with BI tools and OpenAI Sites for supported publishing workflows.

09

Supported source does not mean unrestricted access

OpenAI says existing permissions apply. This is especially important with shared warehouses and BI systems. A Data-agent conversation inherits the practical boundaries of the connected user's access rather than creating a new superuser layer.

If the agent cannot answer a question, check whether the relevant source is connected, whether the user can access the required tables or records, whether the semantic definition exists and whether the workspace allows that plugin.

10

Common setup problems

Data plugin is not visible

The workspace may not have the plugin enabled or the user's account may not be eligible for it. Check workspace plugin settings and current product availability.

The correct warehouse is connected but the answer is incomplete

The user may lack access to a table, row set or column required for the analysis. Check permissions in the source system rather than assuming the agent failed to search.

The numbers disagree with the BI dashboard

Compare metric definitions, timezones, filters, joins and the exact source used. A semantic layer or trusted business definition may be missing or ambiguous.

The agent picked the wrong meaning for a company term

Add or improve the business-context source instead of repeatedly correcting the wording in individual chats. The strongest setup makes authoritative definitions reusable.

A connector exists in OpenAI's documentation but is not available in the workspace

Connector and plugin availability can vary by workspace, role and rollout state. Treat the live workspace as the final source for access.

11

Security and governance checklist

Use the minimum source access required for the job. Keep authoritative metric definitions in controlled systems. Review who can publish dashboards and Sites. Separate analysis from consequential actions when a downstream system would send messages or change records. Re-check permissions when a user's role changes.

OpenAI's permission inheritance is an important control, but teams remain responsible for how they configure source access and where generated reports are shared.

12

Bottom line

The best Data-agent setup is not the one with the most connectors. It is the one with the right approved sources, clear metric definitions, least-necessary permissions and an easy path to validate important results. Configure those foundations first; then use Data to reduce the manual work of moving among warehouses, documents and dashboards.