Blog
Data Agents Need More Than Data Access
Data agents are leaving the dashboard and becoming an interface for analysis and decisions. OpenAI introduced the Data agent in ChatGPT Work, while Microsoft describes GPT-6 Astra in Copilot working with enterprise context and existing permissions. For companies, model quality is therefore only one part of the equation. The order behind the data matters just as much.
What OpenAI is changing
The Data agent connects to approved sources such as Snowflake, Databricks, BigQuery, Redshift, and SharePoint. It investigates changes, builds interactive dashboards, and can continue the work in existing BI tools such as Power BI, Tableau, or ThoughtSpot. Administrators choose which connections and roles are available. Queries are designed to respect existing table-, row-, and column-level permissions.
OpenAI also points to semantic layers, business definitions, custom calculations, and data relationships as context. That is the critical detail: a convincing report is not automatically a reliable report. It depends on whether terms such as “revenue,” “active customer,” or “margin” are defined clearly and versioned inside the organization.
The enterprise test comes before the prompt
Microsoft highlights the same control point for GPT-6 Astra in Copilot: Work IQ grounds responses in files, meetings, chats, and business data within existing permissions. Delegating larger tasks increases the value of good context — and the cost of a poorly modelled access boundary.
For DACH companies, a first Data agent pilot should therefore not start with a spectacular question. Check three things first:
- Are metrics, data owners, and refresh times documented?
- Are role, table, row, and column permissions actually enforced?
- Does every generated analysis have a traceable source and a human approval point?
The consequence for CIOs and CFOs
The economic measure is not the number of dashboards created. More useful metrics are cost per reviewed decision, avoided rework, and the share of analyses that can be used without a correction cycle. In finance and regulated environments, teams must also define which data connections are allowed and how results are recorded.
Start with a narrowly scoped process, a maintained metric definition, and a named owner. Only when access, meaning, and approval fit together does natural language become a dependable enterprise process.