OpenAI’s new Data agent shortens the path from a business question to a decision. In ChatGPT Work, a user can investigate company data, build a dashboard, and move toward an approved action in the same conversation.
The launch also makes the data-readiness question more concrete. A connected source can establish where the agent is allowed to look. A semantic layer can help it use the organization’s definitions. Before the result supports a real decision, the team still needs to know whether the selected data, definitions, conditions, and review criteria fit that specific use.
OpenAI’s own guidance points in this direction. It recommends connecting a data warehouse and semantic layer, then tells users to check the source, time period, filters, and metric definition before relying on a result. As analysis gets faster, the enterprise context behind each result matters more. Leaving that context implicit makes the decision harder to defend.
What OpenAI’s Data Agent changes
OpenAI introduced the Data agent in ChatGPT Work on September 10, 2026. According to the announcement, it can connect to approved sources including Amazon Redshift, Google BigQuery, Databricks, MongoDB, and Snowflake. It can also bring files and documents from Google Drive and SharePoint into an analysis.
OpenAI says the Data agent can use business terms, metric definitions, custom calculations, and data relationships supplied through semantic layers and other trusted sources. Administrators control which connections and roles are available, while queries continue to enforce the connected account’s table, row, and column permissions.
This combination can shorten the path from a business question to a decision. A metric that once passed through an analyst’s queue can now reach a decision-maker directly. That speed makes the quality and applicability of its context easier to overlook.
The enterprise assets behind a useful answer
A semantic layer can provide authoritative metric definitions and relationships. The knowledge behind a business decision may still be spread across more places.
It can appear in a finance workbook that adjusts revenue for a contract type, a support document that defines an escalation exception, a customer-specific field mapping, an approval matrix, or the judgment of the person who knows why the official process is sometimes overridden. These are enterprise assets because they shape how work is actually done.
These assets are candidates, not automatic ground truth. A formula may encode a current rule, a one-off exception, or a workaround for a problem that no longer exists. A document may be authoritative for one region and obsolete in another. Human judgment may be essential only under conditions that were never written down.
This is why the upstream work cannot be reduced to data cleaning or ETL. The team has to identify the data, relationships, definitions, decision logic, exceptions, approvals, and expert knowledge that matter to one AI task. It then has to decide who owns each element, when it applies, and how its validity will be checked.
Our earlier article on the enterprise assets behind an AI pilot examines this problem in more detail. The Data agent launch makes it immediate for everyday analytics: the answer can only be as decision-ready as the operating knowledge used to interpret it.
Access and semantic context do not settle Target Fit
Four evidence layers help separate what a data-agent workflow has established from what remains open.
| Layer | What it establishes | What still needs evidence |
|---|---|---|
| Access | Which sources and records the connected account can reach. | Whether those sources are the right ones for the business question. |
| Business context | Which definitions, calculations, relationships, and policies should guide interpretation. | Whether they are current, authoritative, and applicable to this case. |
| Qualification | Whether a defined data state supports one target under recorded conditions and evaluation criteria. | Whether the result remains valid outside the tested scope. |
| Assurance | What data state and conditions were used in actual operation, and what changed afterward. | Whether a material change requires review or requalification. |
These layers are related, but none should be used as a substitute for another. A permission check does not validate a metric definition. A metric definition does not prove that the underlying population fits the decision. A successful analysis does not show that the same result will hold after the data, model, policy, or operating environment changes.
This is the practical difference between data access and usable data for AI. Access answers whether the system can retrieve the information. Usability depends on the target, the condition of the data, the meaning applied to it, and the evidence required for the decision.
What to verify before relying on a data-agent result
OpenAI’s Data plugin guidance advises users to inspect the source, time period, filters, and metric definition before relying on a result. For a consequential use case, the team can extend that review into a compact decision record.
Define the decision. State the business question, the person or system that will use the answer, and the action that may follow. “Why did retention fall?” and “Which customers should receive an intervention?” are different targets, even when they begin with the same data.
Name and review the authoritative inputs. Record the source tables, documents, semantic definitions, calculations, and relevant operating rules. Check customer-specific mappings, manual adjustments, approval thresholds, and regional exceptions. If two definitions conflict, identify who decides which one applies. Do not automate an observed pattern merely because it appears repeatedly.
Lock the comparison conditions. Preserve the data period, population, filters, model or agent version, prompt, connected tools, and evaluation method that produced the result. Without these conditions, a later comparison may look precise while measuring something different.
Set an acceptance criterion. Decide what evidence is sufficient for the intended action and who can approve it. A dashboard used for exploration may have a different threshold from an analysis that changes a price, allocates a budget, or triggers a customer action.
Record what changes the decision. New data, a revised definition, a policy change, a different model, or a changed permission can make earlier evidence stale. Assign an owner to decide when the result needs another review.
This record does not have to slow every analysis. Its depth should match the consequence of the decision. The purpose is to make reliance deliberate rather than accidental.
Match the evidence to the decision
The evidence required depends on where the team is starting. CUBIG’s current operating model separates three decision areas.
When a team needs to understand its data before choosing a specific target, Baseline examines the common state across usability, integrity, context, consistency, reproducibility, and traceability. Its score is a diagnostic indicator, not a verdict that the data is ready for every use.
When the target is known, Qualification asks whether a defined data state fits that specific use. It records the task, model or agent, evaluation conditions, success criterion, and approval boundary. Comparative evidence then supports a qualified, not qualified, or inconclusive judgment for that scope.
After deployment or real use, Assurance asks which data state and conditions were used, what changed, and whether the earlier qualification still applies. Version history alone cannot answer those questions.
These are distinct operating domains, not maturity levels. A team may begin with a common-state diagnosis, a target-specific question, or a need to reconstruct evidence from an existing workflow.
Weak evidence across these domains can eventually produce a stalled pilot or a Zombie PoC, but that is only one downstream symptom. The larger opportunity is upstream: prepare the enterprise assets behind an AI task before an impressive answer becomes an unsupported action.
The wider market is moving toward data agents
Other same-week launches reinforce the signal. Google Cloud introduced a Data Agent Kit for agentic analytics. Prophecy added AI data-preparation capabilities with a step-by-step review flow. Salesforce described an enterprise AI harness built around context, agency, action, governance, security, and models, while noting that new capabilities and a unified experience are planned to begin rolling out in early fiscal FY28.
These announcements suggest a direction rather than measured market adoption. AI is becoming an interface for querying, preparing, interpreting, and acting on business data. That makes the enterprise assets behind the interface more important.
Connection count will not settle whether a result deserves trust. Teams need to show which definitions and operating knowledge shaped it, whether the data state fit the target, who reviewed the evidence, and when a change made the earlier decision stale.
AI-ready data depends on the decision
A faster path from question to action raises a practical requirement: teams must be able to trace each result to the data state and operating context that produced it. CUBIG’s view is that enterprise data becomes AI-ready only in relation to a specific use. The relevant sources, business definitions, exceptions, target conditions, and review evidence must stay aligned as that use evolves.
CUBIG calls this operating discipline the AI-Ready Data Operating Layer, with Syntitan as the AI-Ready Data Platform. The work begins with a defined decision, not an attempt to automate every rule the organization can uncover. Teams first determine which data and operating knowledge apply, who owns them, and how their validity will be tested.
Start with one consequential decision. Identify the systems and working assets behind it. Resolve the definitions and exceptions that could change the answer, assign a review owner, and set an acceptance criterion. The result should be a bounded judgment about whether the current data state provides enough evidence for a person or agent to act, and what change would require another review.
Explore Syntitan to see CUBIG’s approach to AI-ready data for enterprise AI work.
