OpenAI Says Now Everyone Can Put Data to Work, but Governance Is the Real Test

OpenAI launched its Data agent on September 10, arguing that Now everyone can put data to work without writing database queries or learning another analytics tool. The agent connects ChatGPT Work to company data, investigates business questions, and turns its findings into interactive dashboards. The promise sounds simple. The difficult part is ensuring that every answer uses the right definitions, permissions, and evidence.

That makes this more than a conversational interface for charts. OpenAI is trying to place ChatGPT between employees and the systems that define how their organizations measure performance. The Data agent can reach warehouses, documents, semantic layers, and business intelligence tools from one conversation. It can also refine an analysis after follow-up questions.

Microsoft and Salesforce already bring natural-language analysis into Power BI and Tableau. OpenAI is applying pressure from another direction. Instead of making AI a feature inside one analytics product, it wants ChatGPT Work to coordinate analysis across the tools a company already uses. Whether that approach succeeds will depend less on fluent answers than on governance beneath them.

Now Everyone Can Put Data to Work Through One Conversation

The immediate change is that ChatGPT Work can now coordinate data investigation, dashboard creation, and follow-up analysis inside a single conversational workflow.

OpenAI describes the Data agent as a plugin that works with approved company systems. Supported sources named in the Data agent launch include Amazon Redshift, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake. It can also incorporate files and documents stored in Google Drive and SharePoint.

A user can begin with a business question instead of a query. Examples include asking why weekly active users changed, where spending increased, or which accounts face renewal risk. The agent can investigate the underlying information, identify possible drivers, and show the evidence supporting its findings.

The conversation does not have to end with a written answer. Users can ask the agent to create an interactive dashboard with charts, filters, and organization-specific branding. Teams can then edit, share, and refresh that dashboard as their underlying data changes.

This workflow matters because business questions rarely arrive as complete analytical specifications. A manager might first ask why sales declined, then narrow the question by region, customer segment, or product. Traditional reporting often turns each refinement into another request for an analyst. The Data agent keeps those refinements in the same conversation.

OpenAI also says the agent can build and interact with dashboards in established platforms, including Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. That positioning suggests cooperation rather than immediate replacement. Existing visualization and governance systems remain part of the workflow, while ChatGPT becomes the conversational control layer above them.

The agent relies on more than raw tables. It can use business definitions, custom calculations, metric relationships, and trusted data sources to interpret a request. OpenAI points to semantic layers from systems such as Databricks Genie Ontology, dbt, and Snowflake Horizon.

A semantic layer is a governed set of business definitions that explains how data should be interpreted. It can establish, for example, which transactions count as revenue or which activity qualifies a customer as active. Without that context, an agent might produce a technically valid calculation that answers the wrong business question.

Administrators control whether Data is available and which roles can install or use it. OpenAI’s setup guidance recommends connecting at least a warehouse and a semantic layer. It also advises users to check sources, time periods, filters, and metric definitions before relying on a result.

The installation flow reflects those dependencies. An administrator makes Data available through Workspace settings and configures the relevant source plugins. A user can then install the plugin, connect authorized accounts, and address a request to @Data.

Calling the interface conversational does not remove the underlying infrastructure. A useful answer still depends on well-maintained data, documented metrics, and correctly configured connections. The agent changes who can initiate an investigation and how quickly they can iterate. It does not eliminate the work required to make company data trustworthy.

The Data Agent Pressures the Analytics Request Queue

OpenAI’s strongest challenge is not to analysts themselves, but to the queue separating business questions from analytical answers.

Many organizations divide analytics work into two stages. Business teams identify a question, while data specialists translate it into queries, calculations, visualizations, and explanations. That division protects quality, but it can also create delays when the number of questions exceeds the available analytical capacity.

The OpenAI Data agent targets that bottleneck. It gives sales, finance, operations, and product teams a way to begin routine investigations independently. Analysts can spend less time rebuilding common reports and more time validating definitions, improving data models, and handling questions that require deeper judgment.

OpenAI says internal adoption already spans much of its own organization. According to the company, nearly all members of its product team and more than two-thirds of its go-to-market organization use related data-agent capabilities in ChatGPT Work. Those figures are company-reported adoption signals, not independent measures of analytical accuracy or business impact.

The customer examples reveal the work OpenAI expects the product to absorb. NTT DATA says non-engineers in sales and corporate roles have built and updated dashboards with plain language. Thermo Fisher Scientific reports using the agent to prepare teams and identify opportunities involving its supply base.

ServiceTitan offers a more specific example. The company says it used the agent to build a dashboard comparing customers who used Atlas, its AI assistant, with those who did not. The analysis found that Atlas users launched campaigns at roughly three times the rate of nonusers. ServiceTitan says the finding is influencing its onboarding decisions.

That example shows both the attraction and the limit of self-service analysis. A threefold association can point toward a useful product strategy. It does not, by itself, prove that Atlas caused customers to launch more campaigns. More active customers might simply be more likely to adopt the assistant.

An analyst would normally test that distinction by examining customer size, tenure, prior activity, and other possible explanations. A conversational agent can help perform those checks, but a person must still recognize why they matter. Faster access to analysis does not automatically create statistical discipline.

Other alpha users describe similar productivity gains. CookUnity says its growth team used the agent to refine a seasonal conversion dashboard and checked the result against internal reports. Micro1 says its operations team rebuilt performance dashboards in about half an hour while identifying errors in the originals.

These accounts are informative because they describe concrete workflows. They remain customer testimonials selected by OpenAI, however. They do not establish a general error rate, a standardized time saving, or a measurable return across different data environments.

The pressure on analytics teams will therefore be uneven. Organizations with mature metric definitions and clean permission structures can delegate more exploratory work. Companies with fragmented systems may discover that the agent exposes unresolved disagreements about whose numbers are authoritative.

That outcome would still be valuable. When two teams calculate retention differently, a polished dashboard can hide the dispute. A conversational workflow that shows its sources and definitions can bring the disagreement into view. The real productivity gain may come from resolving that ambiguity once, rather than answering one question faster.

For knowledge workers, the shift resembles the broader move toward an AI knowledge base. Finding information is only the first step. Useful systems must connect facts with context, provenance, and the decisions people need to make.

How the OpenAI Data Agent Works Across Existing BI Tools

OpenAI is competing through orchestration: one agent can cross warehouses, documents, business definitions, and visualization platforms without requiring every task to begin inside one BI product.

That is the central mechanism behind the launch. A warehouse stores structured company data. A semantic layer explains business meaning. Documents add qualitative context. A BI platform presents approved reports and visualizations. ChatGPT Work provides the interface that coordinates those components.

Consider a sales leader investigating a drop in renewals. The warehouse might contain contract dates, product usage, support activity, and account attributes. SharePoint could hold account plans, while a BI dashboard shows the company’s approved retention metric. The Data agent can bring these sources into one investigation, subject to the available connectors and permissions.

The leader can ask for the affected segments, compare them with prior periods, and request evidence for the strongest drivers. A follow-up could add support-ticket themes from documents or ask for a dashboard organized by region. The value comes from maintaining the analytical thread as the question evolves.

That differs from a basic text-to-query feature. Translating a sentence into SQL solves only one part of the workflow. The user must still choose the correct source, interpret the result, compare alternatives, and communicate the finding. OpenAI is packaging those steps as an agentic process, meaning the software can plan and execute several connected tasks toward a stated goal.

The launch does not make existing BI vendors irrelevant. Microsoft’s Copilot data questions already query Power BI semantic models and return answers as visualizations. Microsoft advises model authors to use clear field names, sound model structures, and business-specific synonyms to improve results.

Tableau follows a similar path within its own environment. Tableau Agent can create visualizations, calculate fields, filter data, and support natural-language exploration. Its documentation also describes boundaries, including source constraints and tasks that users should divide into separate steps.

These products hold an important advantage. They operate inside established analytical environments where organizations may already manage certified sources, dashboards, permissions, and reporting workflows. Users can move between AI assistance and direct visual editing without leaving the platform.

OpenAI’s advantage is breadth. The Data agent can start in ChatGPT Work and reach several data systems, documents, and BI products. That can help when a question crosses departmental boundaries or when relevant context lives outside a formal dashboard.

The strategic contest is therefore not simply Data agent vs Power BI, or Data agent vs Tableau. It is a contest over the starting point for analytical work. Microsoft wants many questions to begin with a Power BI semantic model. Salesforce wants users to explore through Tableau’s governed environment. OpenAI wants the business question to begin in ChatGPT Work and route outward.

That distinction has consequences for vendors and customers. If ChatGPT becomes the common interface, data platforms risk becoming less visible infrastructure behind the conversation. If BI products retain the trusted presentation and review layer, OpenAI remains a coordinator rather than the final system of record.

OpenAI is acknowledging that reality by supporting existing tools. The agent can interact with Power BI and Tableau rather than forcing customers to recreate every dashboard. This reduces adoption friction and lets organizations preserve investments in their current data stack.

It also creates technical complexity. Each connection can expose different capabilities, metadata, and permission models. A dashboard action supported in one platform may not be available in another. OpenAI’s guidance notes that connected-tool actions depend on both the tool’s capabilities and the user’s access.

The same issue applies to unstructured context. A document may explain why a metric changed, but it might be outdated or speculative. The agent must distinguish authoritative definitions from supporting commentary. Administrators must decide which sources deserve trust and how conflicting evidence should be handled.

The best implementation will not connect every available system immediately. It will begin with a bounded workflow, a known source of truth, and an explicit review process. That lets teams measure whether the agent reproduces approved metrics before expanding its reach.

A knowledge blending approach can help users understand this design. Structured metrics and qualitative documents serve different purposes. Combining them becomes useful only when the system preserves enough context to show where each conclusion came from.

Natural Language Does Not Remove the Governance Problem

The Data agent lowers the skill barrier at the interface, but it raises the importance of definitions, access controls, validation, and responsible sharing.

OpenAI says queries enforce the existing permissions of a connected account, including table, row, and column restrictions. That is essential because a conversational interface should not grant access that a user lacks in the source system. Administrators also choose which connections are available and which roles can use them.

Those controls reduce one class of risk. They do not answer every governance question. A user might have legitimate access to sensitive data but still create a dashboard intended for a broader audience. The act of publishing or sharing can introduce a second permission boundary.

OpenAI’s help documentation warns that data used in an analysis can be copied into a published site. It tells users to consider permissions when choosing recipients. This means source-level authorization and output-level distribution must be reviewed separately.

A dashboard can also reveal sensitive information through aggregation. Small groups, narrow filters, or unusual combinations may expose details that individual rows would otherwise obscure. Organizations need publication policies that address what a visualization communicates, not only whether the query succeeded.

Accuracy presents another challenge. Large language models can interpret vague requests confidently, while business data often contains silent assumptions. “Revenue,” “customer,” and “active user” can each have several valid definitions. A natural-language question may conceal those choices rather than resolve them.

Semantic layers are OpenAI’s main answer. By grounding the agent in approved calculations and relationships, an organization can reduce ambiguity. The product can also show evidence and let users question a finding within the same conversation.

Yet semantic layers require ongoing maintenance. New products alter metric definitions. Teams rename fields, change attribution rules, and migrate databases. If the governed context falls behind the business, the agent can deliver a well-explained answer based on obsolete assumptions.

The risk becomes more serious when the agent recommends action. A diagnosis might influence staffing, spending, customer prioritization, or product investment. Those decisions demand more than a plausible chart. Users need to know which data was included, what period was examined, and which alternative explanations remain.

OpenAI’s own guidance encourages that review. Users should verify sources, filters, time ranges, and metric definitions. If an answer conflicts with an existing report, they should ask the agent to compare those details. That is a practical acknowledgment that conversational convenience does not guarantee correctness.

Competitors make similar qualifications. Microsoft notes that generative outputs can be nondeterministic, so people should evaluate and validate them. Tableau documents source and workflow limitations, while retaining existing row-level and column-level security policies.

These caveats point toward a shared industry conclusion. AI can make analytics more accessible, but dependable self-service still rests on curated models and human review. The conversational layer changes the experience. It does not repeal the principles of data governance.

Organizations evaluating how the Data agent works should test reproducibility, not just presentation quality. Can two users with the same permissions obtain consistent calculations? Does the agent cite the same definitions used by official reports? Can reviewers reconstruct the steps behind a surprising result?

They should also test permission separation. A useful pilot would include users with different regional, departmental, and managerial access. The goal is to confirm that answers and dashboards respect those distinctions throughout analysis, editing, refresh, and sharing.

Another test involves adversarial ambiguity. Teams can ask underspecified questions and observe whether the agent requests clarification or silently chooses a definition. A trustworthy system should surface consequential assumptions before presenting a recommendation.

Data quality remains a final constraint. Duplicate records, delayed ingestion, missing values, and poorly documented joins can mislead traditional dashboards and AI agents alike. Natural language may make these problems less visible because the user never sees the underlying query.

This is the main tension behind Now everyone can put data to work. Wider access can shorten the distance between a question and a decision. It can also multiply the number of people making decisions from data they did not model themselves.

The launch will succeed only if companies treat governance as part of the product experience. Definitions should appear alongside findings. Evidence should remain inspectable. High-impact decisions should require review, and published dashboards should have clear owners.

Three Signals Will Show Whether the Promise Holds

The next test is not whether the Data agent can draw a convincing dashboard, but whether organizations can trust and repeatedly use its analysis.

The first signal is verified adoption beyond selected alpha examples. OpenAI has named customers using the agent for sales, spending, operations, supply planning, and product adoption. The stronger evidence will come from recurring use across multiple departments, with documented checks against established reports.

If organizations keep using the agent after initial pilots, the workflow is solving a durable problem. If usage concentrates among technical specialists, the claim that Now everyone can put data to work will look narrower than the headline suggests. Adoption should be measured by completed, reviewed decisions rather than prompt counts alone.

The second signal is how well OpenAI standardizes provenance and validation. Users need visible sources, applied filters, metric definitions, refresh times, and calculation logic. Reviewers should be able to distinguish a sourced finding from the model’s interpretation.

Clear provenance would strengthen OpenAI’s position as an analytical control layer. Weak or inconsistent provenance would push cautious companies back toward their existing BI environments for consequential work. A conversational interface earns trust when it makes verification easier, not when it merely hides complexity.

The third signal is the response from Microsoft, Salesforce, and major data platforms. Power BI and Tableau already support natural-language questions and visual analysis. Warehouse providers are also adding AI interfaces grounded in their own governed metadata.

Competitors can answer OpenAI by expanding cross-source reasoning, improving agent workflows, or making their products easier for nontechnical users. They can also emphasize that trusted analysis should remain inside the platform where data models and permissions are already managed.

OpenAI’s partner strategy may soften that conflict. Snowflake, Databricks, MongoDB, AWS, ClickHouse, and G2 contributed supportive statements to the launch. Their participation indicates that major data providers see value in making their systems accessible through ChatGPT Work.

Partnership does not remove competition over user attention. If employees ask ChatGPT first, OpenAI controls the starting experience even when another company stores the data or renders the final chart. That position can become strategically important as agents take on longer analytical workflows.

For enterprise buyers, a measured pilot is the sensible next move. Choose one decision process with stable definitions, such as weekly pipeline review or product adoption reporting. Compare the agent’s outputs with approved dashboards, record disagreements, and test every sharing path before expanding access.

Give the pilot a clear success standard. Useful measures include agreement with certified metrics, time saved on repeat questions, the number of required corrections, and whether decision-makers can inspect supporting evidence. Avoid treating a polished demonstration as proof of operational reliability.

For individual knowledge workers, the practical opportunity is faster exploration. Use the agent to formulate questions, compare segments, and create an initial dashboard. Then examine definitions and evidence before turning the output into a recommendation.

The Data agent makes an ambitious bet about the future of business intelligence. It assumes people will prefer to state their goal and let an agent coordinate the systems underneath. That is a credible direction, especially when questions cross warehouses, documents, and established dashboards.

The harder question is whether organizations can make that convenience dependable. Watch how OpenAI exposes provenance, how customers report sustained adoption, and how incumbent BI vendors respond. Those signals will determine whether everyone can put data to work, or whether trusted analytics still depends on a smaller group standing behind every answer.