OpenAI Data Agent Enters an Enterprise Analytics Race Built on Trust

OpenAI launched the OpenAI Data Agent on September 10, bringing governed company analysis into ChatGPT Work as the enterprise analytics contest intensifies.

The new agent connects to approved business systems, investigates changes, and builds interactive dashboards from natural-language requests. It also recommends follow-up actions and can send approved outputs through connected workplace tools.

That scope puts OpenAI into direct competition with Google’s increasingly mature conversational analytics stack. Yet the contest is not simply about which model writes better SQL. It is about which platform can understand business definitions, respect permissions, expose evidence, and earn enough trust to influence real decisions.

OpenAI enters with a useful advantage. The company already tested a related internal system across a complex data environment before packaging its capabilities for customers. However, Google, Microsoft, Snowflake, Databricks, and established business intelligence vendors control important parts of the existing analytics stack.

The resulting battle has an unusual shape. OpenAI wants ChatGPT to become the place where employees ask, investigate, visualize, and act. Its partners still want their warehouses, semantic layers, and dashboards to remain the authoritative foundation underneath that experience.

The OpenAI Data Agent Moves From Answers to Action

The important change is not natural-language querying. It is the attempt to place an entire analytics workflow inside one ChatGPT conversation.

According to the Data Agent launch, users can connect approved sources such as Amazon Redshift, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake. Files from Google Drive and SharePoint can also inform an analysis.

The agent does more than translate a question into a database query. OpenAI says it can investigate why a metric changed, inspect supporting evidence, accept follow-up directions, and convert its findings into an interactive dashboard.

That workflow targets a familiar delay inside large organizations. A sales leader notices that renewals weakened, asks an analyst for an explanation, and waits while the analyst finds appropriate tables. The analyst must then confirm definitions, check joins, reconcile dashboards, and package the findings.

The OpenAI Data Agent compresses those steps into a conversational loop. A user can ask why weekly active users fell, request comparisons with earlier periods, and direct the agent toward likely causes. The agent can then propose additional checks instead of ending with a static chart.

OpenAI also says the agent can create or interact with dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. This detail matters because OpenAI is not asking every customer to abandon its current visualization layer immediately.

Instead, ChatGPT Work becomes an orchestration surface. The agent can reach across data stores, business documents, semantic systems, and existing dashboards while preserving those systems as sources of truth.

A semantic layer is the governed set of definitions that explains how a company calculates measures such as revenue, retention, and active users. Those definitions prevent separate teams from attaching different meanings to the same label.

OpenAI says its agent can use metric definitions, custom calculations, relationships, and business terminology from sources including dbt, GitHub, Snowflake Horizon, and Databricks Genie Ontology. This grounding separates the product from a generic text-to-SQL interface.

Administrators retain a central role. They choose available connections, configure data-source plugins, and decide which roles can use them. Queries are supposed to enforce the connected user’s existing table, row, and column permissions.

The agent appears as Data in the ChatGPT Work Plugins directory. Once an administrator installs it and configures relevant connections, employees can invoke it by addressing @Data inside a conversation.

OpenAI says the final step can extend beyond analysis. The agent can identify people who should participate, recommend next actions, share findings through Slack or email, and perform actions that a user approves.

That progression changes the competitive stakes. A dashboard traditionally helps someone understand what happened. An agent that analyzes a change and initiates a response starts competing for control of the decision workflow itself.

Enterprise Analytics Is Becoming a Fight for the Front Door

OpenAI is pressuring analytics vendors by making ChatGPT the first place an employee turns with a business question.

Business intelligence platforms have spent years moving beyond static reports. Tableau, Power BI, Looker, ThoughtSpot, and newer data platforms already support natural-language questions, automated explanations, and assisted visualization.

The OpenAI Data Agent does not arrive before that shift. It arrives after conversational analytics has become a strategic priority across the industry.

Google made Conversational Analytics in BigQuery generally available on June 30, 2026. Its system lets technical and business users run multi-step analyses and generate visual reports with natural language inside BigQuery.

Google’s conversational analytics can also reach cross-cloud sources, including Databricks Unity, AWS Glue, SAP, and Salesforce. Data teams can publish specialized agents into Gemini Enterprise, Data Studio, or custom applications.

That gives Google a clear position in the contest. It controls a major cloud data warehouse, the Gemini model family, enterprise interfaces, and the infrastructure where many queries execute.

OpenAI is taking a different route. It wants ChatGPT Work to sit above multiple data platforms, including competitors’ systems. The product can become the common conversational layer even when the underlying warehouse belongs to Google, Snowflake, Databricks, or Amazon.

This is why the launch creates pressure beyond traditional business intelligence. The platform that owns the employee’s starting point can influence which tools remain visible and which become infrastructure hidden behind a conversation.

A finance leader might no longer open a dashboard catalog, choose a report, and adjust filters. The leader could start in ChatGPT, describe the decision, and let the agent locate relevant datasets and dashboards.

That shift does not automatically eliminate BI platforms. OpenAI’s announced integrations suggest a layered market where ChatGPT handles intent and orchestration while established systems provide governed metrics, storage, visualization, and access control.

The strategic question concerns who captures the durable user relationship. If most analysis begins in ChatGPT, an existing dashboard product risks becoming a rendering destination rather than the primary workspace.

OpenAI’s partners recognize that possibility, yet participation also expands their reach. Microsoft says Power BI users can create dashboards by describing business questions. Tableau emphasizes that its semantic layer can ground answers while ChatGPT supplies another place to work.

Snowflake positions its governed data and access controls as a shared foundation, regardless of where an employee asks the question. Databricks similarly presents Genie as the context layer that bridges enterprise data and trustworthy answers.

These statements reveal a negotiated boundary. OpenAI wants to own the conversational experience, while partners want to preserve authority over data, definitions, and governance.

Google is the clearest primary opponent because it can combine those layers internally. Its analytics agents live close to BigQuery data and can move into Gemini Enterprise without depending on an external conversational interface.

OpenAI’s cross-platform approach offers flexibility, but it also creates integration work. Every additional system introduces different schemas, permissions, metadata conventions, refresh schedules, and operational limits.

The race therefore turns on more than feature lists. Customers must decide whether an independent orchestration layer offers enough convenience to justify another critical control point in their analytics architecture.

Business Context, Not SQL, Is the Hard Part

The OpenAI Data Agent succeeds only when it can distinguish a technically valid answer from the answer a company actually means.

Large language models can generate syntactically correct SQL while choosing the wrong table, grain, filter, or business definition. The resulting answer may look persuasive because the query runs and the chart appears coherent.

OpenAI encountered this problem inside its own operations. In January, the company described an internal data agent serving more than 3,500 users across over 600 petabytes and 70,000 datasets.

Its internal architecture illustrates why enterprise analysis requires much more than database access. The system combines schema metadata, historical queries, human annotations, code-derived context, institutional knowledge, memory, and live warehouse inspection.

Schema metadata tells the agent about columns and data types. Table lineage describes how datasets connect and which upstream processes created them. Historical queries reveal common joins and established usage patterns.

Those ingredients still cannot fully explain an organization. Two tables can contain similar fields while applying different exclusions. An apparently simple metric can depend on undocumented conventions, experimental filters, or product-specific logic.

OpenAI’s internal agent addresses that ambiguity with human-written descriptions and code-level analysis. Codex examines pipelines to infer a table’s ownership, granularity, dependencies, join keys, and update behavior.

The agent can also retrieve institutional context from workplace systems. Internal documents may explain product launches, incidents, codenames, and canonical metric definitions that database schemas cannot express.

Memory adds another layer. When users correct a subtle filter or clarify an ambiguous term, the internal system can preserve that lesson for later conversations. OpenAI says users can review and edit these memories.

The commercial OpenAI Data Agent follows the same central idea, even though OpenAI has not publicly established that every customer deployment reproduces the internal architecture. Company-specific context must accompany raw access.

This mechanism explains the integrations with semantic systems. OpenAI needs Databricks Genie Ontology, dbt models, Snowflake Horizon, BI dashboards, and other trusted sources because the model cannot safely infer every business rule.

It also explains why established data vendors remain difficult to displace. Snowflake’s Cortex Analyst already grounds natural-language questions in semantic models attached to structured data.

Google lets teams configure data agents with verified queries, custom metadata, glossary terms, and explicit instructions. Its documentation recommends validating generated output because plausible answers can still be factually wrong.

OpenAI faces the same underlying limitation. A reasoning model can explore multiple hypotheses, but its confidence does not guarantee that the chosen dataset represents the organization’s approved truth.

The practical test is not whether an agent produces a dashboard. It is whether an analyst can inspect its evidence, reproduce its logic, and identify the assumption behind an unexpected result.

This is also where knowledge management becomes operational infrastructure. Teams need maintained definitions, documented decisions, and searchable business context before an agent can reason reliably across departments.

A well-organized AI knowledge base can help employees locate relevant context. Enterprise analytics adds stricter requirements around lineage, permissions, approval, and metric ownership.

Without that foundation, self-service analysis can multiply disagreements rather than resolve them. More employees can generate answers, but each answer may depend on a different interpretation of the business.

The Trust Problem Does Not Disappear Behind a Dashboard

A polished chart can make an analytical mistake more dangerous because it gives uncertain reasoning the appearance of settled evidence.

OpenAI says users can review evidence behind each finding and refine an investigation conversationally. That design is preferable to a one-shot answer, but public details do not yet establish error rates across varied customer environments.

The launch materials include promising customer examples. NTT DATA says non-engineers in sales and corporate roles built and updated dashboards using plain language. The company points to licensing, effort, and technical expertise as barriers to broader dashboard use.

ServiceTitan says its team created a dashboard indicating that users of its Atlas assistant launched campaigns at roughly three times the rate of nonusers. The result reportedly informed changes to onboarding.

CookUnity says its growth team built a seasonal conversion dashboard, checked the output against internal reports, and reduced planning time. That verification step is significant because it preserved an independent reference point.

Another evaluation partner, micro1, says its operations team rebuilt performance-tracking dashboards in half an hour and found errors in the originals. These accounts show useful scenarios, but they remain customer statements presented by OpenAI.

They do not provide a standardized comparison across accuracy, latency, total operating cost, permission failures, or analyst review time. OpenAI has not published broad independent evidence showing how often the commercial agent reaches a correct conclusion without intervention.

The distinction matters because analytics errors have several layers. The generated query can be invalid, the selected table can be inappropriate, or the business question can be too ambiguous for one defensible answer.

Even a correct calculation can produce a misleading recommendation. A decline in renewals may correlate with support volume, but that relationship does not prove which operational change will improve retention.

Giving the agent action tools raises the consequence of these mistakes. Sending a report is lower risk than changing a campaign, reallocating staff, or modifying a customer workflow.

Human approval provides a control point, but approval quality depends on visibility. A manager cannot meaningfully supervise an action without understanding the data, assumptions, confidence, and alternative explanations behind it.

Permissions present another challenge. OpenAI says queries follow existing access rules, including row-level and column-level restrictions. That protects direct retrieval, but enterprises must also consider what a generated dashboard reveals through aggregation.

An employee may lack access to individual records yet infer sensitive information from a small group. Security teams will need to test whether existing policies behave as intended across follow-up questions and combined sources.

Cross-system reasoning creates further exposure. A question that joins sales, product, finance, and support data may be valuable precisely because those domains were previously separated.

It can also produce an answer that no single system was designed to authorize. Administrators must examine whether the composition of individually permitted facts reveals a restricted business conclusion.

Memory deserves similar scrutiny. Remembering corrections can improve future analysis, yet organizations need clear ownership for shared memories. A mistaken rule should not quietly become a durable source of truth.

Teams also need change controls when a metric definition evolves. The agent must know which definition applies to earlier periods, current reporting, regulated disclosures, and experimental analysis.

OpenAI’s own usage offers encouraging evidence that the architecture can operate at substantial scale. The company says nearly all of its product team and more than two-thirds of its go-to-market organization use data agents in ChatGPT Work.

However, OpenAI is an unusually AI-native environment. Its employees, data teams, internal tools, and tolerance for experimentation do not represent every healthcare provider, bank, manufacturer, or government agency.

The company also built its internal system around its own workflows. A customer connecting several mature platforms inherits organizational complexity that a product integration cannot automatically clean up.

For buyers, the safest early deployments have clear metrics, bounded data sources, reviewable queries, and reversible actions. High-stakes recommendations need stronger evaluation than a dashboard that summarizes campaign activity.

The OpenAI Data Agent should therefore be judged as an analytical collaborator, not an unquestioned decision engine. Its value depends on how well it exposes uncertainty and supports verification when the evidence is incomplete.

Three Signals Will Decide the Enterprise Analytics Race

The next stage will be decided by verified adoption, governance under real workloads, and the response from platforms that already control enterprise data.

The first signal is sustained use beyond alpha programs. OpenAI has named organizations testing sales, spending, reporting, staffing, and operational analysis, but initial enthusiasm does not establish durable adoption.

Watch whether nontechnical employees keep using the agent after early demonstrations. Repeated weekly use, shared dashboards, and reduced analytical backlogs would strengthen OpenAI’s claim that conversational investigation can become routine work.

The quality of questions also matters. Simple summaries are widely available. Durable differentiation requires users to trust the agent with multi-step investigations that cross datasets and challenge existing reports.

A rise in those workflows would reinforce OpenAI’s front-door strategy. If usage remains limited to basic chart creation, established BI interfaces will retain more of their current role.

The second signal is evidence about accuracy and governance. Enterprises should look for published evaluation methods, audit capabilities, correction workflows, and clear boundaries around actions.

OpenAI’s own enterprise data indicates that adoption is moving toward deeper delegation. As of June 2026, Codex produced 64 percent of combined Codex and ChatGPT output tokens among enterprise customers.

The company also reported that its highest-usage enterprise customers generated 8.3 times more output tokens per active user than typical firms. OpenAI notes that token volume is an imperfect measure of business value.

That caveat applies directly to data agents. More queries, longer investigations, and additional dashboards do not prove better decisions. Buyers need measures tied to correctness, analyst review, time saved, and business outcomes.

Incident reporting will be equally revealing. A transparent account of failed joins, stale context, permission mistakes, or misleading recommendations would help customers assess operational maturity.

If OpenAI supplies reproducible evaluations across different warehouses and semantic layers, its cross-platform position becomes more credible. If evidence remains dominated by selected customer stories, the trust gap will remain open.

The third signal is how Google and the data-platform vendors respond. Google already offers generally available conversational analytics inside BigQuery and distributes agents through Gemini Enterprise and APIs.

Its data agent controls include verified queries, custom instructions, glossary terms, access restrictions, and audit logging. Google also prevents these agents from running write operations or data-manipulation queries in the documented BigQuery experience.

That boundary creates a clear contrast. OpenAI emphasizes movement from analysis toward approved action, while Google’s documented BigQuery controls emphasize a constrained analytical environment.

Neither approach is universally better. A narrower system can be easier to govern, while a broader agent can remove more workflow friction.

Snowflake, Databricks, Microsoft, Tableau, and other partners must also decide how much of the user experience to concede. They can deepen ChatGPT integrations, strengthen their own agents, or make semantic governance their main source of differentiation.

Partner support currently helps OpenAI reach enterprise data without demanding a full migration. Over time, those same partners could resist becoming interchangeable back-end services behind ChatGPT.

The decisive competitive response will therefore appear in product behavior, not launch statements. Watch where dashboards are created, where definitions are edited, where permissions are administered, and where users return for follow-up questions.

Enterprises should begin with one consequential but reviewable workflow. A product team might investigate an adoption decline, compare findings with an approved dashboard, and record every correction required.

That test should include ambiguous terminology, stale documentation, restricted columns, and conflicting metrics. It should also measure how much analyst time moves from query writing into review and governance.

The OpenAI Data Agent makes enterprise analytics easier to access, but access was never the only bottleneck. Trusted context, careful permissions, and accountable decisions remain the harder work.

The central question for buyers is now practical: does the agent help more people reach defensible answers, or does it simply let more people produce convincing dashboards?

Organizations that test that distinction carefully will learn where conversational analytics belongs. Those that treat fluent output as verified truth risk automating the most expensive kind of mistake.