OpenAI calls it Data: a new agent available through the ChatGPT Work Plugins directory, designed to investigate business questions using approved company data. OpenAI's announcement, published September 10, 2026, describes Data as turning company data into answers, interactive dashboards, and action — connect ChatGPT to your company's data, ask a business question in plain language, investigate what changed, and turn the result into an interactive dashboard or report. OpenAI's own examples include questions like why sales slowed, where spending increased, and which issues threaten renewals in major accounts. From here on, I'll use "Data agent" as reader-friendly shorthand.
That sounds like a better analytics interface. I think the more interesting shift is deeper than that.
The important change isn't that AI can make another dashboard. It's that analysis is moving closer to the person making the decision.
Traditionally, a business question begins with someone noticing a problem and waiting for somebody else to translate it into data work: a manager asks what changed, an analyst pulls the data, someone builds or updates a report, the business reviews it, another question appears, the loop starts again. OpenAI's Data agent is aimed directly at shortening that loop. If systems like this work well, the future of analytics may become less about navigating predefined dashboards and more about asking increasingly specific questions of trusted business data until you understand what's happening well enough to act.
What is OpenAI's Data agent?
The Data agent is a plugin for ChatGPT Work and Codex that connects to approved company data sources and lets users investigate business questions conversationally. It can analyze connected data, incorporate company-specific context, answer follow-up questions, and create interactive dashboards or reports.
OpenAI's documentation says supported sources include systems such as Amazon Redshift, Google BigQuery, Databricks, Snowflake, ClickHouse, and MongoDB, plus documents from Google Drive and SharePoint. It also works alongside existing BI environments including Power BI, Tableau, Sigma, Oracle BI, Omni, and ThoughtSpot.
At a high level: business question → connected company data → business context and definitions → analysis → findings → follow-up questions → dashboard or report → decision. The interface is conversational. The underlying problem is not.
The real value is reducing the distance between question and analysis
Imagine a sales leader notices revenue slowed. The traditional path — question → request to the analytics team → query and data prep → report → review → follow-up → new query — isn't inherently broken, but it creates friction: the person with the business context often isn't the person with the technical ability to interrogate the data directly.
The Data agent compresses those steps. A leader can ask why sales slowed, then whether it was concentrated in a segment, then whether it's a pipeline, conversion, or expansion problem, then which accounts contributed most, then ask for a shareable dashboard. OpenAI built the agent around this kind of iterative investigation rather than one-shot reporting, inspecting evidence, refining the analysis, and turning findings into dashboards as you go.
That changes the interaction. A traditional dashboard usually starts with someone deciding in advance which metrics and views people will need. A conversational agent lets the question come first. That doesn't mean dashboards disappear; it means the dashboard becomes the result of an investigation rather than always its starting point.
This is more than "upload a CSV to ChatGPT"
We've already been able to give AI models spreadsheets and ask questions about them. Enterprise analytics gets harder when data is spread across multiple systems and only makes sense in the context of how a company actually defines its business — "revenue," "active customer," and retention calculations can all vary by team, alongside custom fiscal periods, product groupings, and account hierarchies. Ignore those definitions and the AI can produce an answer that sounds intelligent and is still wrong.
OpenAI's documentation says the Data agent can use organizational context from trusted sources including semantic layers, metric definitions, custom calculations, and data relationships, citing sources such as dbt, Databricks Genie Ontology, Snowflake Horizon, GitHub, and existing BI dashboards as places that context can come from. The goal isn't giving the model access to more rows; it's giving it access to the information required to interpret those rows correctly. That's a much harder problem.
Data without context is not intelligence
A company can have enormous amounts of data and still struggle to answer simple questions, because useful intelligence requires more than storage: raw data → business definitions → relationships → context → analysis → interpretation → decision. If any of those layers are weak, a conversational interface on top doesn't fix them.
This is why OpenAI's description of its own internal use is worth noting: in an internal engineering write-up published January 29, 2026, the company says nearly all of its product team and more than two-thirds of its go-to-market organization use data agents internally, but its data team first had to create shared business definitions, access rules, and safeguards for sensitive data. That may be the most important line in the entire announcement. The AI interface is the visible layer; the trusted data system underneath is what makes the answers useful.
Where Skills fit into this stack
This is also where OpenAI's recent product releases start connecting to each other rather than reading as isolated announcements. The Data documentation notes that organizations can use ChatGPT Skills to modify the Data plugin's default behavior and provide context for specific topics, analysis practices, or dashboard design systems, with reusable templates standardizing recurring analyses and reports.
| Layer | Role |
|---|---|
| Data plugin | Access to business data |
| Semantic layer | Defines what the data means |
| Skills | Sets how the analysis should be performed |
| AI | Reasoning and investigation |
| Tools | Where results go and what actions happen |
That's a more useful way to think about this than treating the Data agent as a standalone feature. Access without meaning is noise; meaning without a defined process is inconsistent; a defined process without reasoning is just a template; reasoning without a destination doesn't change anything downstream. Once you notice this pattern, ChatGPT Skills — as explained in the research Skill I built and tested — stop looking like a separate product and start looking like the layer that makes agents like Data behave consistently rather than brilliantly-but-unpredictably.
What is this actually good for?
The strongest use cases are situations where somebody understands the business question but doesn't know how to query every underlying system. OpenAI gives examples across sales, spending, retention, product performance, operations, and leadership reporting — questions like why conversion fell this month, which customers show the strongest renewal-risk signals, where spending increased unexpectedly, which features correlate with stronger retention, or where process delays are increasing. Those aren't dashboard questions. They're business questions, and that distinction matters.
The dashboard isn't the end of the system
Analytics often stops at data → insight → dashboard. But businesses don't benefit because somebody produced a visualization; they benefit when insight changes a decision or action. The fuller loop is data → signal → analysis → insight → decision → action → outcome → new data.
OpenAI is already gesturing at this broader model: the announcement describes Data as helping teams turn company data into "answers, interactive dashboards, and action," with results shareable through tools like Slack or email. Depending on the connected tools, existing permissions, and required approvals, findings may also be used to trigger approved follow-up actions through connected systems. That starts to look less like traditional BI and more like an intelligence-to-action system, which is where this becomes strategically interesting rather than just a nicer chart generator.
But easier analytics does not remove the hard parts
A conversational interface can make analytics substantially more accessible. It does not repeal the underlying problems of analytics.
Bad data is still bad data. If underlying records are incomplete, inconsistent, or wrong, the model is reasoning over a flawed representation of reality. AI may make bad data easier to query, not more trustworthy.
Definitions still matter. If Sales and Finance define "active customer" differently, a conversational interface on both sides doesn't resolve the disagreement; it may just surface it faster. Somebody still has to decide which definition governs the analysis.
Correlation is not causation. An agent finding that two metrics moved together doesn't mean one caused the other. The easier analysis becomes, the more tempting it is to slide from "these are associated" to "this explains what happened" — not the same claim. Treat an agent's explanation as a hypothesis to investigate, not proof that it identified the cause of a business outcome.
Permissions become extremely important. OpenAI says the Data agent respects the connected account's existing permissions, including table-, row-, and column-level restrictions, with administrators controlling which connections and roles are available. The objective isn't "everyone can access all the company data" — it's people more easily interrogating the data they're already authorized to use.
A confident answer is not automatically a correct answer. Polished explanations get treated as evidence of understanding more readily than they should. The Data agent lets users inspect the evidence behind a finding, which is exactly the habit worth encouraging: answer → evidence → check assumptions → follow-up → decision, not answer → trust it.
Does this eliminate the data analyst?
No, and framing it that way misses the opportunity. If business users can answer straightforward questions independently, that does reduce some dependency on analysts for routine requests. But organizations still need people who understand data architecture, metric design, semantic layers, experimentation, causal inference, governance, and how to tell an interesting pattern from a misleading one.
OpenAI's own internal engineering post suggests strong data teams become more important, not less, because somebody has to build the trusted foundation that makes self-service analysis possible. Its internal system relied on schema metadata, lineage, historical queries, expert annotations, code-level context, and permissions-aware retrieval rather than simply granting raw data access. The analyst role shifts upward: less "build me this chart," more "are we asking the right question, is this metric meaningful, can we trust this conclusion, what should this analysis inform." That's a higher-leverage use of the same expertise.
Who should care about this, and who shouldn't rush toward it?
| Situation | Take |
|---|---|
| Mature data infrastructure, established metric definitions, decision-makers frequently waiting on reports | Strong fit: the delay between question and trustworthy answer is the exact friction this targets |
| Multiple systems that need to be investigated together | Strong fit |
| Nobody agrees on what the key metrics mean | Fix that first: a conversational layer surfaces the disagreement faster, it doesn't resolve it |
| Data scattered across undocumented spreadsheets, inconsistent systems | The bottleneck is infrastructure and governance, not the interface |
| Team already drowning in dashboards nobody uses | Generating dashboards faster isn't the fix |
The first question shouldn't be "how do we deploy a Data agent?" It should be "what actually prevents the right person from getting trustworthy information when they need to decide?" The answer might be access, analyst capacity, fragmented systems, poor definitions, bad data, or organizational behavior. The technology should address the real constraint, not just the visible one.
The bigger shift: from reporting to investigation
Traditional BI is organized around artifacts: dashboards, reports, queries, charts. The Data agent shifts the interface toward a conversation: what happened → why → where → which segment → what changed → what should we check next. That resembles how people actually investigate problems; we rarely know the full analytical path before we start, and each answer creates the next question. A conversational interface makes that loop far more natural than static reporting ever could.
The real metric isn't how many dashboards people create
OpenAI's launch includes examples of customers building dashboards faster and expanding analytical access to nontechnical users. Those are reasonable adoption signals, but they're not the ones I'd actually track. The better questions: did people reach useful answers faster, did those answers hold up under verification, did decisions improve, were problems caught earlier, did analyst time move toward higher-value work, did teams act on what they found, and did that action actually change an outcome. More analysis doesn't equal better decisions. The value shows up further down the chain than dashboard-creation counts can measure.
So does OpenAI's Data agent matter?
I think it could — not because companies desperately need another way to make charts, and not because AI eliminates the hard work of building trustworthy data systems. It matters because it attacks a real organizational constraint: the person who has the question is often separated from the tools and expertise required to investigate it. Data agents can shrink that distance.
The architecture becomes: company data → trusted definitions → AI investigation → business user → follow-up questions → insight → decision → action. That's a real change from waiting on a static report. But the companies that get the most value won't be the ones that install the plugin fastest; they'll be the ones that already understand that AI doesn't turn messy information into a good decision by itself. You still need trusted data, shared definitions, appropriate access, evidence, and human judgment connecting the insight to the action that follows.
The real opportunity isn't another dashboard. It's a shorter, stronger loop between data → signal → intelligence → decision → action. If that loop gets meaningfully better, Data agents become much more than an analytics feature.
Sources & updates
Reviewed September 12, 2026. Product details verified against OpenAI primary sources. No affiliate links.