OpenAI's own employees couldn't get fast answers to basic data questions — a gap that persisted until the company built a fix for itself first, then decided to sell it.

That's the story behind the Data agent OpenAI is launching Thursday inside ChatGPT Work, OpenAI's competitor to Anthropic's Cowork, which launched in July. The Data agent connects to enterprise data sources, investigates what changed and turns the findings into dashboards, recommendations and follow-up actions it can execute through connected tools.

The fix was not just more connectors. OpenAI built an agent that orchestrates context across the tools a business already runs. It pulls from Slack, BI dashboards, data warehouses and file storage in the same breath, instead of asking anyone to move that data into one place first.

"If you asked me a year ago what percentage of our data questions were getting answered end to end without needing anyone intervening, it was a very small fraction of the company," Arpan Shah, general manager of the enterprise technology vertical at OpenAI, told VentureBeat in an exclusive interview. "Today, everyone at OpenAI can self-serve any data questions they have, and that was a gap we kept hearing externally as well."

How OpenAI's own data bottleneck built this agent

The Data agent is not a new build dressed up for launch. Shah described it as the externalized, generalized version of an internal tool. That internal tool was built specifically around OpenAI's own data, permissions and workflows, before any customer touched it.

OpenAI first detailed that internal tool in a January 2026 post on its engineering blog with additional insights disclosed on September 8 on how it is used. The company described a data platform serving more than 3,500 internal users across engineering, product and research. It spans over 600 petabytes of data across 70,000 datasets.

At that scale, the company wrote, simply finding the right table could consume more time than the analysis itself. The internal agent was built to close that gap. It combines schema metadata, human annotations, code-level analysis of how tables are built, and a memory layer that carries corrections forward between sessions.

"It's very heavily inspired, taking all of those things and really externalizing that in the form factor that works well with all of the ecosystem partners that you already have and love," Shah said. "Obviously the internal tool is very opinionated around a certain stack, and we've generalized that so that it works with all of the various tools, whether it's BI [business intelligence] tools or data warehouses or other kinds of connectors and plugins that you might want to bring along, and works super well for all of those as well."

Context orchestration, not just another MCP connection

OpenAI and every other LLM has had the ability to connect with data sources via MCP as well as RAG workflows to pull in sources.

The distinction OpenAI is drawing is about what happens after the connection exists, not the connection itself.

“You can of course connect your AWS Redshift or your Databricks or Snowflake and bring all of those data tools and connectors along," Shah said. "But then when you use the data agent, it can share context and pull in context from your Slack and your Google Drive, and actually orchestrate across all of these different tools to make sure that the right pieces of information are both pulled from and then shared back to those ecosystems.”

Among the data tools and platforms that the agent can connect with are:

  • Data warehouses and platforms: Amazon Redshift, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake and Datadog, with more to come

  • Files: Google Drive and SharePoint

  • Dashboards: Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot

The orchestration decides where an insight ends up. Shah named three examples: reaching out to an account owner armed with data, preparing an update for leadership, and refreshing a deck's charts automatically with the latest numbers. Shah framed all three as part of the same end to end workflow the agent is meant to carry a user through.

"MCP is a starting point, and the data agent actually helps orchestrate that specifically for data workflows," Shah said.

No context layer, no published benchmark, and where rivals still win

For a product built to answer questions about enterprise data, OpenAI is notably specific about what it did not build. Asked whether the agent constructs a persistent context layer across a customer's connected tools, the answer is no.

"It combines all the various contexts in their state, as opposed to generating a net new context layer itself," Shah said, describing the process as stitching context together on the fly rather than building a persistent context layer.

For a retrieval accuracy or correctness benchmark, the kind of number most vendors in this category lead with, there is no published figure. Shah said the internal benchmark compares the agent's results against OpenAI's own data tools to confirm parity with what the company expects externally, without citing a published number. That is an internal comparison, not an independently verified benchmark. OpenAI has not published one for the external product.

This week, Databricks published research claiming its Adaptive Instructed-Retriever matches the answer quality of Claude Sonnet 5, GPT-5.6 Luna and DeepSeek-V4-Flash while answering more than twice as fast — an average of 5.8 seconds. OpenAI has no published benchmark for the Data agent. What it offers instead is a claim about process, not a claim about accuracy. Shah calls that process hill climbing, the same iterative internal refinement that built the tool before any customer touched it.

"We've gone through the journey internally of hill climbing, and that in turn has made our models really good at this stuff already," Shah said. "As we've internally hill climbed on making our tools really good at data work, we're distilling all of those principles into this data agent and product, which is super easy to install from the plugin directory."

That said, the data agent is not intended to be a tool that will replace an organization’s use of a specific data tool it might already be using, be it from Databricks, Snowflake or anyone else.

OpenAI's bet is that as the number of tools a business runs grows, orchestration across all of them beats depth in any single one. OpenAI is not necessarily competing directly against those vendors, either; rather, it's partnering with them. As such, the data agent is just another option for enterprise users.

"The way I think about it is that as the breadth of things that you want the agent to span increases, I think ChatGPT Work really shines because it can go across all those surfaces and do the right thing," Shah said.

What this means for enterprises

For a generation of business people, getting access to the right data and then being able to use it required a data analyst. Those days are now increasingly in the past.

That shift does not remove the need for a data engineering team. It moves the work, from writing one-off queries to maintaining the access controls and audit trails that let a self-serve agent answer safely without a human checking every result. 

"My hope and expectation is that a lot of businesses are getting the same out of their data as we get here, and data is truly democratized within these businesses,” Shah said. “My hope is that the salesperson or marketing person no longer has to go to a data analyst to get their questions answered.”