The ChatGPT Data Agent Removes the Query, Not the Responsibility

The ChatGPT Data agent is a plugin for ChatGPT Work that connects to your company's approved data sources, investigates why a business metric moved, and builds interactive dashboards colleagues can share. OpenAI introduced it on 9 September 2026, and the pitch is blunt: ask why sales slowed in plain English, and get an answer without writing a query or opening a separate analytics tool.

If you run marketing measurement, this is the most consequential release of the week, and it is not because of what it can do. It is because of who can now do it.

What the Data agent connects to

The connector list is the part that makes this credible rather than a demo. The Data agent reaches warehouses and databases including Amazon Redshift, Google BigQuery, Snowflake, Databricks, ClickHouse, MongoDB and Datadog. It can pull files and documents from Google Drive and SharePoint into an analysis.

It also works with existing dashboards in Power BI, Tableau, Sigma, Omni, Oracle BI and ThoughtSpot, which matters because it means the agent sits alongside the business intelligence stack rather than asking anyone to replace it.

Administrators install it and control access through Workspace settings, where individual data-source plugins can be enabled, configured and restricted. That is the right shape for anything touching a warehouse, and it is worth checking who in your organisation holds that switch.

Why this is a governance story before it is a productivity story

Self-service analytics has been promised for twenty years and has failed for the same reason every time. The tool was never the bottleneck. The bottleneck was that answering a business question correctly requires knowing which table is authoritative, which metric definition is current, and which joins quietly double-count.

A natural language agent removes the syntax barrier completely. It does not remove the semantics barrier at all.

So the risk is not that people cannot get answers. It is that everyone can now get an answer quickly, confidently and in a shareable dashboard, from a data model that may not mean what they assume. Wrong numbers used to be slow and rare because producing them took SQL. They are about to become fast and plentiful.

That is not an argument against the Data agent. It is an argument for doing the thing most organisations have deferred: writing down what your metrics actually mean, which tables are canonical, and which are staging. The agent will be exactly as trustworthy as the semantic layer underneath it.

What it changes for marketing measurement specifically

Three effects worth planning for.

Attribution arguments get faster and louder. Anyone can now produce a chart that appears to prove a channel does or does not work. Whoever brings the dashboard to the meeting sets the frame, and the person who has to unpick a flawed join is always slower than the person who generated it.

This is why correlation-based reporting is a weak foundation for budget decisions. If a channel refuses to show up in last-click, more dashboards will not rescue it. The answer is experimental, which is why geo-incrementality testing matters more as query tools get easier, not less.

The definition of a metric becomes a political object. When one team's agent says organic conversions rose and another's says they fell, the disagreement is almost never the data. It is two definitions of conversion, both defensible, neither written down. Publish the definitions before the agent gets rolled out, not after the first contradictory dashboard.

Your own analysis stops being scarce. If your value to the business was pulling numbers, that value just fell sharply. If your value was knowing which numbers matter and why, it just rose. That shift is uncomfortable and it is worth naming honestly in your team rather than letting people discover it in a reorganisation.

The question to ask before you switch it on

One question separates a useful rollout from a mess: can you name, today, the authoritative source for your top five business metrics?

If the answer is yes, and it is documented somewhere a newcomer could find, the Data agent will make your organisation faster.

If the answer is "it depends who you ask", you are about to industrialise that ambiguity. Fix the definitions first. It is a two-week job that most teams have postponed for two years, and this release is the forcing function that finally justifies it.

What I would not conclude from this

Two things the announcement does not establish.

It does not establish accuracy. OpenAI has published no error rate for the agent's analyses, and "investigates what changed in a metric" is a task where being confidently wrong is easy and hard to spot. Until there is a published benchmark, treat every output as a hypothesis that needs checking against a known-good query.

It also does not establish that this replaces your BI tool. The integration list points the other way: it reads your dashboards rather than replacing them. Anyone using this release to argue for cancelling a BI contract is moving faster than the evidence.

What to do about the ChatGPT Data agent

Four steps, in order.

Find out who controls Workspace settings in your organisation and whether data-source plugins are already enabled. Many teams will discover this is live before they have discussed it.

Write down the canonical source and definition for your top ten metrics, and put it somewhere the agent's users will actually see.

Pick three questions you already know the correct answer to, ask the agent, and compare. That is your accuracy baseline, and it costs an afternoon.

Then decide what stays human. Some decisions should not be made from a dashboard generated in ninety seconds by someone who does not know how the table is built, and saying so early is easier than saying so after a bad call. The ChatGPT Data agent removes the effort from asking. It does not remove the responsibility for knowing whether the answer is right.