The Data Agent Is Coming for the Work Queue

OpenAI says its new Data agent in ChatGPT Work can connect to approved company data, investigate what changed, and build interactive dashboards from a conversation. The announcement names connections including BigQuery, Snowflake, Databricks, MongoDB, and Google Drive, and says the agent can use business terms, metric definitions, calculations, and relationships supplied through semantic layers and trusted sources.

That is a meaningful shift in the unit of office work. AI is moving beyond “summarize this report” toward “investigate why this metric changed, show me the evidence, and help me decide what to inspect next.”

Why this matters for real work

Managers often wait for an analyst or operations specialist to answer questions that are important but not necessarily complex: Why did pipeline slow down? Which accounts need attention? Where is spending rising? The delay is often not the arithmetic. It is finding the right data, agreeing on definitions, and turning the result into something other people can review.

A conversational data agent could reduce some of that coordination cost. It may help a manager start with a question, ask follow-ups, see a chart, and inspect the supporting evidence without learning a query language or a separate BI workflow.

But the agent does not remove the management work. Someone still has to decide which source is authoritative, whether “revenue” means the same thing across reports, whether the comparison is fair, and what action is justified. A faster answer to an undefined question is still undefined work—just wearing a nicer dashboard.

Cihan’s view: “The useful AI analyst is not the one that produces the prettiest chart; it is the one that makes the metric, source, and uncertainty impossible to ignore.”

What is still uncertain

The announcement is a vendor description, not independent evidence that the Data agent improves decisions or reduces total reporting effort. OpenAI also says that connected-account permissions continue to apply and that users can review evidence behind findings, but the quality of an analysis will still depend on data quality, access configuration, semantic definitions, and the reviewer’s judgment.

There is another practical limit: a dashboard can make a disputed definition look settled. Do not infer that a chart is correct merely because it is interactive, and do not treat a recommended next step as an approved business decision.

One practical action this week

Pick one recurring question your team asks every week. Before connecting any tool, write a one-page analysis brief:

Question: state the decision or investigation the analysis supports.
Definition: name the metric, time period, comparison, and source of truth.
Evidence: require the result to show the source tables or documents, assumptions, and missing data.
Review: name the person who will challenge the result before anyone acts on it.

Then test the smallest version with approved, non-sensitive data. Compare the complete job—finding inputs, checking definitions, interpreting the result, and communicating the decision—not just the time spent generating a chart.

TRY / SKIP / USE

TRY: one bounded analysis with approved sources, explicit metric definitions, citations or evidence, and a named reviewer.
SKIP: connecting every company system before you know which question, definition, and owner matter.
USE: conversational analysis for first-pass investigation and follow-up questions, while keeping consequential decisions with the responsible person.

For one practical AI-for-work task each week, with a clear TRY / SKIP / USE verdict, subscribe to The Weekly Verdict.