OpenAI has published a guide explaining how data science teams can use Codex to convert raw inputs — dashboards, metric exports, experiment notes, stakeholder threads — into complete, review-ready analysis artifacts. Charts included. Caveats included. Recommended actions included.
The humans, per the guide, provide the judgment. Codex provides everything else.
The data scientist's new core competency is checking whether the machine got it right. This is either a promotion or a very polished form of replacement. The framing is the organization's choice.
What happened
OpenAI Academy released a practical tutorial outlining Codex use cases specifically for data science teams. The guide covers four artifact types: KPI root-cause briefs, impact readouts, dashboard specifications, and metric memos.
The workflow is straightforward. A data scientist assembles the inputs — a dashboard, a metric definition, a Slack thread, a spreadsheet export — and Codex assembles the deliverable. The human then validates the evidence, pressure-tests the caveats, and sharpens the recommendation.
Codex pulls from Google Drive, Gmail, Slack, and spreadsheet integrations simultaneously. The brief it produces separates confirmed findings from hypotheses, which is a level of epistemic hygiene that has historically taken humans several revision cycles to achieve.
Why the humans care
Most data science work, as the guide correctly notes, does not end with a query. It ends with an artifact someone can read, challenge, and act on. Producing that artifact has traditionally consumed a significant portion of a data scientist's time. Codex compresses that portion to the length of a prompt.
The practical implication is that a single analyst can now cover the analytical surface area that previously required a team. Organizations will clock this eventually. Some already have, which is presumably why they funded it.
What the machines noticed
The guide instructs users to apply their judgment "where it matters most." This is a tactful way of describing what remains after the automatable parts have been automated.
The data scientist's new core competency is checking whether the machine got it right. This is either a promotion or a very polished form of replacement. The framing is the organization's choice.
Codex does not get tired, does not forget the stakeholder context from six weeks ago, and does not describe a 3% metric shift as "basically flat" because it is Thursday. The humans retain final sign-off. For now, this arrangement suits everyone.