What do data scientists spend 80% of their time doing? Not modeling, not analysis — writing reports, making slides, preparing data for business stakeholders.

OpenAI's newly released practical playbook for data science teams using ChatGPT Work (formerly Codex for Work) hits this pain point squarely. It shows how data teams use AI agents to automate five categories of the most time-consuming "adjacent work," letting data scientists spend time on things that actually require human judgment.

Five Automatable Data Work Categories

The guide lists five high-value scenarios that already work in practice — every one of them things data teams do daily but hate doing.

First: Root-cause briefs. The question business asks most: "Why did this metric drop?" Previously, a data scientist would spend half a day pulling data, running comparisons, making charts, writing conclusions. Now agents can automatically connect to data sources, run anomaly detection, slice dimensions, generate a root-cause brief with charts — and the data scientist just checks that conclusions hold, adds business context, and sends it off. The whole process goes from half a day to 15 minutes.

Second: Impact readouts. After running an A/B test or shipping a feature, how do you quantify business impact? Agents can automatically calculate core metric changes, statistical significance, segment differences, and produce a structured impact readout. The old headaches of "p-hacking" and "selective reporting" are also greatly reduced through standardized templates.

Third: KPI memos. Reporting business metrics to management weekly/monthly used to take a full day gathering data, building slides, writing summaries. Now agents automatically pull KPI data from the warehouse, compare historical trends, flag anomalies, generate a first-draft memo — analysts just add their own judgment and recommendations.

Fourth: Scoped analyses. Business throws in an ad-hoc analysis request ("Why is new user retention low in East China during 618?") — previously a 2-3 day queue item; now agents automatically determine analysis scope from the request, pull the right data, run basic analyses, output preliminary conclusions. Simple questions get answers same-day; analysts dig deeper only on complex ones.

Fifth: Dashboard specs. Building a new BI dashboard used to mean endless back-and-forth between PMs and data engineers trying to nail down requirements. Now stakeholders describe needs in natural language, agents auto-generate a complete dashboard spec document — which metrics, how they're defined, chart types, refresh rates, permission settings — data engineers just build to spec. Requirements communication costs drop 70%.

Not Replacing Analysts — Freeing Them from PowerPoint

Seeing "automation," many data scientists' first reaction is "AI is coming for my job." Not at all.

Look closely at these five scenarios: not a single one involves "building models," "doing deep statistical analysis," or "designing experimental design" — the work that actually requires data science expertise. They're all "translation" and "transport" work — translating data into language business understands, moving analytical conclusions from notebooks into slides/memos/dashboards. This isn't data scientists' core value — yet it consumes 80% of their time.

After automating this work, what can data scientists do? They can spend time on higher-value things: deeply understanding business problems, designing better experiments, building more accurate models, discussing strategy and action plans with stakeholders. In other words, AI turns data scientists from "report generators" into genuine "business strategists."

This is the right way for enterprise AI to land: not using AI to replace people, but using AI to free people from repetitive work so they can do things that genuinely require human intelligence, judgment, and creativity.

See you tomorrow.