How People Use ChatGPT at Work: OpenAI Signals, Japan and a Practical Planning Exercise

ByMedia Spec Lab運営者
Two colleagues planning a community event together

日本語版を読む →

Checked November 4, 2026. What do people bring to ChatGPT at work besides writing? OpenAI Signals v3.0 classifies work-related messages from individual ChatGPT plans into 165 activities across nearly 100 countries. Japanese AI news outlet Ledge.ai reported on November 3 that planning and decision-making was the largest broad category in Japan.

This guide explains what the figures do and do not mean, then offers a fictional exercise for exploring an event plan. The prompt and examples below are teaching material created by MSL, not a recorded ChatGPT test.

What to know first

  • The data covers work-related messages from individual plans; enterprise messages and Codex use are excluded.
  • Japan’s figures describe classified messages, not the proportion of all Japanese workers using AI or a measure of productivity gains.
  • Check the definitions and classification method, and test any proposed workflow in your own context before adopting it.

What was released?

On October 26, OpenAI released quarterly data from the first quarter of 2025 through the third quarter of 2026. Activities such as preparing instructional material, developing marketing content and programming are mapped to intermediate work activities shared across occupations. The interactive charts let readers explore countries, activities and changes over time.

Across 74 countries observed in every quarter, the average share of each country’s classified work messages represented by its top ten activities fell from 56.0% in Q1 2025 to 51.1% in Q3 2026. This indicates a broader distribution across classified activities; it does not establish that AI has spread across every kind of work.

How to read the Japan figures

Ledge.ai reports that planning and decision-making was the largest broad category for Japan in Q3 2026. Its components included creative thinking at 17.0% and decision-making and problem-solving at 15.9%. The denominator is classified work-related messages. A person who asks several questions contributes several messages, so these percentages are not shares of individual users.

OpenAI says automated systems classify messages after identifying information is removed. It withholds small groups and adds statistical noise to published aggregates to protect privacy. Classification may be imperfect, and people who do not use ChatGPT are absent from the data.

Try it: plan a fictional event announcement

Instead of adopting AI simply because a use is common, test whether it gives you useful decision material. Imagine a free event on November 20, at a venue five minutes’ walk from a station. Capacity is undecided; the audience is local families.

First write down the decision: in this example, the order of the announcements and what must be confirmed before publishing. Try this prompt with fictional information:

You are helping prepare an announcement. Using only these facts, propose three ways to announce a family event. Facts: admission is free; the event is on November 20; the venue is a five-minute walk from the station; capacity is undecided. Audience: local families. For each option, give its aim, a headline and points that need confirmation. Do not guess the start time, venue name, registration method or capacity; mark them as unconfirmed. List the decisions a person must make before publication.

Check whether the response invents a time or venue, keeps capacity undecided, and offers alternatives that fit the audience and goal. A person must verify the facts and make the final publishing decision.

Record the comparison

Do the same task once without AI, then try it with AI. Record the total time including review and corrections, what you changed, and which questions remain. One response can vary, so repeat the exercise with different tasks before drawing a conclusion.

  • Consider adoption: you can verify the facts, the work remains useful after correction, and you have permission to enter the information.
  • Revise the approach: if the answer fills in unknown details, make the constraints and requested output clearer and try again.
  • Do not use it: if customer information cannot be entered, errors cannot be checked, or corrections take too much work.

Check the original data

  1. Read OpenAI’s announcement for the coverage period and definition of work-related messages.
  2. Select a country, period and activity in the charts, and check the unit and denominator.
  3. Read the data and methodology for small-group suppression and classification limitations.
  4. Compare reporting with the original data; do not confuse a share of messages with a population adoption rate or an outcome at work.

Continue reading and trying

Try a beginner’s announcement exercise · Read our audio upload test · Build a work prompt (Japanese interface) · Check an AI answer against source material (Japanese interface)

Sources

Similar Posts