The quick read
- More messages, tokens or AI-assisted code do not by themselves prove better results.
- Choose one recurring task and compare complete effort before and after the change.
What the tools show
OpenAI’s September 16 post describes analytics across ChatGPT Work and Codex, including usage, costs, task categories and engineering outcome indicators. It explains how administrators can inspect adoption and how business owners can relate those observations to the work their teams perform.
Activity is not an outcome
More messages, tokens or AI-assisted code do not by themselves prove better results. A team may be experimenting, correcting mistakes or handling more difficult work. Meaningful evaluation needs a starting point and measures such as accepted output, delivery time, defects or rework.
A useful measurement plan
Choose one recurring task and compare complete effort before and after the change. Include setup, review and ongoing support. Treat hypothetical return calculations as illustrations, and keep actual measured outcomes separate from estimates of what saved capacity might be worth.
Sources & notes
AI-assisted editorial content checked against the linked sources.
openai.com — official reference
Sources reviewed for the September 2026 launch edition.