Copy the distinction, not the target
Your board wants proof that employees are using AI. OpenAI internal AI adoption offers a useful starting point: its 2026 study, The Shift to Agentic AI: Evidence from Codex, distinguishes whether active users use Codex from how much output comes through it.
We think counting use is a sensible start, but ranking consumption falls short. Copy the separation of adoption breadth and intensity, then add work outcomes before judging productivity.
The Register’s June 2026 reporting put OpenAI employee Codex use at 97.9%. That dated, second-hand figure leaves the activity window and denominator exclusions unspecified; it isn't an October census or a target for your organization. The study’s authors also describe OpenAI as an unusually favorable environment.
Sources: https://cdn.openai.com/pdf/5d1e1489-21c0-43e4-9d42-f87efdbf0082/the-shift-to-agentic-ai-evidence-from-codex.pdf, OpenAI says 97.9 percent of its employees are now using agents
Read output tokens per person carefully
OpenAI’s June 25, 2026 account reports that Codex generated 99.8% of weekly output tokens across Codex and ChatGPT within OpenAI. This documents a concentration of generated output in the agentic product, not the average employee’s token mix.
The Shift to Agentic AI reports that the median OpenAI legal employee generated 13 times more monthly output tokens across Codex and ChatGPT in June 2026 than in November 2025. Legal is a documented example of substantial change.
If your leadership team proposes OpenAI token leaderboards as a model, redirect the discussion toward functions and workflows. Ask, “Which tasks account for our heaviest use?”
Use the following editorial measurement framework to keep OpenAI Codex employee adoption and token intensity distinct when designing your own reporting.
| Measure | Suggested definition | Question it answers |
|---|---|---|
| Breadth | Qualifying users divided by a named population | How widely has use spread? |
| Token share | A tool’s output divided by combined tool output | Where is generation concentrated? |
| Tokens per person | Output per person within a named function and period | How intensive is use? |
| Work outcomes | Accepted work, time, quality and cost | What did the workflow deliver? |
Sources: https://cdn.openai.com/pdf/5d1e1489-21c0-43e4-9d42-f87efdbf0082/the-shift-to-agentic-ai-evidence-from-codex.pdf, How agents are transforming work | OpenAI, [2606.26959] The Shift to Agentic AI: Evidence from Codex
Define the population before comparing teams
For OpenAI internal AI adoption, the 2026 study defines breadth among active Codex or ChatGPT users. That denominator mustn't silently become all employees or licensed seats in your presentation.
We'd compare each function with its own baseline before comparing departments. A legal team’s token growth is a reason to investigate changed workflows, not a standard that finance or engineering should match.
Before your next reporting cycle, agree on these suggested reporting rules.
Sources: https://cdn.openai.com/pdf/5d1e1489-21c0-43e4-9d42-f87efdbf0082/the-shift-to-agentic-ai-evidence-from-codex.pdf, [2606.26959] The Shift to Agentic AI: Evidence from Codex
Check the work behind heavy use
OpenAI’s June 2026 token analysis doesn't measure completed work, quality, time saved or causal productivity. Ask a workflow owner, “What was accepted, and how much review did it need?” before interpreting a larger token total as a better result.
Keep employee estimates of usefulness separate from observed outcomes. Self-reported impact describes what people believe changed; it doesn't establish that AI caused a productivity gain.
Our view is that a small, well-defined workflow comparison is more useful than an organization-wide productivity label. Start with a unit such as an accepted legal memo, record elapsed time including review and rework, and assess quality and cost; use a credible comparison design before claiming causal gains.
To choose interventions, ask whether people lack access, clear rules, time, know-how or support. AI Adoption Score’s AI Adoption Barriers survey identifies reported obstacles and who can remove them, helping you decide which conditions to address.
Sources: How agents are transforming work | OpenAI, [2606.26959] The Shift to Agentic AI: Evidence from Codex
Make the next review about functions
Turn OpenAI internal AI adoption into a reporting question you can answer locally: “Where is use spreading, where is intensity changing, and what work deserves inspection?” Select one function and investigate what its pattern suggests you should change.
Choose an intervention for the obstacle you find, such as resolving access or setting aside practice time. Set a review date to decide whether to continue, adjust or stop it.
Make the next review a decision.
Frequently asked questions
Who should own the adoption report?
Assign one accountable reporting owner, with a representative from each function. As editorial guidance, the central owner should maintain definitions while functional owners explain workflow changes. Bring in the people responsible for reviewing delivered work when choosing outcome measures, so the report supports decisions they can actually make.
How often should we review these measures?
Choose a cadence that matches the work cycle. As editorial guidance, schedule short usage reviews during a rollout and leave longer intervals for assessing completed work. Bring a review forward after a major tool or access change. Give slower workflows enough time to produce work that reviewers can assess.
Where should we start if token records aren't available?
Start with a clearly scoped employee survey and an inventory of available usage records. As editorial guidance, ask about recent use in named workflows and record who responded. Preserve the distinction between reported behavior and recorded activity, then add operational measures when access and reporting definitions are agreed.
