The Supervision Gap: Why Manager AI Training Starts With Practice

Your youngest employees send up to 9 more messages a week to AI than the average person in your company. Your managers, directors, and executives send fewer.

That finding comes from OpenAI’s new study, How Organizations Use AI: Evidence from ChatGPT, which analysed 17.4 million real workplace messages across 1,764 companies, measured six months after each company adopted ChatGPT Enterprise. Because it draws on actual usage records rather than self-reported surveys, it is one of the clearest pictures yet of how organisations use AI day to day.

The Numbers Behind the Supervision Gap

The researchers describe a “negative seniority gradient”: an empirical pattern where the intensity of AI use falls as seniority rises. Early-career workers and trainees are the most intensive users in the data, while managers, directors, and executives send the fewest messages.

The split is striking. Six months after adoption, managers and directors make up roughly 24% of weekly active users, the largest block by headcount. Executives account for about 10%.

Yet when the researchers measured message intensity rather than headcount, the order inverted: the groups with the most seats were sending the fewest messages per person.

At Radiant Institute, we call this the Supervision Gap:

The people guiding AI-assisted work are barely using the tools themselves.

The Gap is Compounding

Usage is not standing still while organisations deliberate. The study found aggregate output tokens grew sevenfold between June 2025 and March 2026, and grew roughly fourfold within companies that had already adopted by June 2025.

But here’s the thing: AI use is deepening with or without a plan.

Look at what the tool is being used for. More than half of active users (56.3%) use AI for documentation and technical writing, and nearly half (49.6%) for technical digital work. This is precisely the category of work that lands on a manager’s desk for review.

And the work is not confined to technical teams.

Engineering and technical practitioners account for only 11% of weekly active users; the rest is spread across executives (9%), finance (5%), marketing (5%), and sales (4%). Every department’s output now carries an AI-assisted layer that someone senior must evaluate.

The paper’s own warning frames the stakes:

“Initial adoption does not imply effective deployment: realizing value requires experimentation, complementary investment, and organizational change.”

The Gap Runs One Level Deeper Than We Thought

We have written before about the adoption gap at the executive level, where leaders mandate AI but rarely use it. The OpenAI evidence shows the same pattern one level down, inside the manager layer where daily work actually gets reviewed, corrected, and approved.

Last week we described five gaps AI is revealing at work, the AI Hall of Mirrors. This study points to a sixth.

There is an honest caveat in the research worth holding onto. Messages measure activity, not quality. A junior sending 40 messages a week is not automatically producing better work, and a manager sending none cannot tell the difference. That is exactly the problem.

What Bruce Lee Rebuilt After Winning

In 1964, Bruce Lee won a private challenge match against Wong Jack Man, a master of traditional kung fu.

However, the victory troubled him.

The fight ran far longer than he expected, and he finished it exhausted, feeling that techniques he had drilled for years had nearly failed him under pressure.

His response was unusual for a man who had just won. He admitted his training had failed him and rebuilt everything from the fundamentals up. That rebuild became Jeet Kune Do, and the Bruce Lee the world remembers.

He later distilled the lesson into one line:

“I fear not the man who has practiced 10,000 kicks once, but I fear the man who has practiced one kick 10,000 times.”

Bruce Lee’s lesson can be applied for managers in the AI era. Just as you cannot coach a kick you have never thrown, and you cannot guide AI-assisted work you have never done yourself.

What Good Manager AI Training Includes

Closing the Supervision Gap takes a different kind of training for a different kind of role.

Three elements matter most:

1. Their own reps. Managers need hands-on practice on real tasks from their real week, because fluency follows usage, not seniority.
2. Live judgment checks. Short, verbal tests that reveal whether someone owns the reasoning behind AI-assisted work. Four questions cover most situations: What changes if the budget halves? What did you rule out? What evidence supports this? What would change your recommendation?
3. Coaching conversations. Managers are the translation layer between frontline experimentation and organisational outcomes. Training should teach them to turn a junior’s solo AI experiment into a team standard, with quality bars attached.

Ending Thought:

The OpenAI study confirms something many leaders have sensed but could not prove: AI use is concentrated at the bottom of the org chart, while review authority sits higher up. Left alone, that gap compounds, because usage is deepening fourfold inside the companies that have already adopted.

The good news is that the fix is refreshingly human. What managers need is unglamorous: reps, a way to test judgment, and the habit of coaching the work. No engineering degree required. That is a training design problem, and it is solvable.

If your organisation is working on the layer between the people who use AI and the people accountable for its output, Radiant Institute designs manager-focused AI enablement programs for exactly that. Reach out to explore how we can help.