Four out of five organizations are running AI training. Three out of five still report an AI skills gap. The training is happening; the gap isn’t moving. That isn’t a quality failure — it’s a targeting failure. This episode works the argument from chapter two of AI Empowered: the three dominant responses to the AI […] Below is the full script of this episode as an article.
Picture the person at your company who is furthest ahead on AI. Three certifications. Can explain how the models actually work. Ran the lunch-and-learn. Posted the thoughtful thing about responsible adoption and it did numbers. Everybody agrees she’s the one who gets it.
Now pull three of her deliverables from the last quarter and set them side by side.
They could have come from anyone with the same subscription.
That’s the puzzle at the center of chapter two. She isn’t undertrained. She is the training industry’s ideal graduate. She knows more about AI than almost everyone she works with… and she operates exactly like someone who has had no training at all. Type a request. Skim what comes back. Ship it.
Her knowledge about AI is real. Her relationship with AI never changed.
Three responses, and the wrong layer
There have been three big responses to this, and I want to be fair to all of them, because the people building them are not fools.
The first is education. Teach people how the models work, where they make things up, what they’re good at, and they’ll work with them well. The second is technique. Prompt engineering. Learn the right formulas, the right structure, the right way to ask. The third is process. Change management. Executive sponsor, phased rollout, champions in every department, a support channel nobody opens.
Each one addresses something real. Each one gets you more usage. And the honest assessment of all three, as most organizations actually run them, is that the work product comes out looking about the same as it did before.
Take the middle one for a second, because it’s the one people defend hardest. The prompting workshops aren’t wrong about what works. They’re wrong about why it works. Watch two people who sat through the identical workshop attack the identical problem. One runs the template. Role assignment, structured instructions, output format. She gets back something clean and professional that any competent person with the same subscription would have gotten. The other one brings something the template has no slot for. That this client’s board got burned on an acquisition three years ago. That the finance chief’s questions in the last meeting were budget anxiety wearing a strategy costume. Same template. Unrecognizable output. The difference was never the syntax. It was what the person put on the table that the tool could not have invented.
And the syntax expires anyway. The model updates, the formatting trick stops mattering, and the people who memorized templates are stranded while the people who brought expertise never noticed anything happened.
The book’s argument is that they’re not wrong about the problem. They’re wrong about the layer.
What the new research shows
Here’s the part I didn’t have when I wrote that chapter.
A team at Microsoft ran a field experiment with three hundred eighty-eight employees at a Fortune five hundred retailer. The paper came out this spring. Everybody in the study got the same AI tool. The only thing the researchers changed was the structure around how people were asked to work with it.
One group got what the paper calls behavioral scaffolding. A protocol. Work in pairs, go into the tool together, follow the steps. This is the process answer, executed carefully, by people who study this for a living.
It went worse. Lower document quality than the group with no structure at all, and substantially less work finished.
I love that finding, and not because I enjoy watching a protocol fail. I love it because it’s the cleanest version of the argument I’ve seen anywhere. The intervention was competent. The people were willing. And bolting process onto the outside of the collaboration made the collaboration worse.
The other group got cognitive scaffolding. Partnership training. The entire point of it was to move people from treating AI like a search box to treating it like something you think with. That group turned in higher quality work… at the top of the distribution.
That phrase matters, and I’m not going to skate past it. Not everyone improved. The best work got better. The paper also found that the trained group’s beliefs about AI moved in a more positive direction, and then the authors ran their sensitivity checks and said, plainly, that the shift probably reflects recovery from a carry-over effect rather than the training itself. They flag a morning-versus-afternoon session confound. Differential attrition. Grading sensitivity to document length. One company. One day.
So the paper settles nothing on its own. It points. And where it points is the layer.
Why training aims everywhere else
Why does most of the training industry keep aiming everywhere else, then?
Because the other layers are measurable and they fit on a calendar, right? You can put eighty people through a workshop on a Tuesday and report a number on Wednesday. Nobody has ever been promoted for reporting that their people’s mental models are unchanged.
DataCamp ran a survey this year with YouGov, more than five hundred enterprise leaders across the United States and the United Kingdom. Eighty-two percent said their organization offers some form of AI training. Fifty-nine percent said they still have an AI skills gap. Only thirty-five percent said their upskilling program is mature enough to call organization-wide.
Sit with the first two numbers together. Four out of five are running training. Three out of five are still short. The training is happening. The gap is not moving.
That is not a quality failure. Most of those programs are fine at what they set out to do. It’s a targeting failure.
And here’s the thing the rollout plan never accounts for. Reif, Larrick and Soll ran four preregistered experiments with more than four thousand four hundred people and found that using AI got you rated nine percent less competent by your peers. For women in engineering, thirteen percent. And when men who didn’t use AI themselves were the ones doing the evaluating, the penalty on women who did use it hit twenty-six percent.
Think about what that number does to a decision. A woman engineer is sitting there deciding whether to mention that she prototyped the solution with AI. That isn’t reluctance. That’s arithmetic. She’s weighing an efficiency gain against a career cost somebody went and measured. No communication plan survives that. No executive sponsor reverses it.
The exception
So I went looking for the exception.
I wanted to find the product that teaches this at the layer underneath. Not what AI is. Not how to phrase a request. How a person actually relates to the thing. Trust calibration. Knowing when to override. What you personally bring that the tool cannot generate on its own.
So I mapped the market, and here’s what’s in it.
There are the enterprise AI coaching platforms, the ones with real funding and real deployments at companies you’ve heard of. Those use AI to coach you on leadership and communication. You are what’s being coached. AI is just the delivery mechanism. Then there are the AI skills companies. Cohort programs, hundreds of thousands of people trained, genuinely good instruction. Those teach the tool. Then there are the simulation platforms, which honestly have the best format in the entire market. Practice, role-play, feedback, a safe room to be bad in. Pointed almost entirely at interpersonal skills.
Three crowded categories. And as far as I could find, nobody is selling the fourth one.
I’ll be honest about what that might mean. It’s either an opening or it’s a market that doesn’t want the thing. I don’t know yet, and I’d rather say that than pretend. But it lines up uncomfortably well with the pattern: the layer nobody is addressing is also the layer nobody has figured out how to package.
What to do on Monday
So what do you do with this on Monday.
If you’re the one who has to make AI work in your organization, the question to stop asking is “have our people been trained.” They probably have. The question to start asking is narrower and more uncomfortable. When somebody on your team opens an AI tool, what do they believe their job is in that exchange?
If the answer is “ask correctly and check what comes back,” you have an operator. Operators get generic work back, because they put generic work in. If the answer is “bring the thing the model can’t have” — this client’s history, the reason the last proposal died, the pattern you’ve watched four times that nobody has ever written down — you have a collaborator.
And that’s the part nobody can hand you in a workshop. It’s a belief about your own role.
You can’t put it on a slide. You can ask about it out loud, in a real conversation, which almost nobody does.
I run Auspicious, a fractional marketing practice, and here’s the version I watch for in my own work. Whether the first thing I bring to a session is a request… or something only I know. When it’s the request, I get back exactly what anyone else would get. Every single time.
It’s the same tool for everybody. Whatever is making the difference, it isn’t the tool.
—
This show extends the book AI Empowered. Read the first chapter free at https://www.aiempoweredbook.com, or get the book on Amazon.
Aaron Douglas runs Auspicious, a fractional marketing practice.
Blog: https://auspicious.llc/insights/
X: https://x.com/aaron_douglas
LinkedIn: https://www.linkedin.com/in/aaronddouglas/
