This week’s AI training principle is “Invest more in the people who are already moving.”

I had an entire article drafted about how investing in your champions, super-users, and early adopters is the way to get results quickly. I still believe that, with a catch.

Part of my process every week is to form a hypothesis on training based on what I’ve seen in organizations, then test it against data and research. This week, the data pushed back.

I have seen first-hand how much adoption happens from a single super-user showing their peers how they are using AI. Those demos accelerate usage and help connect the dots for those struggling to understand how AI applies to their own workflows.

Organizations need to identify their super users, make sure they have access to the training and upskilling they need, and create a path for them to share their learning with their organization.

Sounds like a good principle, and it is. But the data and research point to some interesting nuances.

I couldn’t find a study that tests champions against universal training head to head. The closest research points to needing both.

Early adopters are disproportionately people who already have more access and advantage, and the clearest evidence is economic. 

In the New York Fed’s November 2025 survey, 66 percent of workers earning over $200,000 used AI in their job, compared with 16 percent of those earning under $50,000. College graduates were more than twice as likely to use it as workers without a degree (59 percent versus 23 percent). Only 16 percent of workers said their employer offers any AI training. The Fed’s economists point out that some of the workers who value training most, including those without a college degree, have the lowest use and the least access to it.¹

Gender shows up too. In Denmark, women were 16 percentage points less likely than men in the same occupations to have used ChatGPT for work.² Where employers encouraged use, provided tools, and offered training, that gap nearly disappeared, according to the researchers’ 2025 working paper.³ Race is less clear, so I gave it its own note below^a.

I couldn’t find a study that measures who gets named a champion, but champions usually come from the early users, and early users skew toward people with more income, more education, and more room to experiment. If your champions emerged on their own, expect them to look like that.

Early research also shows that novices gain the most when AI is built into a structured workflow. In a study of more than 5,000 customer support agents, the AI tool raised novice and low-skilled workers’ productivity by 34 percent, compared with 14 percent on average.⁴

Yes, people say they learn AI from their peers. But if none of your peers are using AI, you’ll never catch up.

We need both. We can support our early adopters and super users while also making sure the technology and training required to become a super-user is widely available and accessible to everyone.

A few guiding principles to follow:

AI is not going to be a one-shot for training. The concepts keep expanding and existing tools keep changing.

And when your approach is wrong, or new data supports a different way of thinking, take the lesson from this article: do better and try again.

Notes

a. I also looked at race. In Pew’s February 2026 survey, employed Black, Hispanic, and White adults used AI chatbots for work tasks at nearly identical rates (36 percent, 37 percent, and 36 percent), while Asian workers were far ahead at 60 percent. Pew didn’t ask about employer-provided tools or training in that survey, so access and training by race is still an open question. The few other studies that break out race, including a large multi-wave worker survey and a Jobs for the Future survey focused on Black workers and learners, don’t show Black or Hispanic workers lagging on use, and JFF found Black workers more likely than respondents overall to report employer AI training. What surprised me is how rarely race shows up at all. The Federal Reserve’s household survey, the New York Fed’s survey, and the St. Louis Fed’s adoption tracker all collect race, but they mostly publish their AI results by education, income, age, or occupation. The one exception I found: the New York Fed noted that non-white workers were willing to pay more for AI training. Gender, by contrast, has its own research base, including a review of 76 studies. I’d like to see these trackers publish race breakdowns, especially for employer access and training, because that’s where gaps could still be hiding.

This week I talked with Mike Nakhimovich, who led Airbnb’s Dev AI team through a no-mandate rollout that, he says, took agentic coding from 5 percent to 98 percent adoption. He’s now building Hive, a place to co-work with your agents that feels more like a farm than a factory. We get into where to keep humans in the loop and why 50 to 80 percent faster “in a happy way” beats 4X.

Listen here

I vibe-coded my most recent conference talk. Not a PowerPoint, not a slide deck, but a live website I used to present. It had a presentation mode I could toggle on for easier navigation, plus fallbacks for every interactive element in case the wifi went down. I’m happy with how it turned out and will share more soon. Here’s the take-home one pager I gave participants.

When AI first clicked for you at work, did it come from someone you work with, or from somewhere else (a course, a video, trying it yourself)? Hit reply and tell me which. I’d love to hear from you.

Molly