The official AI training is scheduled for next quarter. It will not matter much. By the time it runs, the real curriculum will already have been taught, informally, at desks and in chat threads, by the three or four people in each team who started tinkering months ago and can't stop showing colleagues what they found.
This is how AI skill actually moves through an organization: not down the hierarchy but across it and up it, from the curious to everyone else. The pattern is visible in the data. When 78% of AI users bring their own tools to work, as the 2024 Work Trend Index found, they are not waiting for a program; they are running their own, and pulling neighbors along.
The best workers become the curriculum
There is a striking finding buried in the research on AI and productivity. In a study of more than 5,000 support agents, generative AI raised output most for novices, up to roughly 34%, and almost not at all for the most experienced. The mechanism was diffusion: the AI captured the patterns of the highest performers and made them available to everyone else. The system turned the tacit expertise that used to live in a few heads into something a beginner could reach.
The same shape appears with software developers. In the GitHub Copilot experiments, developers finished a coding task about 55.8% faster with the tool, and follow-up trials found the largest gains among the less experienced. In both cases AI works as a diffusion engine, compressing the distance between the best and the rest. But the tool only diffuses expertise once someone brings it into the room. The first mover is human.
Adults coordinate without being told
As Ricardo Semler has told it, Semco let factory workers set their own shifts in 1985. Management predicted the assembly line would collapse because the stations would never line up. Instead the workers coordinated among themselves, parents taking mornings, single people taking nights, with a single instruction from above: talk to the person next to you. The line held. The lesson was that adults, given a shared goal and the freedom to arrange the details, organize themselves better than a plan imposed from the center.
Bottom-up AI diffusion runs on the same principle. The person who figured out how to make the model draft a decent first-pass report will teach the colleague beside them if the environment lets them. No rollout deck teaches faster than a peer saying "watch, do it like this." The manager's job is less to be the teacher than to make sure the teaching that is already happening is allowed to happen in the open.
What leaders get wrong
Two mistakes stall diffusion. The first is insisting all learning flow through an official channel, which starves the fast, sideways transfer that does the real work and makes early adopters feel like rule-breakers. The second is ignoring diffusion entirely and assuming a license is a skill, which leaves the curious carrying the whole company on unpaid, invisible effort until they burn out or leave.
The fix is neither mandate nor neglect. It is to find the people already teaching and give them room and recognition: time to share, a channel to post what works, credit for lifting the team. The World Economic Forum's 2025 survey found 77% of employers planning to upskill their workforce; the ones who succeed will mostly be amplifying the diffusion that started without them, not replacing it with a program.
There is a further reason to let skill travel sideways rather than down. The people best placed to teach a tool are rarely the most senior; in the support-center study the veterans gained least from AI and the novices most, which means the natural teachers are often junior, curious, and low in the hierarchy. A rollout that only lets learning flow from the top down inverts the actual expertise, putting the people who understand the tool least in charge of teaching it. Following the current means being willing to learn from whoever figured it out first, whatever their title.
Follow the current
AI literacy has a current, and it runs from the curious outward, not from the top down. Organizations can fight that current with rigid, centralized training, or they can ride it by naming their early adopters, protecting their time, and letting the person next to you do what people have always done best: show a colleague how it's done.
Sources
- Brynjolfsson, Li & Raymond (2025). Generative AI at Work (Quarterly Journal of Economics). Field study of 5,172 support agents: AI raised resolutions per hour ~15% on average, up to ~34% for novices and near zero for the most experienced. https://academic.oup.com/qje/article/140/2/889/7990658
- Peng, Kalliamvakou, Cihon & Demirer (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (arXiv 2302.06590). Developers with Copilot finished a coding task ~55.8% faster than the control group; later RCTs replicate larger gains for less-experienced developers. https://arxiv.org/abs/2302.06590
- Microsoft & LinkedIn (2024). 2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part. 75% of knowledge workers use generative AI at work and 78% of AI users bring their own tools (BYOAI); survey of 31,000 across 31 countries. https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part
- World Economic Forum (2025). Future of Jobs Report 2025. 63% of employers name the skills gap as the top barrier to transformation; nearly 40% of required skills will change by 2030 and 77% of employers plan to upskill their workforce. https://www.weforum.org/publications/the-future-of-jobs-report-2025/