Most AI training isn't working. Here's what the research tells us and what Vinitra Swamy shared in our May 28 webinar.
There is no shortage of AI courses right now. Platforms, providers, and internal L&D teams are all racing to upskill their workforces in time to matter. And yet, completion rates are low, transfer to real work is minimal, and nobody is quite sure whether any of it is sticking.
This is not a content problem. It is a learning design problem.
The real issue with generic AI training
The dominant approach to AI upskilling today is essentially the same as it was for digital transformation ten years ago: identify a topic, find a course, assign it to everyone, measure completions. The subject matter has changed. The approach has not.
What learning science has known for decades – and what a growing body of research on AI in education is now confirming – is that this approach was never particularly effective for durable skill development. It is even less effective when the skill in question is something as contextual and fast-moving as working with AI.
Generic AI courses teach people about AI. They do not teach people how to use AI in their specific role, with their specific tools, on the tasks they actually do every day. That distinction sounds simple. Its implications for learning design are profound.
What the research says about skill transfer
Skill transfer - the ability to apply learning in a new context – is one of the most studied topics in educational psychology. The consistent finding across decades of research is that transfer is dramatically higher when learning is situated: when it happens close to the context in which the skill will be used, when it is built around real tasks rather than abstract concepts, and when learners receive feedback that is specific to their performance rather than generic.
Role-specific, task-based learning is not a new idea. What is new is the ability to deliver it at scale – to personalise a learning journey not just at the content level, but at the level of role, task, pace, and prior knowledge, for every individual in an organisation simultaneously. This is what agentic AI makes possible. And this is what most current AI upskilling programmes are not yet doing.
Why this matters now for L&D leaders
The organisations that will build genuine AI capability over the next two to three years are not necessarily the ones spending the most on training. They are the ones designing learning that is closest to work – that meets people where they are, builds on what they already know, and develops the specific skills their roles actually require.
For HR and L&D professionals, this is both a challenge and an opportunity. It requires letting go of the course-completion model as the primary measure of success, and building towards a model where learning is embedded in the flow of work, personalised at the individual level, and continuously adapted as both AI capabilities and organisational needs evolve.
What Vinitra Swamy shared in the session
On May 28, LearnChamp hosted Vinitra Swamy, co-founder and CEO of Scholé and one of the first PhD graduates from EPFL's Machine Learning for Education Lab, for a live conversation on what AI-native learning looks like in practice.
Among the most striking insights from the session: dropout rates on platforms like Coursera and edX sit at around 90%. As Vinitra put it, these platforms are "not meeting learners exactly where they are and what they need." The solution is not a chatbot added to the side of a course but a fundamentally different approach to how learning is designed and delivered.
In Scholé's first pilots, 100% of learners on the AI mastery use case reported a new adoption of AI that saved them more than three hours a week. That is the outcome AI-native learning is designed to produce.
Watch the full session
The recording covers the science, a live demo of Scholé's agentic learning engine, and a Q&A with Vinitra. It is free to watch and runs just over an hour.
Watch the recording:
