There is a particular kind of learning that happens when something breaks and you need to fix it. This month I needed to automate a workflow, so I turned to AI and asked it to walk me through Power Automate. I broke it, pasted the error, moved forward, met my next hurdle and started the process again until it ran successfully.
When a structured course is designed and taught, someone has thought carefully about the order in which knowledge should arrive. The foundational concepts and context for understanding come first, so that the scaffolding can be used to create problem solvers and not parrots. That architecture is invisible when it works well. You move through a course and ideas click into place because someone spent time thinking about the sequence.
When you learn from AI, none of that exists. The curriculum that emerges is entirely reverse-engineered from your goal. You bring a goal, and the learning takes the exact shape of the problems that goal brings. You learn in the order difficulties arise, and not in the order that concepts are built. You encounter the knowledge you needed for this, and you never encounter the knowledge you did not know to ask for.
The consequence is a particular kind of competence gap. If you learn a tool through a structured course, you might spend the first module understanding how its core logic works before you ever build anything. You would find it slow. But when something breaks six months later in a way you have never seen, you would have the foundational knowledge to reason through it.
If you learn the same tool through AI, you learn each component the moment you need it. Your knowledge is immediately applicable and completely task-shaped. You are competent at your specific flow, until it breaks in a new way, and then you are standing in front of a problem you do not have the architecture to reason through, because you never built the architecture. It is like learning chess by studying specific plays without ever understanding how each piece moves independently. You can follow a script, but you cannot play a new game.
That critique, while real, is also incomplete. Most traditional curricula are badly designed for actual application. The theory-first model has its own profound failures. People complete courses and cannot do the work. People sit through lectures and retain almost nothing because nothing was at stake.
There is genuine value in problem-first learning. When you need the answer because something is broken and you need to fix it, you are paying attention in a way that no amount of academic motivation quite replicates. AI-assisted learning is essentially applied learning by default.
All of this is to say that the experience got me thinking about the kind of knowledge I had acquired, and whether it would be compoundable. Would I be able to teach someone else how to build a Power Automate flow? And what was I missing out on learning because of the AI-generated tunnel vision?
AI gets you from zero to functional faster than almost anything else. It removes the friction of asking questions. It meets you exactly where you are and lets you learn at the pace your actual work demands.
What it does not do is teach you what you did not know to ask. It does not build the connective tissue between concepts, tell you when you are missing something foundational, or know what your skill level actually is, because it only knows what you have told it.
There is also a subtler problem. AI will explain a wrong answer with the same confidence it uses to explain a right one. A human teacher hesitates, qualifies, says they are not sure and will check. The frictionless confidence is part of what makes AI useful as a learning tool and part of what makes it worth being careful with.
Each problem you solve with AI is still a brick, even if you do not know where it fits yet. Enough of them, and you start to develop an instinct for the shape of the wall. The issue is that instinct alone is not the same as a blueprint. The people who get the most out of AI-assisted learning will be the ones who eventually couple it with something structural like a beginner course, a book, a colleague who can explain the underlying logic. Not to replace the applied learning, but to give it somewhere to land. Each AI-trained event covers ground. A short, deliberate pass through the basics afterwards lets you organise that ground into something you can build on and hand to someone else.
The conversation about AI and jobs is loud and it is real. The more interesting one, about what it means to actually know something in an era where you can become functional at almost anything in an afternoon, is the one worth sitting with longer. Understanding is what compounds.
There were a few moments, learning with the AI, where it felt like I was the one getting prompted.
← Back to Insights