Guided progressive mastery: what if AI should make you better, not just faster?

A simple question

Is AI really making you better at your job? Or do you just think so?

Almost everyone can point to a tool that saves them time. Far fewer people could say, with any real confidence, whether they've become more capable because of it, or just more dependent on it. Answering that honestly starts with understanding what these tools are actually built to do.

What AI solutions actually do today

Work AI splits into two families. Augmented tools assist, but you still make the call, a chatbot that helps you draft something, a copilot that suggests a formula. Automated tools take a task from start to finish, and you're no longer part of the decision at all.

The problem is how success gets measured. For augmented tools, what actually gets tracked day to day is usage and churn, not whether the person got any better. Automated workflows changed that in one specific way: because they take on an entire task instead of a single action, it becomes possible to prove a real return on that investment, by comparing how long the task took before and after. That's genuine progress. But it has a blind spot. It proves a task got done faster, never that the person doing it grew. And a skill that stops being practiced doesn't stay intact, it fades.

The real cost, invisible in the dashboards

This cost has a name in the automation literature, deskilling. The mechanism was described as early as 1983 by researcher Lisanne Bainbridge, in a paper that's still a reference point in the field today (Bainbridge, 1983). Her argument: automate routine work and leave only the exceptions to the human, and you strip away the practice needed to keep judgment sharp, so the person is left unprepared exactly when an exception shows up.

A recent study from Microsoft Research and Carnegie Mellon measured this directly in knowledge workers using generative AI (Lee et al., 2025). The pattern held across hundreds of real work examples, higher confidence in the AI's output meant less critical thinking applied to it, even on tasks where accuracy mattered.

This isn't hypothetical. A 2026 survey of 2,400 executives and employees by Writer found that a large majority of senior leaders are already deliberately building an "AI elite" inside their teams, while 60 percent are planning to lay off employees who don't adopt these tools (Writer, 2026). On the employee side, a comparable share say they're genuinely afraid of losing their job over this shift. The real divide isn't between people who use AI and people who don't. It's between people who are actually getting better because of it, and people who are quietly becoming easier to replace.

What the research already says about a better way

None of this starts from zero. The idea of an AI that guides without doing the work for you has existed, in several forms, for decades.

In the 1990s, researchers John Anderson and Kenneth Koedinger built cognitive tutors, systems that track a student's reasoning step by step, compare it against an expert model, and step in only when the student drifts off course, offering a hint rather than the answer (Koedinger and Corbett, Carnegie Mellon University). A second technique, knowledge tracing, goes further, continuously estimating how well the student has actually mastered a given skill instead of just counting how often they show up. That's exactly the kind of metric missing from most workplace AI tools today.

Psychologist Albert Bandura described something remarkably close back in 1993, under the name guided mastery, walking someone through progressively harder tasks while gradually withdrawing assistance as their real competence grows (Bandura, 1993). It's almost exactly the mechanism a good AI tool should reproduce.

Proof that none of this is wishful thinking, OpenAI launched Study Mode in July 2025, built with pedagogy experts from more than 40 educational institutions (OpenAI, 2025). Instead of handing over the answer, it asks what the learner already knows and offers hints toward the solution, giving more support to people who are struggling and more challenge to those moving fast.

Guided progressive mastery: the principle

This is the logic I think should apply to any AI tool used at work, under the name guided progressive mastery.

The principle is simple. An AI that watches how you actually work, and only steps in when it's genuinely needed. Rather than carrying out the action for you, it walks you through the steps that build real understanding, do, see the problem, search why, understand, correct. The system knows the best possible path to the outcome, but it calibrates how much it intervenes to your actual level of knowledge, and it withdraws that assistance as you improve, exactly the pattern Bandura described.

How this could actually work

Technically, this is no longer science fiction. A new generation of software agents can now watch a screen and its surrounding context continuously, instead of waiting for someone to type a command, and they're becoming a standard building block of enterprise software. Gartner expects 40 percent of enterprise applications to include task-specific agents by the end of this year, up from less than 5 percent in 2025 (Gartner, 2025). The listening half of the system is no longer theoretical.

What's left is the measurement half. That's where knowledge tracing becomes genuinely useful. Instead of tracking how often you open a tool, you could track how your actual mastery of a task evolves, week over week.

The win on both sides

This framing answers two concerns that used to seem incompatible. For the person using the tool, real and measurable growth, an actual skill built, a career that keeps moving forward. For the business, performance it can finally prove with solid data, instead of a culture built around the fear of layoffs.

What shouldn't be underestimated

Two limitations deserve to be said plainly rather than glossed over.

The first is a genuine engineering problem. Classic cognitive tutors only worked because researchers spent years hand-modeling the expert path through one narrow, well-defined domain, like algebra. Building that same model for open-ended, wildly varied business software is a much bigger problem than anything cognitive tutors ever solved.

The second touches trust directly. A system that watches everything you do, and generates data about your progress, can very easily be read as a surveillance tool serving management rather than a coach serving you. If that distinction isn't made explicit at the design stage, the tool recreates the exact fear it was supposed to fix.

Neither of those problems gets solved this year, which is exactly why it's worth knowing you don't have to wait for them to be.

You don't have to wait for someone to build this

Here's a quick way to check where you actually stand today. Next time you use an AI tool to get something done, a chatbot, a copilot, an agent, doesn't matter which, ask yourself right after. If this tool disappeared tomorrow, would you be less capable than before, exactly the same, or better?

Researchers have a name for the failure mode here, the AI ghost learner effect, succeeding at a task with AI's help while never noticing you didn't actually learn anything from it (Rausch, Frontiers in Organizational Psychology, 2025). Most people never ask themselves this, because the output itself, the email sent, the report finished, the code that runs, feels enough like progress on its own.

Noticing the problem isn't the same as fixing it. And you don't have to wait for a vendor to ship guided progressive mastery as a feature to start getting it. You can build it into how you prompt, starting with your very next message.

Nearly every popular framework people use to write good prompts is built to get you a better output, a clearer answer, a cleaner draft, a more useful result (Fello AI, 2026). None of them ask what you should walk away knowing. But the pattern behind OpenAI's Study Mode, and a related technique called model guided prompting, explicitly telling the AI to ask you what it needs rather than assume, both point to the same fix (Educative). State your intent to learn in the first message, not as an afterthought once you're unhappy with the answer.

Concretely, this means adding one clause to whatever you were already about to ask, regardless of which framework you use.

Before you give me the final result, tell me the one part of this where getting it right actually depends on understanding something, not just execution. Ask me what I already think about that part. Give me a hint there instead of the full answer, and only give me the complete answer if I'm still stuck after trying. When we're done, tell me in one line what I now know that I didn't before.

That's the whole change. Paste it onto the end of any request, to any chatbot, and the interaction stops being a pure output machine. It starts doing, in miniature, exactly what this piece has been arguing for, do, see the problem, search why, understand, correct. You don't need a company to build guided progressive mastery into a product for you to start living it. You can prompt your way into it today.

My conviction

I don't think the right goal for workplace AI is to choose between making you faster or replacing you. I think it should exist for one reason, making sure you never stop getting better. I hope future AI solutions take this seriously, and make it their guiding star.