TL;DR: For half a century, the science of expertise rested on one finding: people get dramatically better at a skill through deliberate practice, meaning repeated effort at the edge of their ability paired with immediate feedback. AI breaks the delivery mechanism for that practice. When the tool does the reps, you keep the output but lose the struggle that used to build the skill, and controlled studies already show AI raising a novice’s throughput by 34 percent without the novice necessarily getting better underneath. This guide gives you the Rep Allocation Matrix, a named CEOtudent framework for deciding which reps to keep by hand and which to safely offload, and a four-part protocol for rebuilding the deliberate practice loop on top of AI-assisted work. A CEO protects the assets that compound while automating the rest; a student keeps doing the hard reps on purpose, because that is where the growth still lives.
Anders Ericsson spent his career showing that expert performance is built, not born. His 1993 research established that the amount and quality of deliberate practice, not raw talent, best explained who reached elite levels in music and other demanding domains. The word deliberate is doing heavy lifting there. It is not repetition on autopilot. It is effortful, uncomfortable, feedback-rich practice aimed precisely at what you cannot yet do. That mechanism is now colliding with a technology whose entire value proposition is to remove effort and hand you the finished result.
The reps problem
Here is the tension in one sentence. Deliberate practice requires you to do the difficult thing repeatedly, and AI exists to do the difficult thing for you. Every time you accept an AI-generated draft, run an AI analysis you could not perform yourself, or ship a fix you did not understand, you complete the task and skip the practice. The work gets done. The learning does not happen. Over months, this produces a specific and dangerous profile: a person whose visible output keeps rising while their underlying capability quietly stalls or decays.
We already have a hint of this in the data. In the largest controlled study of AI at work, access to an AI assistant raised customer support throughput by 14 percent on average and by 34 percent for novices, largely by feeding them the phrasing and solutions that skilled colleagues had learned the hard way. That is a genuine productivity win. But notice what it is: the tool disseminating expertise the novice did not build. Whether the novice internalizes any of it, or simply leans on the tool forever, is an open question the throughput number cannot answer. The same ambiguity sits inside every AI-assisted profession right now.
It helps to remember that deliberate practice was never equally powerful everywhere. The most careful meta-analysis of the field found that how much practice explains performance depends heavily on the domain, and specifically on how structured and feedback-rich that domain is.
Verified data: how much deliberate practice explains performance, by domain
| Domain | Variance in performance explained by deliberate practice | Feedback structure |
|---|---|---|
| Games | 26 percent | Tight, immediate, unambiguous |
| Music | 21 percent | Tight, immediate |
| Sports | 18 percent | Fairly tight |
| Education | 4 percent | Loose, delayed |
| Professions | Less than 1 percent | Loose, noisy, delayed |
Source: Macnamara, Hambrick and Oswald, Psychological Science (2014).
The pattern is the lesson. Practice pays off most where feedback is tight and immediate, and least where feedback is loose and delayed. AI-assisted knowledge work sits, by default, at the bottom of that table, because the tool absorbs the feedback loop and hands you a polished result before you have felt where you went wrong. The strategic response is not to abandon AI. It is to deliberately reintroduce the tight-feedback reps that the tool would otherwise erase, and to be selective about which reps those are.
The Rep Allocation Matrix
Not every rep is worth keeping. Some difficult tasks build skills that compound and stay valuable, so you should keep doing them by hand even when AI could take them. Others are drudgery that builds nothing durable, so offloading them is pure gain. The mistake is treating the choice as all-or-nothing, either doing everything manually out of nostalgia or offloading everything and hollowing out. The Rep Allocation Matrix forces the choice rep by rep.
CEOtudent editorial framework: The Rep Allocation Matrix
| Builds a compounding skill | Builds little durable skill | |
|---|---|---|
| Core to your value | KEEP the reps. Do these by hand, use AI only to critique your attempt after you make it. | AUTOMATE with review. Let AI do it, but check the output so your judgment stays sharp. |
| Peripheral to your value | SAMPLE the reps. Do them by hand occasionally to keep the capability alive, offload the rest. | OFFLOAD fully. This is what AI is for. Reclaim the time for the top-left quadrant. |
The top-left quadrant is sacred. These are the reps where difficulty is the point, because struggling through them is what builds the judgment that makes you valuable. If you are a writer, drafting the hard argument yourself belongs here. If you are an analyst, reasoning through the ambiguous case belongs here. Use AI in this quadrant only as a critic that responds after you have committed to your own attempt, never as a substitute that answers before you have tried. That sequencing preserves the feedback loop, which is the whole mechanism, and it connects directly to the evaluation skill of judging AI output, because you can only evaluate the model well if you can still do the work yourself.
The bottom-right quadrant is where AI earns its keep. Formatting, boilerplate, first-pass research collation, and routine transformation build no lasting skill and should be offloaded without guilt. The reclaimed time is not meant for leisure. It is meant to be reinvested into the top-left reps you would otherwise never have time for.
The four-part practice protocol
Once you know which reps to keep, you still have to make them deliberate rather than mindless. Ericsson’s loop had four elements, and AI erodes each one in a specific way. Restoring them is the protocol.
CEOtudent editorial framework: restoring the deliberate practice loop under AI
| Loop element (Ericsson) | How AI erodes it | How to restore it deliberately |
|---|---|---|
| A specific goal at the edge of ability | The tool defaults to easy, finished output, so you never reach your edge | Set the task one level above what AI would hand you, and attempt it before prompting |
| Full concentration and effort | Accepting a suggestion takes one keystroke, so effort collapses | Do the first attempt with the tool closed, then open it |
| Immediate, informative feedback | The tool gives you a polished answer, not a diagnosis of your error | Use AI as a critic: ask it to find the flaw in your attempt, not to redo it |
| Repetition with refinement | One good output feels like completion, so you stop | Redo the same class of problem across a week, comparing your attempts over time |
The practical shape of this is a rhythm, not a rule. Attempt first, with the tool closed. Then invite the tool to critique, not to replace. Then repeat the class of problem enough times that your unaided attempts visibly improve. This is slower than pure automation on any single task, which is exactly why most people will not do it, and exactly why the few who do will pull ahead as skills keep expiring. We covered how fast that expiry runs in the half-life of skills in 2026, and the broader learning toolkit in what the evidence says about learning.
Why this is a CEO-and-student problem
A CEO thinks in terms of assets that compound versus expenses that do not. Your capability is a compounding asset, and the danger of AI is that it lets you draw down that asset invisibly, spending your existing skill on today’s output while investing nothing back into it. The Rep Allocation Matrix is capital allocation for your own mind: protect the reps that compound, expense the ones that do not.
The student half is the discipline to keep doing hard things when an easier path is one keystroke away. It is genuinely harder to stay a student now than at any point before, because the tool removes the friction that used to force learning on you. The people who thrive will be the ones who reintroduce that friction on purpose in the quadrants that matter, treating every AI-assisted task as a choice between finishing faster and getting better, and choosing to get better often enough that it compounds. If you want the map of which capabilities are worth this investment, the ten cognitive skills AI cannot automate is the place to start, and the meta-skills that make every other skill easier covers the practice habits that transfer across domains.
Frequently asked questions
Does using AI actually make my skills worse, or just stop them improving?
Both are possible, and which one happens depends on how you use the tool. If you offload reps you never do by hand again, capability decays through disuse, the same way any unpracticed skill fades. If you keep doing the core reps and use AI as a critic, your skills can improve faster than before, because the tool gives you feedback you could not otherwise get. The tool is neutral; the usage pattern decides the outcome.
Is deliberate practice still relevant if AI can do the task better than I can?
Yes, for the reps in the top-left quadrant, because the goal there is not to beat the AI at the task. It is to keep the judgment that lets you direct and evaluate the AI. A manager does not need to type faster than their team, but they do need to judge the work, and that judgment atrophies without practice.
How do I find time to do reps by hand when AI is so much faster?
You buy the time by offloading the bottom-right quadrant aggressively. The reclaimed hours from automating drudgery are the budget for deliberate practice on what matters. If you automate everything and reinvest nothing, you have optimized for output at the direct cost of capability.
Can I use AI as the feedback source in deliberate practice?
Yes, and this is one of its best uses. Asking a model to critique your attempt, find its weaknesses, or compare it to a stronger version supplies the immediate, informative feedback that deliberate practice requires and that is often missing in professional work. The key is that you attempt first and let the model respond to your attempt, rather than letting it produce the attempt for you.
Which skills should I protect first?
Protect the ones that are core to your value and that compound over time, the top-left quadrant of the matrix. In most knowledge roles these are the reasoning, judgment, and communication skills that determine which work is worth doing, rather than the mechanical execution skills that AI now handles well.
Kaynakça
- K. Anders Ericsson, Ralf Th. Krampe and Clemens Tesch-Romer, “The Role of Deliberate Practice in the Acquisition of Expert Performance,” Psychological Review, 1993.
- Brooke N. Macnamara, David Z. Hambrick and Frederick L. Oswald, “Deliberate Practice and Performance in Music, Games, Sports, Education, and Professions: A Meta-Analysis,” Psychological Science, 2014.
- K. Anders Ericsson and Robert Pool, Peak: Secrets from the New Science of Expertise, Houghton Mifflin Harcourt, 2016.
- Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work,” Quarterly Journal of Economics, 2025.
- World Economic Forum, Future of Jobs Report 2025.
- OECD, OECD Skills Outlook 2023.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
