10,000 hours - The Automaticity Trap for AI
Ericsson’s Research on Deliberate Practice Explains the Difference Between Organisations That Have Used AI for a Month and Organisations That Have a Month’s Expertise in AI
Most organisations that have been “doing AI” for six months months have six months months of experience. They do not have six months of expertise. The distinction is not semantic. It is the central finding of one of the most important bodies of research in the psychology of performance and it explains why enterprise AI adoption is slow to unlock genuine capability.
K. Anders Ericsson, a Swedish psychologist who spent his career at Florida State University, devoted four decades to studying how people become genuinely excellent at things. His conclusion, supported by studies across chess, music, surgery, sport, and dozens of other domains, overturns a comforting assumption in professional life: that doing something for long enough makes you good at it. It turns out that it does not. For someone with three decades of experience that makes me wince.
Doing something for long enough makes you automatic at it - fast, fluent, and permanently stuck at whatever level you reached when you stopped effortfully improving. Experience produces automaticity. Only deliberate practice produces expertise. And the difference between the two is the difference between an organisation that uses AI and an organisation that is genuinely good at it. The “10,000 hours” figure that became attached to his name was Malcolm Gladwell’s simplification, not Ericsson’s finding, and it has caused considerable confusion.
This short post focuses on the core insights most directly relevant to anyone designing the learning pathways that AI transformation requires and on why the standard approach to “upskilling” systematically fails to produce the expertise it promises.
1. The Automaticity Trap: Why Experience Is the Enemy of Expertise
Ericsson’s most important finding is also his most counterintuitive: beyond a certain point, experience actively prevents improvement.
The mechanism is automaticity; the process by which repeated performance becomes effortless, unconscious, and resistant to change.
When you first learn to drive, every action requires conscious attention: mirror, signal, gear, steering, speed (or at least until we started driving electric cars...) Within months, the process became automatic. You drive while thinking about other things. You have moved from what Giddens calls discursive consciousness - knowledge you can articulate and examine, to practical consciousness, knowledge embedded in bodily routine that operates below awareness. This transition feels like mastery. It is actually the end of learning.
The same process occurs in every professional domain. The developer who has written Java for fifteen years has not improved at Java for thirteen of those years. They improved rapidly during the first two years when every task required conscious attention, when errors were frequent and feedback was immediate, and then reached a level of automaticity that felt like competence but was actually a performance plateau. They can write Java fluently. They cannot write it brilliantly. And improvement is hard, because the fluency itself has made the activity (relatively) effortless, and effortless activity does not trigger the learning mechanisms that produce improvement.
This is the finding that most organisations refuse to absorb. The assumption embedded in every resourcing decision, every promotion criterion, and every rate card is that experience equals expertise; that ten years of doing something means ten years of getting better at it. Ericsson’s research questions this assumption. Ten years of doing something means two years of getting better and eight years of performing at the same level with increasing speed and decreasing effort. The experienced person is not more skilled than they were eight years ago - necessarily. They are more automatic. And automaticity, in any domain undergoing fundamental change, is not an asset. It is a liability because the automatic responses that served the old domain will be applied, without conscious examination, to the new one.
Argyris identified the organisational expression of this individual phenomenon. His “skilled incompetence” — the highly developed ability of smart people to avoid learning, is the professional equivalent of Ericsson’s automaticity. Argyris’s Model I behaviours (unilateral control, suppress negative feelings, win, do not lose) become automatic through years of professional socialisation, and their very automaticity makes them invisible to the person displaying them. The senior architect who “just knows” the right technical approach is not exercising expert judgment. They are exercising automatic pattern-matching that was calibrated for a world that perhaps no longer exists. Ericsson explains why they cannot improve through additional experience: the same automaticity that makes their current performance effortless prevents the conscious, effortful engagement that improvement requires.
2. Deliberate Practice: The Structure That Produces Expertise
If experience does not produce expertise, what does? Ericsson’s answer is deliberate practice. Deliberate practice is not “working harder” or “trying your best” or “learning by doing.” It has specific, non-negotiable characteristics that distinguish it from every other form of practice:
Targets a specific, well-defined aspect of performance. Not “get better at specification writing” but “improve the precision of constraint specification in the payments domain.” The target must be narrow enough that feedback can identify what specifically is working and what specifically is not.
Designed by a teacher or coach who understands expert performance. The practitioner alone cannot design effective deliberate practice, because they do not yet know what expert performance looks like in sufficient detail to identify the specific gaps between their current performance and that standard. This is a finding that somewhat challenges the “self-directed learning” paradigm: without external guidance, people practice what they are already good at and avoid what they find difficult. That said, there are increasingly expert curators of learning online that, with some assiduous research, can provide the structure you need. I hope this series is one such.
Demands full concentration and effort. Deliberate practice is cognitively exhausting. It is not enjoyable in the moment. Ericsson found that even elite performers can sustain deliberate practice for only a few hours per day before cognitive fatigue degrades its quality. This stands in stark contrast to the standard training model, which assumes that a full day of workshops produces a full day of learning. It does not. It produces two hours of learning and six hours of diminishing returns.
Involves immediate, informative feedback. Not “your design was good” but “your design handled the happy path precisely, but the error conditions were underspecified - here are the three modes of failure your specification does not address, and here is how an expert would have addressed them.” The feedback must be specific enough to guide the next attempt and timely enough that the connection between action and result is vivid.
Builds on existing knowledge through progressive challenge. Each practice activity is slightly beyond current capability - hard enough to require effort, achievable enough to permit success. Csikszentmihalyi’s flow research maps the emotional terrain where too much challenge produces anxiety, too little produces boredom. But Ericsson adds a crucial nuance; that deliberate practice operates at the edge of the flow channel, in the zone where the challenge slightly exceeds current skill. It is not comfortable. It is the productive discomfort from which expertise grows.
The implications for AI transformation in particular are clear. For instance, specification writing, being the ability to articulate domain knowledge with sufficient precision for AI to act on it, is a learnable skill with a definable expertise trajectory. It is not a talent that some people have and others lack. But learning it requires all five characteristics of deliberate practice, and most organisations provide few of them. They provide exposure to the concept (a training day), occasional unstructured practice (using AI in a new project), no expert-designed progression, no specific feedback, and no coach. They provide, in Ericsson’s terms, naive practice - repetition without design, effort without structure, experience without improvement.
3. Mental Representations: What Experts Actually Have That Novices Lack
Ericsson’s most theoretically important contribution and the one most relevant to understanding what makes good specification writers different from poor ones, is the concept of mental representations. These are the internal models that experts build through years of deliberate practice, and they are what make expert performance qualitatively different from novice performance.
A chess grandmaster does not see a chessboard the way a novice does. The novice sees thirty-two pieces in various positions. The grandmaster sees patterns, structures, threats, and possibilities. When shown a position from a real game for five seconds, the Magnus Carlsen can reconstruct the board perfectly. When shown random pieces in random positions, the grandmaster performs no better than the novice. The expertise is not in memory. It is in the mental representations, or the patterns that organise perception, that make certain moves “obvious” and certain dangers “visible” that apply to meaningful configurations.
Weick’s sensemaking framework illuminates the broader significance. Sensemaking is the process by which people impose order on ambiguous situations by selecting certain cues and fitting them into a plausible frame. Mental representations determine which cues are selected. The novice, writing a specification for a loan approval process, selects the obvious cues: income, credit score, loan amount. The expert selects the same cues plus the non-obvious ones: the regulatory edge cases, the fraud scenarios, the boundary conditions where the rules conflict, the implicit assumptions about data quality. The expert does not think harder. They see more because their mental representations, built through deliberate practice, organise their perception differently.
This has a direct implication for training design. Mental representations, in Ericsson’s research, cannot be transferred through instruction. They must be built through practice. You can explain the concept of edge cases. You cannot install the perceptual habit of noticing them. That habit develops only through the repeated experience of writing specifications that fail (for instance) because edge cases were missed, receiving feedback that identifies the specific gaps, and revising. Each cycle deepens the mental representation. Each failure, properly examined, makes the practitioner’s internal model of “good specification” slightly richer. Over time, what required conscious analysis becomes perceptual and the expert “sees” the edge cases the way the grandmaster “sees” the threatened square. But this perceptual expertise cannot be shortcut. It is the accumulated product of deliberate practice, and nothing else produces it.
Mintzberg’s Craft School of strategy describes the same phenomenon at the organisational level. The master strategist does not deduce strategy from analysis. They develop an intimate feel for the material through sustained engagement, what Mintzberg called “knowing in the fingers.”
4. The Flow Paradox: Why Deliberate Practice Hurts and Why That Matters
There is a tension at the heart of Ericsson’s work that connects to one of the most important practical questions in any transformation: will people actually do this?
Elite violinists, in Ericsson’s studies, rated deliberate practice as the least enjoyable of their musical activities even as they rated it the most important for their development. They did it not because they liked it but because they had internalised the connection between this specific kind of effortful practice and the improvement they valued.
Csikszentmihalyi’s flow research describes the opposite experience though. Flow - the state of complete absorption, intrinsic reward, and effortless concentration - occurs when challenge and skill are matched at a high level. People seek flow. They organise their lives around activities that produce it. Developers, in particular, experience flow while coding. For those of us who have experienced it; the immediate feedback of compilation, the exultant feeling of passing tests, the progressive challenge of increasingly complex problems; fills us with joy.
The paradox is that deliberate practice and flow occupy different psychological territories.
Flow is what sustains engagement. Deliberate practice is what produces improvement. Performing at your current level in a state of flow does not make you better, it makes you faster and more automatic at your current level. Practicing deliberately, at the uncomfortable edge of your capability, is what makes you better but it does not feel like flow. It feels like struggle.
Deci and Ryan’s self-determination theory resolves the paradox or at least maps the conditions under which it can be managed. The competence need is not satisfied by flow alone. It is satisfied by the felt experience of growing mastery; of being able to do today what you could not do yesterday. This experience comes from deliberate practice, not from fluent performance. But sustaining deliberate practice requires that the other two needs - autonomy and relatedness - are also satisfied. The person who is forced to practice a new skill through a mandated training programme experiences controlled motivation, which Deci and Ryan have shown produces compliance with less effective learning than when they choose to practice because they can see how it connects to problems they care about. They experience autonomous motivation, which sustains the effortful engagement that deliberate practice demands.
Tom Peters intuited this without the psychological framework: excellence requires passion, and passion cannot be mandated. Ericsson provides the mechanism: the passion sustains the practice, and the practice produces the expertise. But the passion must be genuinely connected to intrinsic interest and autonomous choice or the practice degrades into the rote repetition that produces automaticity, not expertise.
5. The Coach Problem: Why Self-Directed Learning Has Structural Limits
Ericsson’s insistence that deliberate practice requires a teacher or coach is the finding that creates the most discomfort for organisations committed to “self-directed learning” and “empowered teams.” The discomfort is understandable: coaches are expensive, scarce, and difficult to scale.
The function of the coach is not motivation, though motivation matters. It is design. The coach understands expert performance in the domain well enough to identify the specific gaps between the learner’s current performance and the expert standard. The coach designs practice activities that target those specific gaps, not generic exercises that cover the domain broadly, but focused activities that force the learner to confront exactly the aspect of performance where they are weakest. And the coach provides feedback that is specific, immediate, and actionable.
Practicing strengths produces the flow and competence feelings that Deci and Ryan describe. Practicing weaknesses produces frustration, anxiety, and the threat to self-efficacy that Bandura documents. Left to their own devices, people will choose flow over frustration every time. The coach’s role is to redirect effort toward the areas where improvement is possible and away from the comfortable repetition where it is not.
For AI transformation, this means that “give everyone access to AI tools and let them learn” is not a learning strategy. It will produce widespread familiarity and not necessarily expertise. The teams that develop the most capability will be the teams that have access to someone who can design their practice, examine their product with expert eyes, and provide the specific feedback that drives improvement.
A coach might be someone like an internal champion, an external consultant, a more experienced peer, or - and this is one of AI’s most exciting possibilities - the AI tool itself, configured to provide formative feedback on quality rather than merely executing whatever specification it receives.
Edmondson’s psychological safety is the precondition for coaching to work. Coaching involves identifying weaknesses and providing feedback on them. If the environment is unsafe; if having your product examined and critiqued carries social risk, people will avoid coaching, avoid feedback, and practice in private where their errors are invisible. The coach-learner relationship requires precisely the interpersonal risk-taking that Edmondson defines as the essence of psychological safety: the willingness to appear incompetent, to ask naive questions, and to have your work scrutinised.
Carol Dweck’s mindset research determines how feedback is received. In a growth mindset culture, the coach’s feedback is received as developmental information: here is where you can improve. In a fixed mindset culture, the same feedback is received as a verdict on ability: you are not good at this. The culture determines whether coaching produces learning or defensive withdrawal, and no coaching methodology can overcome a culture that treats feedback as judgment.
6. Designing the Practice Architecture: From Training Events to Learning Systems
The synthesis of Ericsson’s research with the thinkers explored across this series produces a design framework for how organisations can build genuine AI expertise, not the cosmetic familiarity that six months of unstructured experience provides, but the deep, flexible, transferable capability that deliberate practice develops.
The framework has five elements, each grounded in a different body of research:
Progressive challenge sequences, not generic training. Ericsson requires that practice be structured with increasing difficulty targeted at specific aspects of performance. The first exercise should address a clearly bounded, well-understood domain problem. The second should introduce ambiguity — edge cases where the rules conflict, boundary conditions where domain knowledge becomes uncertain. The third should tackle a problem where Snowden’s Cynefin framework would classify the domain as Complex: where the task must be framed as a probe rather than a solution, where the validation criteria must include “what would tell us this answer is wrong?” The progression builds mental representations by systematically confronting the learner with the gaps in their current model of what a “complete specification” looks like.
Specific, immediate feedback on specification quality. Generic feedback does not drive deliberate practice. Effective feedback identifies specific gaps: “The specification you produced handles the approval workflow but does not address what happens when approvers are unavailable. In this domain, approver absence is the primary source of process delays. An expert specification would include timeout conditions, escalation rules, and the specific state the application enters when no approval is received within the SLA.” This level of specificity requires either human expertise or, increasingly, AI systems provided sufficient context to supply formative assessment rather than mere execution.
Coaching relationships, not self-directed exploration. Someone must design the practice sequence, deliver the feedback, and adjust the challenge level based on the learner’s development. This might be a human coach, a peer working at a more advanced level, or a structured AI feedback system but it cannot be nothing. “Give teams access to AI tools” without coaching support will produce the same plateau that Ericsson found in every other domain: rapid initial improvement followed by permanent automaticity at a mediocre level.
Psychological safety as the operating environment. Deliberate practice requires confronting weaknesses, receiving critical feedback, and performing at the edge of competence where failure is frequent. Edmondson’s research is a prerequisite: without safety, people will not submit their work for critique, will not attempt tasks where they might fail, and will not engage in the effortful struggle that deliberate practice demands. They will instead practice in private, avoid feedback, and perform only at the automatic level where their competence is unthreatened, which is exactly the automaticity trap that Ericsson warns against.
Autonomous motivation to sustain the effort. Deliberate practice is cognitively exhausting and emotionally uncomfortable. Sustaining it requires the intrinsic motivation that Deci and Ryan describe. Autonomy (the person chooses to engage), competence (they experience growing mastery despite the difficulty), and relatedness (they are practicing within a community of others making the same journey). Mandated practice sessions, compliance-driven participation, and performance-linked incentives produce the controlled motivation that generates compliance without learning. The practice might look the same from outside. The learning will not be the same on the inside.
This five-element architecture is not a training programme. It is a learning approach, an ongoing structure embedded in the way work is done, rather than an event that interrupts work temporarily and is then forgotten. The distinction matters because Ericsson’s research shows that expertise develops over months and years of sustained deliberate practice, not in a two-day workshop. The workshop might provide the initial mastery experience that Bandura identifies as essential for building self-efficacy. But the self-efficacy is the starting condition, not the destination. The destination is expertise — and expertise requires a practice architecture that persists long after the workshop is over.
(An Organisational Prompt is something you can do now…)
Organisational Prompt
Find someone in your organisation who has been using AI for at least three months. Ask them to show you their most recent AI-generated output. Then ask a question
“Are you working on harder problems now than three months ago?” If not, if they are applying AI to the same level of problem with greater speed, they have reached the plateau.
Experience is accumulating. Expertise is not. The challenge level has not increased, so the practice is no longer deliberate. How could you provide the support they need to move to the next level?
Further Reading
K. Anders Ericsson, Robert Pool: Peak: Secrets from the New Science of Expertise - The accessible version of Ericsson’s research, written for a general audience. Read it for the core framework and the case studies. The distinction between naive practice, purposeful practice, and deliberate practice is the single most important framework in the book.
Mihaly Csikszentmihalyi: Flow: The Psychology of Optimal Experience - The essential companion to Ericsson. Flow explains why people continue doing things; deliberate practice explains how they improve at them. Understanding the tension between the two is critical for designing learning systems that sustain engagement while driving genuine development.
Albert Bandura: Self-Efficacy: The Exercise of Control - Ericsson explains how expertise develops. Bandura explains why people attempt or avoid the practice that develops it. Self-efficacy determines whether someone will engage in the deliberate practice that would make them genuinely good at it.
Amy Edmondson: The Fearless Organization - Deliberate practice requires confronting weaknesses, receiving critical feedback, and failing frequently in the service of improvement. None of this happens without psychological safety. Edmondson provides the environmental condition without which Ericsson’s framework cannot operate in an organisational setting.
Disclaimer
I write about the industry and its approach in general. None of the opinions or examples in my articles necessarily relate to present or past employers. I draw on conversations with many practitioners and all views are my own.