Six articles. One question every organization is avoiding: Are you building capability or just building dependency?

Two Organizations, Five Years From Now

Imagine two organizations, both operating in the same industry, both with broadly similar AI investments, both with broadly similar adoption metrics today.

At Organization A, investing in AI has significantly boosted speed and productivity. This has led to higher-quality outputs and better-informed decisions, creating real and sustained efficiency gains. Over the past five years, the organization has developed an impressive AI capability, complete with sophisticated tools, well-trained users, and optimized workflows.

While their people are better at using AI than they were five years ago, their underlying capabilities have not grown. When conditions change, a novel threat arises, a client requires judgment the AI wasn’t trained on, or a strategic turning point demands thinking no tool can generate, the organization is left with the same human skill it had five years ago, only now, it’s trapped within a more sophisticated dependency on the AI tools.

Similarly, Organization B’s investment in AI has enhanced the speed and productivity of its workforce. While the efficiency gains are on par, over five years, the organization has cultivated a more profound asset - deepened human capability. The professionals who use AI daily are not merely more productive; they have become better at their jobs. Their judgment has matured, their pattern recognition has sharpened, and their ability to navigate novel complexities has expanded. In this environment, the AI tools are valuable precisely because the humans using them are excellent, and both are continuously improving together.

Fast forward five years, and the two organizations are in vastly different positions. While Organization A has grown in one dimension, the other has quietly stagnated. In contrast, Organization B has achieved simultaneous, compounding growth across both dimensions.

The crucial distinction wasn’t the tools they used, their adoption rates, or their utilization metrics. Instead, it was the strategic decisions their leaders made about integrating those tools into the very fabric of how their people work, learn, and grow.

Every organization has the power to make these decisions, though most have not done so deliberately. This article explores what happens when they do.

What the Series Has Built To

Across five articles, a single, fundamental problem has been examined from multiple angles: organizations are deploying AI at scale but failing to invest in the human development required to unlock its long-term value.

In this series, we’ve explored the persistent gap between technology adoption and true business transformation.

• Article 1 established the core issue: this isn’t a technology problem, but a human capability design problem.

• Article 2 defined the leader’s role in deliberately shaping work to build, rather than erode, human capability.

• Article 3 provided individual professionals with the key behavioral habits needed to secure their future relevance.

• Article 4 offered organizations the metrics to identify the capability gap before it becomes critical.

• Article 5 highlighted the team-level failure modes that are invisible to individual-level analysis.

This article builds the architecture that addresses all of them.

It’s not a theory but a design system, but rather a set of six elements that, when integrated into an organization’s AI strategy, can transform it from Organization A to Organization B within a five-year timeframe.

The Design System: Six Elements

Element One: The Capability Integration Audit

Before integrating AI into any significant workflow, it’s crucial to assess the human capabilities developed by the current process. This assessment should distinguish between two categories of skills:

Category 1: Skills that can be automated without any strategic loss.

Category 2: Skills that must be preserved because the expertise and judgment gained from performing them are a core competitive advantage.

If you outsource Category 2 work to AI, your team will gradually lose the ability to evaluate the quality of the AI’s output, becoming unable to differentiate between high-quality work and superficial results.

This audit should answer three questions for every significant AI integration decision:

• What capability does this workflow currently develop in the humans doing it? Not what outputs it produces but what cognitive muscles it exercises, what judgment it builds, what expertise it develops through repetition.

• Does that capability belong in Category 1 or Category 2? Using the three diagnostic questions from Article 3 of this series, can the output be fully specified in advance, does competitive advantage come from the result or from what is built in producing it, and would clients or stakeholders care whether a human or AI produced this?

• If Category 2: how does the redesigned workflow preserve and develop that capability rather than substituting for it? What specific design choices ensure that AI integration sharpens human capability?

The AI audit is not a barrier to adoption, but a design practice that strengthens an organization rather than making it more fragile. By conducting an audit before integration, companies can avoid the common pitfall of discovering capability costs only after those capabilities have been lost, a pattern highlighted throughout this series.

Element Two: The Sharpening-First Workflow Design

Every significant AI-assisted workflow should be designed around the sharpening pattern rather than the replacement pattern as its default sequence.

The replacement pattern is where you encounter a problem, query AI, review output, and submit. It’s the path of least resistance. It is faster, easier, and produces good outputs in the short term. However, it also systematically bypasses the cognitive process that builds genuine expertise.

The sharpening pattern is where you first form your own analysis, use AI to challenge it, compare positions, update where it is earned, and submit. It requires more time and more discipline, but it also produces a professional whose capability compounds with every AI interaction rather than one whose capability migrates into the tool.

Designing for the sharpening pattern means making it the default sequence, not the exception. This involves three specific design decisions.

• First, workflow design should build in a distinct step for “independent analysis” before making AI assistance available for any important work. While the AI is accessible, the default process should not be to use it first.

• Second, we must foster a culture where professionals use AI not as a substitute for original thought, but as a partner for intellectual sparring. Instead of asking, “What is the best analysis of this situation?” one should propose, “Here is my analysis; what am I missing?” Although the tool remains the same, this simple shift in approach can have a profound developmental effect over time.

• Third, professionals should engage in evaluation practices where they articulate the reasoning behind their AI-assisted work. This isn’t for auditing purposes, but rather a developmental exercise. It cultivates a deeper understanding that enables them to make AI output genuinely useful rather than just passively consuming it.

Element Three: The Deliberate Difficulty System

Robert Bjork’s research on desirable difficulties is unambiguous. The conditions that feel hardest in the moment produce the strongest capability over time. The conditions that feel easiest (smooth, immediate, frictionless) produce the weakest long-term retention and skill development.

AI-optimized workflows systematically remove friction, which is their immediate value. However, this convenience, if left unchecked, can become their long-term cost.

Designing for deliberate difficulty means building structured experiences where professionals regularly encounter the productive struggle that AI typically removes, not all the time, and not in ways that impair productivity, but deliberately and regularly enough that the capability muscles are exercised rather than atrophied.

In practice, this looks like three things built into how work is organized.

• To maintain and assess your team’s core skills, hold regular “scaffold-removal sessions.” These monthly or quarterly meetings challenge teams to solve significant problems without defaulting to AI. The goal isn’t to test or punish, but to deliberately practice and keep underlying skills sharp. This also provides leaders with a clear view of their team’s true capabilities.

• Assign stretch projects that challenge professionals to build upon the AI’s output, treating it as a starting point for independent development rather than a finished product for review. These tasks should push them just beyond their current comfort zone, encouraging them to reach for new skills rather than rely on existing ones.

• Offer transfer practice, which deliberately exposes professionals to new problem structures. This requires them to apply their domain knowledge in unfamiliar contexts, a skill that Article 4 identifies as the clearest measure of genuine expertise versus sophisticated tool fluency.

Element Four: The Capability Feedback Loop

The metrics from Article 4 (the Capability Gap Index, Reasoning Depth Score, Transfer Performance Index, and Development Trajectory Score) should be integrated into the organization’s performance architecture, not as annual assessments but as ongoing feedback mechanisms.

The focus should be on development, not evaluation. If a professional’s Capability Gap Index is widening, meaning the performance gap between their AI-assisted and unassisted work is growing, it signals a need for a development conversation, not a performance review.

• The question becomes: where is this dependency forming, and what specific practice can be implemented to address it?

When a team’s challenge rate declines, and its members begin confirming each other’s AI-assisted analyses rather than genuinely challenging them, it’s a clear signal that a conversation about workflow redesign is needed.

• The question then becomes: which practice from Article 5 can restore the productive disagreement this team requires?

For a feedback loop to be effective, it must lead to action. This means managers need to be skilled in conducting developmental conversations, not just performance reviews. They must be able to distinguish between saying “your output is below standard” and “your performance is declining, so let’s build a plan together.”

Crucially, they also need the right development tools to address the issues the metrics reveal.

Element Five: The Learning Integration Model

The Outlearn Loop (NOTICE, BUILD, HARDWIRE, PERFORM) applied at the organizational scale is the connective tissue that holds the design system together.

At the organizational level, the first phase, NOTICE, involves creating a diagnostic infrastructure to reveal capability gaps. This is achieved through audits before making integration decisions, applying the metrics outlined in Article 4, and implementing the team-level diagnostics from Article 5. After all, you can’t design solutions for problems you can’t see.

The second phase, BUILD, at the organizational level means designing AI-integrated workflows around the sharpening pattern, ensuring that the cognitive work that builds genuine expertise is preserved within AI-assisted processes rather than bypassed by them. Independent analysis first. Challenge-based prompting. Explanation requirements for significant AI-assisted work.

To HARDWIRE at the organizational level means to embed structures that reinforce learning and turn it into a lasting capability. This includes integrating spaced-retrieval practice into development programs, adding reflection prompts to workflows, and establishing team-level practices from Article 5. These measures transform productive challenge from a social risk into a cultural norm.

At the organizational level, the final phase, PERFORM, involves removing the AI “scaffolding.” This means creating regular opportunities for professionals and teams to demonstrate their capabilities without AI assistance. This isn’t a test, but rather a diagnostic tool and a commitment to development. Performance without the scaffold reveals the truth: is the system genuinely building new capabilities, or merely borrowing them?

Element Six: The Leadership Operating System

Each of these elements requires leaders to ask different questions about the organizations they lead.

• Not just “Are our people using AI effectively?” but “Is our use of AI making our people more capable, or more dependent?”

• Not just “Are our outputs improving?” but “Are the people who produce those outputs improving?”

• Not just “Are we getting more from our people with AI?” but “Are we building people who can get more from AI because they are growing more capable themselves?”

The leadership operating system isn’t a collection of new processes. Instead, it’s a framework of consistently asked new questions woven into work reviews, performance evaluations, development investments, and integration decisions.

By making it standard practice to ask, “What was your thinking before you used AI?” in every significant review, a leader installs a crucial element of this system. When posed consistently, this question transforms critical thinking from an individual discipline into a cultural expectation.

By separating discussions about output from those about development during feedback sessions, leaders can embed a continuous cycle of capability-building into their organization’s daily operations. This distinction illuminates crucial aspects of performance that output metrics alone often miss.

A leader who openly sharpens their thinking—reflecting before they question, using AI to challenge their reasoning rather than replace it, and engaging in the deliberate practice that hones keen judgment—builds the desired culture through their actions long before any policy is written.

The Decision This Series Has Been Building Toward

Across six articles, I have built a single argument. Now, allow me to state it plainly.

The organizations that will lead in the AI era will not be the ones with the highest adoption rates, the best tools, or the most sophisticated AI capabilities. They will be the ones where human capability and AI capability have grown together, where the people using the tools are genuinely better at their work because of how they use them, not just faster at producing outputs.

This outcome is not accidental. It is not the result of good intentions, the right tools, or even effective change management. Instead, it requires deliberate design and specific decisions about the structure of work, the integration of AI, the measurement of capability, team collaboration, and the nature of leadership.

Every element of that design is within reach for any organization. No part of it requires waiting for better AI, superior tools, or a more favorable market. What it does require is starting now, before the capability gap widens to the point where it can no longer be closed by development alone.

The question this series has been building toward is not a diagnosis. It is a decision.

Are you building capability or just building dependency?

The answer to that question is being determined right now. It’s woven into every AI integration your organization considers, every workflow that either fosters or forfeits the cognitive effort required for true expertise, and every leadership conversation that chooses to either probe deeper or settle for the surface-level solution.

Organizations that get this right will see human judgment and capability grow in tandem with AI-driven productivity. Those that don’t will face a widening gap between their AI systems and the human insight needed to unlock their true value.

This disparity is the capability crisis this series began with.

The design in this article is how you close that gap, before it closes options for you and your organization.

This concludes The AI Capability Playbook. If the series has given you a framework for building the organization that gets smarter rather than more dependent, share it with the leaders in your organization who are making the AI integration decisions that will determine which future you are building.

For the individual-level application of everything in this series, The Performance Playbook on this channel is the operational companion, which is the Outlearn Loop applied to your own professional development, phase by phase.

References

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.

Bjork, R. A., & Bjork, E. L. (1992). A new theory of disuse and an old theory of stimulus fluctuation. In A. Healy, S. Kosslyn, & R. Shiffrin (Eds.), From learning processes to cognitive processes: Essays in honor of William K. Estes (Vol. 2, pp. 35–67). Lawrence Erlbaum Associates.

Deci, E. L., & Ryan, R. M. (2000). The “what” and “why” of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268. https://doi.org/10.1207/S15327965PLI1104_01

Gartner. (2025). CIO and technology executive survey. Gartner Research.

Gollwitzer, P. M. (1999). Implementation intentions: Strong effects of simple plans. American Psychologist, 54(7), 493–503. https://doi.org/10.1037/0003-066X.54.7.493

Hadwin, A. F., Järvelä, S., & Miller, M. (2011). Self-regulated, co-regulated, and socially shared regulation of learning. In B. J. Zimmerman & D. H. Schunk (Eds.), Handbook of self-regulation of learning and performance (pp. 65–84). Routledge.

McKinsey & Company. (2025). The state of AI: How organizations are rewiring to capture value. McKinsey Global Institute.

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604. https://doi.org/10.1037/0278-7393.4.6.592

Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745

Stasser, G., & Titus, W. (1985). Pooling of unshared information in group decision making: Biased information sampling during discussion. Journal of Personality and Social Psychology, 48(6), 1467–1478. https://doi.org/10.1037/0022-3514.48.6.1467

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–40). Academic Press. https://doi.org/10.1016/B978-012109890-2/50031-7

Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70. https://doi.org/10.1207/s15430421tip4102_2

Zimmerman, B. J., & Schunk, D. H. (Eds.). (2011). Handbook of self-regulation of learning and performance. Routledge.

Charles Good works with organizations that have already solved the adoption problem and are now asking the harder question. As President of the Institute for Management Studies, he reaches over 20,000 professionals annually. His Outlearn Loop framework, built on behavioral and learning science, is the architecture organizations use to design AI integration that builds human capability rather than substituting for it. He writes The AI Capability Playbook and The Performance Playbook on Substack and hosts The Good Leadership Podcast (Apple / Spotify / YouTube).

Six articles. One question every organization is avoiding: Are you building capability or just building dependency?

The Team That Looked Brilliant

A strategy team at a financial services firm had developed a reputation for sharp, well-integrated thinking. Their recommendations were coherent. Their analyses accounted for interdependencies that other teams missed. Their presentations held up under pressure.

After eighteen months into a broad AI rollout, something had quietly changed.

While the outputs had noticeably improved, becoming more thorough, consistently formatted, and data-rich, the quality of the surrounding conversation had declined. Working sessions saw fewer genuine disagreements, less pushback on underlying assumptions, and a tendency to converge on recommendations that felt predetermined. felt predetermined.

When an executive questioned the team lead, his initial response was to point to the team’s output metrics, which had not only remained strong but had actually improved. Puzzled, the team lead couldn’t see what the problem was.

The problem, which took several more months to name, was this: the team had stopped thinking together. They had started reviewing together, but every member arrived at working sessions having already consulted AI.

Though each member presented an “independent” analysis, their conclusions were all shaped by the same underlying models, pattern-matching, and synthetic outputs. The apparent diversity of perspectives merely masked a fundamental convergence of their initial assumptions.

This team-level capability crisis is more consequential than its individual counterpart. It’s often invisible, both to outsiders and to many within the team itself. This invisibility makes it particularly damaging, as it directly erodes the very reason teams exist: to generate a collective intelligence far greater than the sum of its individual parts.

Why Teams Are Different From Individuals

Articles 1 through 4 of this series have largely addressed the individual level — what AI does to a professional’s capability, what a leader can do about it, and how any professional can measure their own development trajectory.

Teams introduce dynamics that individual-level analysis is not able to capture. When multiple people are each developing AI dependency in parallel, the team-level effects are not simply additive; rather, they interact in ways that amplify the risk.

Research into collective intelligence, the process by which groups pool information, challenge one another’s reasoning, and reach better conclusions than individuals, has identified three conditions that make a team more intelligent than the sum of its parts:

• information diversity (members bring genuinely different starting points),

• psychological safety (members challenge each other’s reasoning without social cost)

• productive conflict (disagreement surfaces assumptions that agreement would leave hidden).

AI dependency at the team level undermines all three of these conditions, often in ways that can make a team appear highly functional from an outsider perspective.

The Three Team-Level Failure Modes

Failure Mode One: The Homogenization Trap

When team members use AI independently to prepare for collaborative work, they often arrive at similar frameworks, analyses, and conclusions. This is not because they have thought the same thoughts, but because they have consulted the same underlying models. As a result, the apparent independence of their preparation is partially illusory.

This is the strategic team’s problem described in the opening scenario. It produces a specific kind of collective intelligence failure: a team that is highly capable of refining and integrating perspectives, yet whose perspectives are less diverse than they appear. The team’s range of thinking has been compressed during the preparation stage, before collaboration even begins.

The research on group information processing, along with Stasser and Titus’s work on hidden profiles, shows that teams systematically over-discuss information that members hold in common and under-discuss information that only one member holds. AI-generated preparation amplifies this bias by increasing the proportion of information held in common, while reducing the distinctive contributions that only individual human judgment and experience can produce.

When every member of a team consults the same AI, their shared, AI-generated starting points can overshadow the unique perspectives, unconventional framing, and nuanced judgment calls that arise from individual experience. The team may converge on a solution faster, but they risk overlooking crucial insights in the process.

Failure Mode Two: The Confidence Asymmetry

AI-assisted work is often polished and confident. Arguments are well-structured, data is properly cited, and recommendations are clearly reasoned. This is a genuine strength of AI, as it significantly raises the baseline quality of the output.

This also creates a problem at the team level. When one person’s AI-assisted analysis appears far more polished than a colleague’s independent work, it can skew the team’s social dynamics in ways that aren’t always productive.

The polished output commands more attention, generates less challenge, and tends to anchor the group’s discussion, not because it is necessarily better reasoned, but because it looks better reasoned.

The member who arrives most prepared with AI assistance tends to occupy an outsized share of the team’s cognitive space, not because their judgment is superior, but because the presentation of their AI-assisted work has a quality that independent thinking can’t match in the moment.

Over time, teams adapt to this new dynamic in ways that inadvertently erode their collective intelligence. When individuals consistently find that their independent contributions are overshadowed by polished, AI-assisted presentations, they may cease to offer their own unique analyses.

Instead of arriving at sessions as active contributors, they become passive reviewers. This cycle gradually diminishes the diversity of thought within the team, shrinking its capacity for genuine, collaborative problem-solving.

Failure Mode Three: The Accountability Diffusion

When AI generates the first draft of an analysis or recommendation, the team’s relationship to the work subtly but significantly shifts. Since no individual has personally originated the argument, the team’s role becomes editorial; they are now tasked with approving, refining, and presenting a work initiated by AI, rather than owning the initial thinking.

This diffusion of intellectual ownership has a predictable effect on team accountability. When the work is incorrect, it’s harder for the team to learn from their mistakes because no single person generated the information. The cognitive process required for learning (the Forethought → Performance → Self-Reflection cycle described in Zimmerman’s research) was never completed by any individual member.

The team becomes a sophisticated review mechanism for AI-generated work rather than a collective intelligence system that genuinely compounds through experience.

A team that reviews AI output in unison isn’t truly thinking together; they are merely confirming together. While the distinction may seem subtle at first, its long-term impact is significant.

What Genuine Collective Capability Looks Like

The failure modes above describe what teams lose when AI dependency develops at the collective level. It is equally important to describe what genuine collective capability looks like when it is being built with AI rather than eroded.

Hadwin, Järvelä and Miller’s research on socially shared regulated learning describes the conditions under which teams develop genuine collective intelligence, which is defined as the capacity to plan, monitor, and reflect on their work as a group rather than as a collection of individuals whose outputs happen to be aggregated.

Teams with genuine collective capability share three characteristics that distinguish them from teams that merely produce good outputs.

• They deliberate before they converge. Rather than arriving at sessions with polished positions and working to reconcile them, they invest time in shared problem framing. This allows them to build a collective understanding of what the problem is before developing solutions. This is harder with AI in the room, because

  • AI tends to accelerate convergence by producing frameworks that everyone can anchor on. However, teams that successfully cultivate genuine collective intelligence deliberately resist this premature convergence.

• They name disagreements rather than smooth them. When one member’s analysis diverges from another’s, the productive response is not to reconcile the outputs but to understand the divergence. What different assumptions are embedded in different framings? What does each analysis reveal that the other misses?

  • AI-generated outputs tend to look so coherent that divergences seem like errors to be corrected rather than signals to be investigated. Teams with genuine collective intelligence treat divergence as information.

• They reflect on their process, not just their output. After significant collaborative work, they ask not only “Did we produce a good recommendation?” but also “Did our process make us smarter?” Did the collaboration surface something that individual work would have missed? Did the challenge and pushback improve the quality of thinking? Did each member leave the session with a sharper perspective than they arrived with?

These are the characteristics that distinguish a team building collective capability from one producing good work while its intelligence quietly plateaus.

Four Practices for Building Collective Capability in an AI Era

These are specific, implementable practices that address one of the failure modes above.

Practice One: Independent analysis before collaborative synthesis.

Before any significant collaborative working session, each member produces their own independent analysis of the problem, without consulting AI first. This doesn’t mean AI cannot be used in preparation. It just means AI is used to sharpen independent thinking, not to replace it.

The expectation is that each member has genuinely thought through the problem and can articulate a position, even a tentative one, that is authentically theirs. When AI is used after this step, it should be to challenge the independent analysis, not to produce it.

This practice directly addresses the homogenization trap by ensuring that the diversity of perspectives around the table reflects genuine differences in how members have thought about the problem, not just differences in which AI outputs they chose to anchor on.

Practice Two: The disagreement audit.

At the end of significant collaborative sessions, spend five minutes on a structured questions such as the following:

• What perspectives or concerns were raised but not adequately explored?

• What disagreements were smoothed over rather than investigated?

• What assumptions in the final recommendation were not challenged?

This practice directly addresses the accountability diffusion failure mode by creating a shared habit of naming what the collaboration might have missed. It also counteracts the natural tendency, which is amplified when AI generates polished outputs, to treat convergence as evidence of correctness rather than a social dynamic that may have suppressed important information.

Practice Three: Rotating the skeptic role.

Designate one team member per session whose explicit responsibility is to challenge the AI-generated or AI-assisted analysis. In other words, they are tasked with finding the strongest argument against the emerging recommendation, surfacing the assumptions the analysis depends on, and identifying the conditions under which the conclusion would not hold.

This practice directly addresses the confidence asymmetry by institutionalizing the challenge function that AI-assisted polish tends to suppress. It makes the skeptical role a social expectation rather than a social cost, and it ensures that the team’s collective intelligence is consistently applied against its own outputs rather than in service of confirming them.

Practice Four: The collective reflection prompt.

After significant collaborative work, build a brief, structured reflection on the team’s process and the quality of the thinking by posing three questions:

• Where did our collective analysis exceed what any of us would have produced individually?

• Where did AI assistance help us think better?

• Where might our process have compressed the diversity of our thinking, and what should we watch for next time?

This practice builds the metacognitive awareness at the team level that Hadwin, Järvelä, and Miller identify as the foundation of genuine collective capability, which is the team’s capacity to observe and regulate its own thinking process, not just the outputs of that process.

The Team Capability Metrics

Article 4 in the AI Capability Series introduced four metrics for tracking individual capability development. Teams require their own diagnostic, which is the following:

Diversity of starting points: Before collaborative sessions, how different are team members’ independent analyses? A team where everyone arrives with essentially the same framing has a homogenization problem, regardless of how good the final output is.

Challenge rate: In working sessions, how often does a team member’s position genuinely change as a result of a challenge from another member, not as a compromise, but because someone surfaced something the first member had missed? This is the most direct measure of whether the team is genuinely thinking together or performing collaboration while individually anchoring on AI outputs.

Post-session capability: After significant collaborative work, are individual team members better positioned to think about this domain or problem than before the session? In other words, have they built something or have they produced something?

These metrics do not require elaborate measurement infrastructure, but they do require leaders to observe how their teams work, not just what their teams produce.

The Team as the Unit of Capability

In an AI-saturated world, sustainable competitive advantage won’t come from having the best individual AI users. Instead, it will belong to organizations with the best teams - groups that can think together, genuinely challenge one another, and generate a collective intelligence far greater than what any person, with or without AI, could achieve alone.

This capacity is built through practice, not tools alone. It requires intentionally creating an environment where intellectual diversity is protected, disagreement is seen as a signal for improvement rather than a source of friction, and the team’s goal extends beyond producing a good output to becoming genuinely better through the process itself.

Are you building that kind of team? Or are you building a sophisticated review panel for AI-generated work that looks like a high-performing team until the situation requires the thing AI cannot provide, which is genuine collective judgment under novel conditions?

If this identifies patterns you recognize in your own team, the most useful starting point is Practice One (independent analysis before collaborative synthesis) applied to your next significant working session. Run it as an experiment. The difference in the quality of the conversation will be immediately apparent.

Next in this series: Article 6 — The Design: How to Build an Organization Where Human Capability and AI Capability Grow Together.

References

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.

Hadwin, A. F., Järvelä, S., & Miller, M. (2011). Self-regulated, co-regulated, and socially shared regulation of learning. In B. J. Zimmerman & D. H. Schunk (Eds.), Handbook of self-regulation of learning and performance (pp. 65–84). Routledge.

Stasser, G., & Titus, W. (1985). Pooling of unshared information in group decision making: Biased information sampling during discussion. Journal of Personality and Social Psychology, 48(6), 1467–1478. https://doi.org/10.1037/0022-3514.48.6.1467

Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., & Malone, T. W. (2010). Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004), 686–688. https://doi.org/10.1126/science.1193147

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–40). Academic Press. https://doi.org/10.1016/B978-012109890-2/50031-7

Charles Good works with organizations that have already solved the adoption problem and are now asking the harder question. As President of the Institute for Management Studies, he reaches over 20,000 professionals annually. His Outlearn Loop framework, built on behavioral and learning science, is the architecture organizations use to design AI integration that builds human capability rather than substituting for it. He writes The AI Capability Playbook and The Performance Playbook on Substack and hosts The Good Leadership Podcast (Apple / Spotify / YouTube).

Six articles. One question every organization is avoiding: Are you building capability or just building dependency? Articles 1 through 3 of this series established the problem, the leadership responsibility, and the individual habits that address it. This article addresses the measurement gap that makes the problem invisible, while providing organizational leaders with the metrics that tell the truth about which future they are building.

The Dashboard That Looks Perfect

Consider an organization twelve months into a successful AI rollout.

The numbers tell a compelling story. License utilization is at 84%. Average task completion time is down 31%. Output volume per employee is up 27%. AI-assisted work scores consistently higher on quality reviews than unassisted work. The CIO presents to the board: the investment is working.

Every one of those numbers is real. Every one of them is measuring something true about the organization’s current state.

And not one of them can tell you whether the people producing those outputs are becoming more capable or more dependent. Not one of them can tell you whether the organization’s collective judgment is compounding or quietly eroding. Not one of them can tell you which future the organization is building.

This is not a failure of data collection. It is a failure of design. The organization built a measurement system that answers the question it knows how to ask, which is the following:

Are our people using AI effectively?

However, it left unanswered the question that determines competitive durability:

Are our people getting better because of how they are using it?

Here is the good news: there is one question you can ask every professional in your organization once per quarter that serves as a leading indicator of the capability gap before it shows up in performance data. We will get to it. But to understand why that question works, you first need to see why the metrics you are currently tracking cannot.

Why the Current Metrics Are Structurally Blind

The problem is not that organizations are tracking bad metrics. It is that the metrics they are tracking are structurally incapable of seeing what matters most.

Every standard AI adoption metric (utilization rate, task completion time, output volume, quality scores on AI-assisted work) shares a common characteristic: it measures the performance of the human–AI system, not the performance of the human within it. It tells you what the combination produces, not what the human alone can produce, what the human’s judgment contributes, or whether that contribution is growing or atrophying over time.

This distinction matters because the long-term value of any human–AI partnership is determined by the quality of the human component. A system in which AI capability increases and human capability remains static, or declines, is not stable. Every efficiency gain from AI doing what humans used to do is also a practice opportunity that humans are no longer getting. Over time, enough missed repetitions produce the pattern this series has described: high adoption, low transformation, and a workforce that performs well with the tool and discovers the gap only when the tool is not there.

Research on cognitive offloading makes this mechanism visible. Sparrow, Liu, and Wegner showed that when people expect information to be stored externally and easily retrievable (e.g., search engines), they remember less of the content itself and focus primarily on where to find it. Risko and Gilbert’s review extends this pattern: as people routinely rely on external systems to carry cognitive load, internal storage and processing demands shift.

When AI consistently handles tasks like analysis, drafting, and information retrieval at work, the professional skills once honed through regular practice will diminish from disuse.

Standard metrics don’t just miss this decline; they hide it by focusing on what the system produces instead of what people are actually putting into it.

Adoption dashboards often focus on the wrong metric. What truly matters isn’t just the task being completed, but the person completing it. The key question is whether each repetition makes that person more skilled or more reliant on the system.

One clarification is essential before building the measurement framework, because applying it to the wrong category of work would make it useless. Article 3 in this series drew a line between two categories of AI-delegated tasks.

Category 1 work is rule-based and fully specifiable, such as formatting, scheduling, summarizing transcripts, generating boilerplate, and pulling data from structured sources. The output is identical regardless of who or what produces it. Therefore, automating this work completely is the right decision. The capability you stop exercising on Category 1 tasks was never building anything worth keeping, and the efficiency gains are real and permanent. No measurement framework should penalize an organization for offloading work that carries no capability cost.

Category 2 work is different in kind, not merely in degree. It encompasses tasks such as strategic analysis, novel problem framing, high-stakes client judgment, and decision-making under genuine uncertainty. In this domain, the quality of the output is determined by the depth of a professional’s thinking. The capability that makes this work valuable is not a static asset; it is practice-dependent. It compounds with use and atrophies with disuse, gradually, invisibly, and consequentially.

Everything that follows in this article, every measurement gap, every metric, every diagnostic, applies to Category 2 work. That is where the capability crisis lives. That is where the current metrics are blind. And that is where the cost of not measuring is highest, precisely because the erosion is hardest to see on the dashboard that shows adoption is going well.

The Three Measurement Gaps

To develop a measurement system that accurately reflects capability development, we must first identify what the current system fails to see. These blind spots can be categorized into three distinct gaps.

Gap One: The Assisted/Unassisted Gap

A crucial measurement is missing from nearly every AI integration: a regular comparison of what people produce with AI versus without it. This isn’t about creating an elaborate research study. Instead, it should be a deliberate operational practice. Periodically, give professionals meaningful tasks to complete under conditions where AI is not the default, and then track these results over time.

The gap between assisted and unassisted performance is your primary capability metric. A narrow gap means human capability is developing alongside AI-enabled productivity. In other words, the person is genuinely learning through the work, not just producing through the tool. A widening gap means capability is migrating into the tool, which means the productivity gains are real, but the person’s independent capability is not keeping pace and likley declining.

Most organizations have never measured this capability gap, leaving them without a crucial baseline. The first step is to establish one, not as a punitive assessment, but as a developmental diagnostic. This process reveals where an individual’s genuine capabilities lie and shows the organization where investment in development is most needed.

Gap Two: The Reasoning Gap

While output quality metrics can assess the final product, they fall short of measuring true understanding. These metrics cannot determine whether a person understands why their work is effective, can defend it under scrutiny, can adapt it to new contexts, or can identify when a confident-sounding AI has made a mistake.

The reasoning gap is the difference between a professional who produces good work because AI produced it and a professional who produces good work because their judgment, informed by AI, shaped it. Both can produce similar outputs on any given day. Their trajectories over twelve to twenty-four months are entirely different.

Measuring the reasoning gap requires a different kind of evaluation, one that asks people to explain their work, such as brief verbal or written explanations of the reasoning behind significant AI-assisted outputs. Questions that probe whether the professional understands the logic of the recommendation, the assumptions built into the analysis, and the conditions under which the conclusion would change. This is not an audit. It is a development practice that makes visible what output metrics are not able to see.

Gap Three: The Transfer Gap

The most consequential capability, and also the hardest to measure, is transfer: the ability to apply knowledge and judgment in situations that are novel, complex, or different from the conditions in which the skill was originally developed.

Transfer, the ability to apply knowledge to new situations, is what separates genuine expertise from sophisticated tool fluency. A professional with true expertise can navigate challenges their AI training didn’t cover. In contrast, someone with mere tool fluency produces excellent results only in familiar conditions, discovering the limits of their capabilities when circumstances change.

The transfer gap is measured by giving professionals problems that are related to their field but structurally different from their usual tasks. These are problems that demand judgment rather than simple retrieval and adaptation rather than rote application. How well they solve these problems, both with and without AI assistance, reveals more about their true capabilities than their performance on any number of familiar tasks.

A senior financial analyst at a mid-sized asset management firm told me a version of this story earlier this year. She had been using AI-assisted modeling tools for about fourteen months. Her quarterly reviews were strong, her output had increased, her models were cleaner, and her manager described her as one of the team’s most effective AI adopters.

Then the firm’s AI platform went down for three days during a critical client deliverable window.

She described the experience as “like forgetting how to drive.” She could still build a model, but the intermediate steps she used to move through instinctively now required conscious effort. Assumptions she once pressure-tested by hand, she found herself waiting to verify with a tool that was not available. The work she produced in those three days was competent. However, it was not close to the quality her AI-assisted output had led everyone, including her, to believe she could produce.

No metric in her organization would have predicted that gap. Her adoption scores were excellent. Her output quality was consistently high. Every number on the dashboard said she was thriving. What no number captured was that her independent judgment had been quietly atrophying through fourteen months of under-use.

Her work, by the way, was entirely Category 2. Every model she built required judgment about assumptions, risk weighting, and client-specific context that no AI could evaluate without her domain expertise informing it. The AI made her faster at producing that judgment, but it did not make her better at forming it. In fact, fourteen months of letting the tool lead the way had quietly widened the gap between what the system could generate and what she could produce on her own.

The three measurement gaps described above are what made her situation invisible. What follows is a framework for making it visible, before the platform goes down and the gap reveals itself.

The Measurement Framework

What follows is a practical framework for measuring capability development alongside AI productivity applied specifically to Category 2 work, where the professional's judgment is the differentiating value and where capability erosion carries the highest organizational cost.

Metric 1: The Capability Gap Index

What it measures:
The difference in performance quality between AI-assisted and unassisted work for a given professional or team, tracked over time.

How to implement it:
Once per quarter, present each professional with one significant task in their domain, comparable in complexity to their normal work,, in conditions where AI is not the default tool for the initial analysis. Track the output quality against their AI-assisted baseline. The ratio is their Capability Gap Index. A stable or narrowing ratio indicates capability is developing, while a widening ratio indicates dependency is building.

What to look for:
The trend matters more than the absolute level. A professional whose unassisted performance is improving, even if it remains below AI-assisted performance, is developing genuine capability. A professional whose unassisted performance is static or declining while AI-assisted performance rises is building dependency.

Metric 2: The Reasoning Depth Score

What it measures:
The quality of a professional’s understanding of the work they produce with AI assistance, their ability to explain, defend, and adapt it without the tool.

How to implement it:
After significant AI-assisted work, conduct a brief structured conversation (10 to 15 minutes) in which the professional explains the reasoning behind their output and evaluate against three criteria:

• Can they explain why the approach was chosen?

• Can they identify the key assumptions and where they might break down?

• Can they describe what they would do differently if one of the key conditions changed?

Score on a simple three-point scale across all three criteria.

What to look for:
Professionals who genuinely understand their AI-assisted work will score consistently across all three criteria. Professionals who are primarily reviewing and approving AI output will struggle with the third criterion, the adaptive reasoning question, because they have not built the underlying mental model that makes adaptation possible.

Metric 3: The Transfer Performance Index

What it measures:
A professional’s ability to apply their domain knowledge and judgment to novel situations that differ structurally from their familiar work.

How to implement it:
Twice a year, challenge professionals with cases or problems adjacent to their field but structured in ways different from their usual tasks. This could involve presenting a familiar type of problem in an unfamiliar context, or an unfamiliar problem within a familiar context. By evaluating their performance both with and without AI assistance, you can measure the gap between their ability to transfer skills and their standard performance. This gap represents their transfer index.

What to look for:
Professionals who possess true adaptive expertise will show a smaller decline in performance on transfer tasks than those whose abilities are merely tool-dependent. Their unassisted performance on these new tasks will be significantly more robust, serving as the clearest diagnostic for distinguishing genuine expertise from sophisticated tool fluency.

Metric 4: The Development Trajectory Score

What it measures:
The direction and rate of change in a professional’s genuine capability over time, independent of AI productivity gains.

How to implement it:
Combine the Capability Gap Index, Reasoning Depth Score, and Transfer Performance Index each quarter to create a single directional indicator. This synthesis will show whether the individual’s unassisted performance, reasoning depth, and ability to transfer skills are improving, remaining stable, or declining. Together, these metrics reveal if their genuine capability is on an upward, flat, or downward trajectory.

What to look for:
The aim isn’t for every professional to perform at their peak at all times. Rather, the goal is to ensure that professional capabilities continue to improve. Organizations should focus on building human capital in tandem with AI productivity, ensuring they enhance each other rather than serve as replacements.

What to Do With the Data

Simply measuring performance is just documentation; true management involves using those measurements to drive change. The capability metrics discussed earlier are only valuable if they shape work structures, guide development investments, and inform leadership’s decisions on AI integration.

Here are three specific applications worth considering.

Pinpoint Where AI Is Replacing, Not Enhancing, Skills

The Capability Gap Index is designed to reveal which teams and functions are most at risk. It highlights areas where the performance gap between AI-assisted and unassisted work is widening—in other words, where employees are becoming more dependent on AI rather than more capable because of it. Interestingly, these may not be the teams with the lowest AI adoption rates; they could even be the ones embracing it most enthusiastically.

The question is: how should the work be structured so that using AI builds human capability rather than substituting for it? The five habits from Article 3, applied at the team level, are the starting point.

Target Development Investment with Precision

Instead of just looking at broad skills, the Reasoning Depth Score and Transfer Performance Index pinpoint exact areas where individuals and teams can improve.

For instance, someone might be great at explaining things and spotting assumptions but struggle with thinking on their feet. This highlights a specific area for them to work on. Likewise, a team might be solid when tackling familiar problems but falter when faced with something new, showing a gap in how they work together.

This level of precision allows investment in development to be highly targeted rather than broad-brushed. It marks a fundamental shift from generic directives like “everyone needs AI training” to specific, actionable insights such as, “this individual requires deliberate practice in adaptive reasoning to improve their declining transfer performance.”

Proactively Evaluate AI Integration Decisions

Article 3 in this series, “Automate Everything You Can,” introduced three diagnostic questions to classify capabilities before automation. This measurement framework now provides empirical support for those questions.

Before expanding AI integration into any new domain, first establish a baseline by calculating the Capability Gap Index and Development Trajectory Score for the professionals involved. After six months of expanded AI use, re-measure these scores. The comparison will reveal whether the integration is enhancing or eroding the essential human capabilities your organization relies on.

This approach transforms capability measurement from a retrospective audit into a proactive design tool. It’s the kind of instrument that empowers organizations to consciously build their desired future, rather than discovering an undesirable one when it’s already too late.

The Number Worth Tracking Above All Others

Even if your organization isn’t ready to implement the full framework, there’s one metric you should start tracking immediately. It’s simpler to collect than other data points and offers more diagnostic value than any current adoption metric.

Ask every professional in your organization, once per quarter, a single question:

On a task you completed this week that required significant judgment ( e.g., an analysis, a complex decision, etc.), did you form your own position before using AI, or did AI go first?

Track the ratio of “I went first” to “AI went first” across the organization. Track it by team, by function, by seniority level. Track it over time.

Think of this ratio as an early warning system for a future skills gap. It shows you whether your organization is focused on improving the skills you already have or replacing them completely. And it gives you this information before any performance issues even show up. This allows you to step in and fix the problem while it’s still just a trend, not a widespread, ingrained habit.

A dashboard tells you where your organization has been, but the right metric tells you where it’s going. That’s the measurement worth having.

The Measurement Is Not the Goal

A final point worth making explicit is that measuring capability development is not the goal; building it is.

The framework in this article is valuable because it makes visible what is currently invisible, the slow drift toward dependency that standard adoption metrics can’t detect. But measurement is only useful to the extent that it informs design. Every metric in this framework should feed back into decisions about how work is structured, how AI is integrated, how development is invested, and how leaders are building the conditions for genuine capability growth.

Future industry leaders won’t be defined by their sophisticated measurement systems. Instead, they will be the organizations that use measurement to build a resilient, integrated system where human and AI capabilities grow in tandem. By compounding each other’s strengths, this synergy creates a system that becomes more valuable, not more fragile, over time.

Measuring the gap is how you know whether you are building that system or drifting away from it.

If this framework provides the metrics your organization has been lacking, the most effective first step is to establish a baseline Capability Gap Index for your most critical functions. This should be done before the gap widens to a point where it cannot be closed by development alone.

Next in this series: Article 5 — The Team Problem: What Happens to Collective Intelligence When AI Does the Thinking.

References

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.

Gartner. (2025). CIO and technology executive survey. Gartner Research.

McKinsey & Company. (2025). The state of AI: How organizations are rewiring to capture value. McKinsey Global Institute.

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604. https://doi.org/10.1037/0278-7393.4.6.592

Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745

Winne, P. H., & Hadwin, A. F. (1998). Studying as self-regulated learning. In D. J. Hacker, J. Dunlosky, & A. C. Graesser (Eds.), Metacognition in educational theory and practice (pp. 277–304). Lawrence Erlbaum Associates.

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–40). Academic Press. https://doi.org/10.1016/B978-012109890-2/50031-7

Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70. https://doi.org/10.1207/s15430421tip4102_2

Charles Good works with organizations that have already solved the adoption problem and are now asking the harder question. As President of the Institute for Management Studies, he reaches over 20,000 professionals annually. His Outlearn Loop framework, built on behavioral and learning science, is the architecture organizations use to design AI integration that builds human capability rather than substituting for it. He writes The AI Capability Playbook and The Performance Playbook on Substack and hosts The Good Leadership Podcast (Apple / Spotify / YouTube).

Six articles. One question that most professionals are avoiding: Are you building capability or just building dependency?

The Moment You Realize You’re Dependent

There is a moment that almost every professional who uses AI regularly has experienced, and almost nobody talks about.

It happens when you’re in a meeting, or on a call, or standing in front of a client, and someone asks you a question you should be able to answer. A question that sits squarely in your domain. A question that, six months ago, you would have answered without hesitation.

And you feel the reach.

Not for the answer itself but for the tool. The reflex to open a window, run a query, let the interface do the work of retrieval that used to be yours. The answer is in there somewhere, you’re sure of it. It’s just not as accessible as it once was.

This article is about that moment when you reach for a tool instead of relying on your own capabilities.

This doesn’t mean you’re incompetent, nor does it mean AI is inherently bad. It simply means you’ve been using AI in a way that subtly transfers your capabilities from yourself to the tool. This gradual shift, if sustained, represents the most significant professional development decision you are currently making, likely without even realizing it.

Articles 1 and 2 of this series addressed organizations and leaders. This article is for the individual professional navigating this in real time, in your actual workflow, with real deadlines and real pressure to produce.

The question is not whether to use AI. The question is how to use it in a way that makes you better rather than more dependent. Those are two entirely different strategies. And the one you’re currently using, whether by design or default, is determining your professional trajectory.

What the Research Predicts About Your Current AI Habits

Before the practical playbook, I want to make sure you understand the stakes, because I’ve found that professionals tend to underestimate this risk until they see the mechanism in action.

Barry Zimmerman's self-regulated learning research describes how genuine capability develops through a three-phase cycle. In the Forethought phase, you define your goals and map out a strategy before you begin. In the Performance phase, you actively monitor your progress and make real-time adjustments. In the Self-Reflection phase, you compare the outcome to your initial plan and decide what to do differently next time.

Each time you complete the full cycle on a meaningful task, you improve incrementally. However, the cycle breaks whenever you skip a phase. This can happen when you outsource forethought to AI, bypass performance monitoring with AI assistance, or skip self-reflection because there was no human performance to evaluate. While you may produce an output, you fail to develop your own capability. Multiply that by every AI-assisted task you complete in a week. Now multiply it by 52 weeks.

Now ask yourself: over the past year, how many times did you complete the full self-regulation cycle on a significant professional task versus how many times did you produce output with AI assistance without completing any phase of it?

That ratio is your professional development ratio. And for most knowledge workers right now, it is moving in the wrong direction without anyone, including them, noticing.

The Sparrow, Liu and Wegner research on cognitive offloading adds the neurological layer: when you consistently expect external access to information and thinking, your recall and retrieval of that information internally declines. You remember where to find things rather than knowing them. You know how to prompt rather than how to think. The capability migrates from you to the system.

You are not just using AI to do your work. You are training yourself — through every interaction — in a particular relationship with your own thinking. The question is whether that training is making you sharper or making you dependent.

Does It Actually Matter Where the Thinking Lives?

The comparison to a calculator is fair when it comes to simple arithmetic. The answer to 347 multiplied by 28 is the same whether it’s calculated by a human or a tool. Offloading this task has no real cost, as the essential human capability, the judgment of what to calculate and how to interpret the answer, remains entirely untouched.

But judgment isn’t arithmetic. When you walk into a difficult conversation, a high-stakes negotiation, or a genuinely novel problem, the output is not the same regardless of who produces it. It depends entirely on the depth of your own thinking.

When you consistently let AI take the lead, it isn’t your memory that suffers. Instead, what withers is your ability to form an independent starting point: to read a situation and generate your own perspective before a prompt tells you what to think.

That capacity is built through use. It degrades through substitution.

Your own judgment matters most in the very moments AI is least helpful: when the situation is genuinely new, the stakes are real, and you’re expected to think on your feet. In these moments, your judgment is the only thing you have.

Knowing where to find answers is powerful. However, there are times when technology isn’t available, or when you’re expected to rely on your own internalized knowledge. In those moments, the only thing that is available are the mental models you have built and cultivated over time.

Two Task Types: Automate One, Protect the Other

Not every task carries the same stakes for your capability development. The mistake most professionals make is treating all AI-assisted work as equivalent, which is inefficient in one direction and dangerous in the other.

Certain tasks lend themselves to complete automation. Formatting, scheduling, summarizing transcripts, generating boilerplate, and pulling data are all rule-based operations; the output is identical regardless of who performs them. Offloading these tasks entirely isn’t creating a dependency; it’s gaining efficiency. The capabilities you cease to exercise in these areas were never building anything of lasting value in the first place.

For other tasks, your judgment is the real value. Strategic analysis, client relationships, reframing complex problems, high-stakes communication, and making decisions under uncertainty; these are all areas where the depth of your thought directly shapes the quality of your work. If you consistently let AI take the lead on these tasks, you aren’t just delegating the execution; you are outsourcing the cognitive effort that builds the very mental models required for sound judgment.

Expertise in judgment-intensive work is not a fixed asset; it is sustained by practice. Without it, skills gradually atrophy, not in a way that is visible from week to week, but measurably over a year and consequentially over two.

To classify the work, ask yourself a simple question: if an AI produced this output and you approved it, could a discerning client or colleague tell it wasn’t created by a human?

• If the answer is no, it’s Category 1 work. Automate it completely and confidently.

• If the answer is yes, it’s Category 2 work. Every time you engage with it, you are building capabilities that truly matter.

Four Questions That Reveal Your Real Capability

Before I introduce the habits for the Category 2 work that matters most to your career, you first need an honest diagnostic. Answer these four questions based on what you actually do, not what you intend to do.

When you face a significant problem at work, what happens first? Do you sit with it, even briefly, even uncomfortably, and form your own initial read? Or does the reflex now run directly to the tool, before you’ve done any independent thinking?

When AI produces an output you’re going to use, do you know why it’s right? Not whether it looks right, or whether it passes a surface check. Do you understand the reasoning well enough to defend it under pressure, adapt it to a slightly different situation, or catch it when it’s confidently wrong?

When you’re asked to explain your work (your analysis, your recommendation, your reasoning), can you do it without the tool? Not reconstruct it with help. Explain it. From your own understanding.

When was the last time you struggled productively on a significant professional task where the difficulty was the point, not something to be eliminated?

If the honest answers to those questions are uncomfortable, that discomfort is useful information. It’s the signal that the shift has already begun, and that designing against it, deliberately, starting now, is worth doing.

Five Habits That Turn AI Into a Capability Engine

These aren’t abstract principles but specific behavioral patterns that often look quite different in practice from what most professionals are used to.

Habit One: Think First, Then Ask

This is the foundational habit and the hardest to build because it runs directly against the efficiency logic that makes AI so appealing. The fastest path to an output is to ask it a question immediately. But the fastest path to improving your capability is to not do that.

Before you open any AI tool on a task that matters (a strategic analysis, a significant communication, a complex problem) spend five to ten minutes forming your own position first. Not a polished answer. A rough one. What do you actually think? What’s your initial read of the situation? What would you recommend if AI weren’t available?

Then ask AI the question and compare your position to what AI produces. Afterward, ask: where were my gaps? Where was my reasoning weak? Where did AI surface something I hadn’t considered, and what does that tell me about what I need to develop?

This “think first, query, then compare” sequence is the generation effect in action. When you make an initial attempt, you begin to encode the learning. This makes the subsequent comparison with an AI’s output genuinely educational, rather than merely confirmatory. The gap between your version and the AI’s is the most specific, actionable feedback you can get. Yet, most professionals miss out on this learning opportunity by turning to AI from the start.

Habit Two: Know Why the Answer Works

One level of AI fluency is operational: knowing how to generate good outputs from the tool. A deeper, more developmental fluency involves understanding the reasoning behind the output: knowing why it’s good, where it might falter, and how to adapt it effectively as circumstances change.

The first version makes you productive, but the second makes you capable. This distinction is most critical when the stakes are highest, such as when a client pushes back, a situation is novel, or your output must be defended and adapted in real time under scrutiny.

Before finalizing any significant AI-generated output, ask yourself one crucial question:

Could I explain the reasoning behind this to someone who hasn’t seen the AI’s response?

If the answer is yes, you’ve developed a true understanding. If it’s no, you’ve merely produced an output. While both may appear identical at first glance, only the former leads to compounding knowledge and growth.

Habit Three: Use AI to Sharpen Your Thinking, Not Replace It

This is the distinction Articles 1 and 2 called the Sharpening Pattern versus the Replacement Pattern. At the individual level, it looks like this:

The replacement pattern: encounter a problem → query AI → review output → submit. Your cognitive contribution is editorial. The thinking was outsourced.

The sharpening pattern: encounter a problem → form your own analysis → use AI to challenge it, find the holes in it, surface what you missed, steelman the counterarguments → update your thinking where it’s earned → submit. Your cognitive contribution is generative. AI made your thinking better.

In the sharpening habit, AI is functioning exactly as Ethan Mollick describes in Co-Intelligence as a genuine thinking partner. The difference is in the order of operations. Human thinking first. AI pressure-testing second. That sequence is what makes AI use developmental rather than substitutional.

Here’s the practical implication: when using AI for a task that requires judgment, don’t ask it for the answer. Instead, ask it to challenge the answer you’ve already formed.

A person using the sharpening pattern might ask, “Here’s my analysis; what am I missing?” In contrast, someone using the replacement pattern would ask, “What’s the best analysis of this situation?” It’s the same tool and the same task, but the effect on your long-term capability is entirely different.

Habit Four: Regular Reps Without the AI Safety Net

If you never work without AI support, you genuinely don’t know where your capability stands. You know where your AI-assisted capability stands. Those are not the same thing, and the gap between them becomes invisible only when the tool isn’t available or when the situation requires judgment that the tool can’t provide.

Incorporate deliberate scaffold-removal into your practice. While you don’t need to do it constantly, you have work to do, after all, do it regularly enough to maintain an honest assessment of your current capabilities.

For instance, you might draft a critical email or document from scratch before using AI to refine it. Or, when preparing for a key meeting, you could formulate your entire argument independently before asking AI for feedback. You could even challenge yourself to tackle one significant problem each week by sitting with it before turning to any tool for help.

The discomfort that accompanies this practice isn’t a problem to be solved—it’s a sign that it’s working. This aligns with Robert Bjork’s research on desirable difficulties, which indicates that the conditions that feel most challenging in the moment yield the greatest capability over time.

The test: if working without the scaffold feels uncomfortable, that is where your development is happening. If it feels easy, the capability is already built."

Habit Five: Study the Gap Between You and the Model

After completing any significant work with AI, take a moment to ask a question most professionals overlook: Where was the gap between my initial thoughts and the AI’s output?

If AI consistently surfaces things in a particular domain that you hadn’t considered, that domain is a development priority. If AI’s framing of a problem is consistently sharper than your initial framing, your problem-structuring skills are a development priority. If AI’s counterarguments to your recommendations are consistently stronger than the objections you anticipated, your anticipation of challenge is a development priority.

The gap between your own thinking and an AI’s output creates a uniquely personalized development curriculum. While most professionals use this gap for a simple quality check, those building the most durable skills use it as a mirror, reflecting opportunities for their own growth and learning.

Two Professionals, Same Tools, Very Different Futures

I want to make the long-term benefits of these habits concrete. I believe most professionals neglect to build them because the consequences of doing so are invisible in the short term.

Consider two professionals with the same role, the same AI tools, the same workload, and similar starting capabilities.

Professional A: She consistently produces excellent work by leveraging AI, which she then reviews and approves. Over the past two years, the quality of her output has been high and continues to improve, demonstrating an exceptional fluency with AI tools. However, because she is no longer regularly exercising her own judgment, pattern recognition, and independent thinking, these underlying capabilities have begun to plateau and atrophy. As a result, she is becoming increasingly dependent on AI.

Professional B: By adopting the sharpening pattern and the five habits mentioned, the quality of her output has shown a clear upward trajectory. Over 24 months, her underlying capabilities have also expanded. She is becoming not only faster but also sharper, continually enhancing her overall competence.

For the first eighteen months, Professionals A and B appear indistinguishable. Their outputs are identical, their performance seems matched, and their adoption metrics are the same.

Then, at the two-year mark, a situation arises that demands their immediate expertise and judgment. It might be a novel client crisis, a strategic inflection point, or a high-stakes decision in a room where their tools are unavailable and the client is watching.

Unable to access her usual tools, Professional A must rely on her internalized knowledge. As she shares her expertise with the people in the room, she uncovers an unexpected gap in her understanding.

Meanwhile, Professional B draws upon her own capabilities, which have been steadily building all along.

These five habits won’t produce better outputs this week, but they will produce a fundamentally different professional in two years. This is the power of the compounding effect, which only works if you start before you can see the return.

The One Question That Reveals Who You’re Becoming

Let me bring you back to where we started.

You’re in a meeting. Someone asks a question that sits squarely in your domain, the kind of judgment call that no one else in the room can answer the way you can. And instead of answering it, you feel the pull. The reflex to reach for the tool. The answer is in there somewhere. It just isn’t in you the way it used to be.

This moment of reckoning is coming for every professional who defaults to letting AI replace their judgment on complex tasks. It’s not because AI is inherently dangerous, but because human capability is built on practice. And with each judgment call quietly delegated away, that practice, and the skill it maintains, erodes.

The solution isn’t to use AI less. It’s to use it in the right sequence: think for yourself first, then turn to AI. The gap between your initial thought and the AI’s output becomes your roadmap for development.

Before you move on, there’s one thing I want you to do.

Open your calendar and find your next significant meeting, presentation, or decision, the kind of task where your judgment is the real value you bring. Block out fifteen minutes before it and label the event: My Position First.

In those first fifteen minutes, before you turn to any other tool, capture your thoughts, analysis, recommendations, and any outstanding questions. These notes can be rough, even incorrect; what matters most is the act of creating them yourself.

Then, use AI to challenge your own thinking. Let it probe your assumptions, find the gaps in your logic, and stress-test your reasoning. Where the counterarguments hold true, refine your position accordingly.

Adopting this single habit for your next important task is worth more than reading this article twice.

Using AI in the right sequence is the strategy.

You can't outperform what you haven't outlearned.

Six articles. One question every professional is avoiding: Are you building capability or just building dependency?

If this reframes how you’re thinking about your own AI habits, share it with one colleague whose professional development you care about and ask them which professional they’re becoming.

Next in this series: Article 4 — The Measurement Problem: How to Track Whether AI Is Building Capability or Dependency.

References

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.

Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher et al. (Eds.), Psychology and the Real World (pp. 56–64). Worth Publishers.

Fogg, B. J. (2019). Tiny habits: The small changes that change everything. Houghton Mifflin Harcourt.

Gollwitzer, P. M. (1999). Implementation intentions: Strong effects of simple plans. American Psychologist, 54(7), 493–503.

Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W., & Wardle, J. (2010). How are habits formed: Modelling habit formation in the real world. European Journal of Social Psychology, 40(6), 998–1009.

Mollick, E. (2024). Co-intelligence: Living and working with AI. Portfolio/Penguin.

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604.

Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778.

Zimmerman, B. J. (1989). A social cognitive view of self-regulated academic learning. Journal of Educational Psychology, 81(3), 329–339.

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–39). Academic Press.

Charles Good works with organizations that have already solved the adoption problem and are now asking the harder question. As President of the Institute for Management Studies, he reaches over 20,000 professionals annually. His Outlearn Loop framework, built on behavioral and learning science, is the architecture organizations use to design AI integration that builds human capability rather than substituting for it. He writes The AI Capability Playbook and The Performance Playbook on Substack and hosts The Good Leadership Podcast (Apple / Spotify / YouTube).

The Wrong Conversation in the Right Room

Picture a leadership team meeting about AI strategy.

The questions on the table are familiar. Which tools should we standardize? How do we drive adoption? What training do we need to close the skills gap? How do we measure ROI?

These are reasonable questions, but they are also the wrong questions.

Underneath every one of them is an assumption that nobody in the room has made explicit: that the leader’s job in the AI era is to deploy capability, to get the right tools into the right hands and get out of the way.

That assumption is understandable. It’s how technology adoption has worked for decades. You introduce the tool, train people to use it, measure usage, and declare success.

But AI is not like previous tools, and the leaders who treat it like one are going to discover, too late, and at high cost, that they optimized for the wrong outcome.

Article 1 of this series established the problem: most organizations are measuring AI usage when they should be measuring whether their people are getting smarter. Those are not the same metric. And in most organizations right now, they are moving in opposite directions.

This article is about who is responsible for closing that gap.

What Great Leaders Have Always Actually Done

Before I make the case for a new competency, I want to make a case for an old one, because I think the AI era is clarifying something about leadership that has always been true but rarely stated.

Over the past decade at the Institute for Management Studies, I’ve observed that the most exceptional leaders I’ve worked with all share one defining trait. It isn’t their strategic vision, although many are visionaries. Nor is it their communication skills, though most are articulate. It’s not even their technical expertise within their field.

It is this: they are architects of environments that make people more capable.

They cultivate an environment, defined by clear expectations, stimulating challenges, constructive feedback, and developmental pressure, that compels the people around them to grow, not just to perform.

Peter Drucker saw this clearly fifty years ago when he wrote that the leader’s job is not to do but to develop. John Wooden, who won ten national championships at UCLA, kept a practice journal that documented not what his players did but what they learned. He was not coaching basketball. He was designing a capability development system that used basketball as the medium.

The leaders who will navigate the AI era most effectively are not the ones who understand the technology best. They are the ones who understand this principle most deeply and apply it with the same intentionality to AI integration that Wooden applied to practice design.

The question is not whether your people are using AI. The question is whether the environment you’ve designed is causing them to grow through using it or simply to produce through it.

Those are two fundamentally different design briefs. And most leadership teams have only written one of them.

The Competency Gap Nobody Is Naming

In the past two years, organizations have invested heavily in AI literacy, teaching people what the tools can do, how to prompt effectively, and how to integrate AI into existing workflows. That investment is real and mostly appropriate.

But there is a second competency that has received almost no attention. I want to describe it precisely, because seeing it clearly is the first step toward developing it. It is the intentional shaping of workflows, AI integration, and performance metrics to amplify human skills, ensuring that as AI productivity increases, the people driving that productivity are getting better, not more dependent.

This is not a technical skill. It does not require deep AI expertise. It doesn’t require a background in learning science, though the science makes it more precise. It requires leaders to ask different questions about the environments they are responsible for.

Not just: Are our people using AI effectively?

But: Is the way our people are using AI making them more capable, or more dependent?

Not just: Are our outputs improving?

But: Are the people producing those outputs improving?

Not just: Are we getting more from our people with AI?

But: Are we building people who can get more from AI because they are getting better themselves?

These questions haven’t been on most leadership agendas, but they should be. The leaders who prioritize them now will build organizations with compounding capabilities, leaving others to plateau in productivity.

Three Leadership Failures the AI Era Is Exposing

To understand what capability environment design requires, it helps to understand the three specific ways leadership is currently failing to provide it.

Failure One: Mistaking output quality for capability development.

When the quality of AI-assisted work improves, it’s natural for leadership to conclude that the underlying capability has also improved. After all, if the tool makes the work better, and the work is what we measure, then it follows that things are getting better.

This is the most seductive failure because it is partially true: output quality can genuinely improve. However, output quality and human capability are not the same thing. In the early stages of AI adoption, the two can improve in tandem, masking the divergence building just beneath the surface.

For years, autopilot-assisted flights were safer and more precise than manually flown ones. The output metric, flight safety, was improving. The capability metric, pilot proficiency in manual flight, was deteriorating. Nobody saw the problem until conditions arose that required the capability the output metric couldn’t measure.

Your AI adoption dashboard is a flight safety metric. You also need a manual flight metric. Unfortunately, most organizations have only built one of them.

Failure Two: Designing for efficiency without designing for development.

Most AI workflow integration has been optimized for one variable: speed. How do we get the task done faster? How do we reduce the friction between the need and the output? How do we make the tool as easy to use as possible?

These are legitimate efficiency goals. They become leadership failures when they are the only goals, when no one in the design process asks, ‘What does this workflow do to the person doing it repeatedly over 12 months?’

This is not a new failure. Frederick Winslow Taylor made the same mistake in 1911 when he designed factory workflows that maximized output per worker while systematically eliminating the skilled judgment those workers had spent years developing. He got his efficiency gains. He also had a workforce that was entirely dependent on the system he designed and incapable of adapting when it changed.

AI-optimized workflows that never require the human to think first are Taylor’s assembly line with better software.

The leader’s job is not to eliminate friction. It is to eliminate unproductive friction while deliberately preserving the friction that drives development. That distinction requires judgment that efficiency metrics alone cannot provide.

Failure Three: Abdicating the developmental conversation.

In most organizations, the question of how people are developing through their AI use has no owner. The CHRO owns adoption metrics. The CTO owns tool deployment. The CFO owns ROI measurement. Nobody owns the question of whether the people using these tools are becoming more or less capable over time.

This gap won’t close on its own. It requires a leader who takes direct ownership of capability development, placing it on the agenda right alongside productivity. This leader must be willing to ask the difficult questions: How is AI-assisted work truly affecting the people performing it? And they must design the evaluation and development systems that reveal the answer before the cost becomes too high.

The greatest challenge facing most organizations isn’t the gap between AI and human capabilities. Rather, it’s the growing divide between the urgent push for AI deployment and the lagging development of human skills to manage it. This gap was created by leaders, and only leaders can close it.

---

What the Research Says About Capability-Building Environments

The learning science here is both well-established and almost entirely absent from the AI leadership conversation.

Edward Deci and Richard Ryan’s Self-Determination Theory, one of the most widely replicated frameworks in motivational psychology, outlines three essential conditions for developing genuine capability: autonomy, competence, and relatedness. Autonomy is the feeling of ownership over your thoughts and decisions. Competence is the experience of mastery achieved through meaningful effort. Finally, relatedness is the sense that your growth and development are valued by those around you.

Now, let’s examine what most AI-optimized work environments provide against those three conditions.

• Autonomy is reduced when AI consistently generates the first draft of thinking. The human is responding, not initiating. Approving, not owning. That shift is subtle but cumulative and Deci and Ryan’s research predicts it will reduce intrinsic motivation and the drive toward genuine mastery over time.

• Competence requires what Robert Bjork at UCLA calls “desirable difficulties,” conditions that feel harder in the moment but drive deeper encoding and stronger capability. When AI systematically removes difficulty, it removes the conditions that produce genuine competence. The work feels easier, but the person does not necessarily get better.

• Relatedness, the sense that development matters, requires leaders to explicitly signal that growth is the goal, not just output. When the only feedback a person receives concerns the quality of AI-assisted output, they get no signal that their personal development is valued, and the conversation around their growth never happens.

Designing for genuine capability development means designing for all three (autonomy, competence, and relatedness) deliberately. It can’t be an accident or a side effect of good management. It requires the same intentionality you would bring to any other strategic priority.

The Four Moves of a Leader Who Builds Capability

What does this actually look like in practice? I want to be specific because the risk with a concept like “capability environment design” is that it can remain abstract. These are the four concrete moves that distinguish leaders who build capability environments from those who don’t.

Move One: They separate the output conversation from the development conversation.

In every significant piece of AI-assisted work, there are two distinct evaluations worth making. The output evaluation: Is it accurate? Is it useful? Does it achieve the intended outcome? The development evaluation: Who did the thinking? Where did the reasoning hold? Where did it break?

Most feedback conversations merge these two evaluations into one. However, effective leaders who foster growth deliberately separate them. They provide distinct feedback for both the output and the development process, making it clear which type of conversation they are having at any given moment.

This isn't about creating more work. It’s about applying a different quality of attention to the work already being done. It requires leaders to care as much about their team's development as they do about its output.

Move Two: Make it standard practice to ask, “What was your thinking before you used AI?”

Asking this one question consistently and without judgment, making it a normal part of work discussions, can shift usage patterns more effectively than any policy or training program. It establishes the expectation that human thought should precede AI assistance. Most importantly, it signals to every team member that their leader values their thinking and capability, not just their output.

The leaders who consistently ask this question are building the Sharpening Pattern as a cultural default — human thinking first, AI as a challenger second. The ones who don't are allowing the Replacement Pattern to become one instead. If you haven't read Article 1 of this series, that distinction is worth understanding before you go further: [link].

Move Three: They design for periodic scaffold removal.

The most important diagnostic available to a leader who wants to know whether capability is growing is simple: create conditions where people have to perform without the tool. Not constantly, not punitively, but regularly enough to tell the truth about where capability actually stands.

This might look like a monthly working session where teams tackle a problem without AI assistance. It might look like a leadership development exercise that requires individual analysis before any collaborative or tool-assisted work begins. It might look like a client presentation prepared by a single person, with their own reasoning, before AI refines it.

The specific form matters less than the principle: if capability is never tested without the scaffold, you genuinely don’t know whether it exists.

Move Four: They make the capability gap visible before it becomes critical.

The scaffold dependency trap is only dangerous when it is invisible. According to Zimmerman's developmental model, professionals progress through four stages on their path to genuine expertise: from Observation and Emulation, where scaffolding is necessary for developing capability, to Self-Control and Self-Regulation, where they can operate independently. A professional at Stage 2 using AI might appear identical to one at Stage 4 on all productivity metrics. However, the difference becomes stark and consequential when a high-stakes moment removes the scaffold.

This requires building evaluation systems that track the gap between AI-assisted performance and unassisted performance, not as a punitive metric but as a developmental one. The question is not “are you dependent?” The question is “where is the gap, and what do we do about it?”

That is a coaching conversation, not a performance management conversation. And it is only possible in an environment where the leader has made it safe to have it.

The Leadership Agenda That Doesn’t Exist Yet

While most organizations are developing an AI strategy, few are creating a human capability strategy to match.

The first perspective focuses on the practicalities of AI: tools, deployment, governance, and ROI. The second, however, delves into a more critical question: what is the human impact of AI integration? This line of inquiry examines the leadership required to ensure that AI empowers individuals rather than fosters dependence.

Here’s what that agenda looks like when a leadership team takes it seriously:

• They review AI integration not just for efficiency gains but for capability impact.

• They ask, as a standing agenda item: where is AI making our people more capable, and where is it substituting for capability they should be building?

• They develop managers in how to design for human capability development, not just in the use of AI tools. The manager who knows how to prompt an AI is useful. The manager who knows how to design the conditions where their team develops through using AI is rare and valuable.

• They create metrics that track the development of people alongside the productivity of outputs. Not as separate initiatives, but as integrated measures of what organizational performance actually means.

• And they model it personally. The leader who asks, “What did I think before I used AI?” who maintains and demonstrates their own deliberate practice, their own thinking-first habits, their own willingness to work through difficulty without immediately reaching for the tool, is building the culture through their own behavior before any policy is written.

The organizations that will lead in five years will not be the ones that deployed AI most aggressively. They will be the ones whose leaders understood that deploying AI and developing people are two different jobs, and they focused on both.

---

The Competency That Changes Everything

I want to return to where this article began, the leadership team in the room with the wrong questions on the table.

The right questions are not complicated. They don’t require deep technical knowledge. They require a leader who has decided that the development of the people in the room is as much their responsibility as the productivity of the outputs those people produce.

Making that decision explicit, maintaining it consistently, and designing it into how work is evaluated and feedback is given, is the new leadership competency the AI era demands.

It is not new in its essence. Peter Drucker saw it. John Wooden practiced it. Every great leader who has ever built a team that became better over time understood it.

What’s new are the stakes. In an era where AI can produce competent work, we risk losing our own ability to do so. Leaders who overlook capability development aren’t just missing growth opportunities; they’re embedding fragility into their organizations when genuine human judgment is more valuable than ever.

The most important question in your next leadership team meeting is not which AI tools to deploy.

It is: what kind of people do we want our people to become, and are we designing for that?

That question has always mattered. In the AI era, it is urgent.

If this reframes the conversation you’re having about AI in your organization, I’d ask you to share it with the leader who owns your people development agenda, and ask them whether they are designing for capability or just for productivity.

Previous article in this series: Article 1 - Beyond AI Adoption: Is Your Organization Getting Smarter?

Next in this series: Article 3 — The Individual Playbook: How to Use AI Without Losing Your Edge.

References

Deci, E. L., & Ryan, R. M. (1985). Intrinsic motivation and self-determination in human behavior. Plenum.

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.

Drucker, P. F. (1967). The effective executive. Harper & Row.

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–39). Academic Press.

Charles Good works with organizations that have already solved the adoption problem and are now asking the harder question. As President of the Institute for Management Studies, he reaches over 20,000 professionals annually. His Outlearn Loop framework, built on behavioral and learning science, is the architecture organizations use to design AI integration that builds human capability rather than substituting for it. He writes The AI Capability Playbook and The Performance Playbook on Substack and hosts The Good Leadership Podcast (Apple / Spotify / YouTube).

The Professional Who Looked Capable

Sarah, a senior strategy consultant at a mid-sized firm, seems to be thriving in the age of AI. By every available metric, she is a model of adaptation. She uses AI daily, producing work that is more polished, better structured, and delivered faster than ever before. Her manager sees an employee who has embraced the future.

Then, one afternoon, the system goes down.

She has to think through a market entry problem on the fly, no AI assist, just her and the problem. She hesitates, not to a complete standstill, but enough to matter. Her analysis is shallow, and the frameworks that should come effortlessly fail to materialize. The confident aura she carries in her AI-assisted work vanishes, leaving a noticeable void. And until that moment, nobody, not her manager, not her firm, not the adoption dashboard, could see the difference.

A professional who performs well with AI and a professional who performs well looks identical until the scaffold is removed.

This isn’t the story of a single consultant. It’s the story of a quiet phenomenon unfolding in organizations everywhere: the problem of measuring the wrong things.

The Numbers Behind the Crisis

Here is where most organizations currently stand.

McKinsey’s 2025 State of AI report, drawn from nearly 2,000 executives across 105 countries, found that 88% of organizations now use AI in at least one business function. Adoption, by any standard measure, is happening. And yet only one-third report scaling AI across the enterprise. However, only 6% qualify as genuine AI high performers, which are organizations achieving meaningful bottom-line impact. And a striking 72% of CIOs told Gartner in 2025 that their organizations are breaking even or losing money on their AI investments.

Read that last number again. Nearly three out of four technology leaders, the people whose job it is to make AI work, are not seeing positive returns.

According to Gartner, the primary obstacle isn’t the technology. It’s the challenge of demonstrating value, cited by 49% of organizations surveyed. Meanwhile, McKinsey’s data shows the factor with the highest correlation to EBIT impact is workflow redesign, yet only 21% of organizations using AI have fundamentally redesigned how work flows around it.

Despite enormous investment, widespread tools, and near-universal adoption, most organizations still struggle to demonstrate durable value from their efforts. The reason for this, as explained by learning science, is not what the industry is currently discussing.

What Ethan Mollick Gets Right And What He Leaves Unresolved

Ethan Mollick’s Co-Intelligence is the book your executive team has likely already read. It’s worth engaging with directly because it provides the most thoughtful framing of the AI adoption conversation.

Mollick’s central argument is genuine: the most effective use of AI is as a thinking partner. In other words, a challenger or tireless colleague who accelerates human work without replacing human judgment. He argues persuasively that well-used AI makes humans more creative, more productive, more informed. Organizations that learn to work with AI, not against it, will have a significant advantage.

That argument is correct, but it leaves one critical mechanism unresolved.

Co-Intelligence describes what the destination looks like. It doesn’t fully explain what happens to human capability on the journey, when it’s consistently faster and easier to ask AI than to think for yourself.

The gap lies not in Mollick’s vision, but in the behavioral design required to bring it to life. Co-Intelligence raises a critical question it doesn't quite answer: when AI assistance is perpetually available (always faster and more polished than our own work), what conditions will foster human capability rather than atrophy?

That is a question for learning science, which happens to have a precise answer.

The Outsourcing Trap: What Happens at the Neurological Level

When a person uses AI to generate an answer they should have worked through themselves, it short-circuits the learning process.

Cognitive scientists call this cognitive offloading: the practice of transferring a mental task to an external system. Some forms of offloading are helpful and largely harmless. Writing reminders in a notebook or using a calendar frees up working memory so we can focus elsewhere.

The concern arises when we begin to outsource thinking itself, which is the analysis, synthesis, judgment, and construction of ideas. When that happens, we may gain efficiency, but we also bypass the very cognitive processes that help information take root in memory. In other words, we don’t just save time; we short-circuit the mental work required for learning and understanding.

The generation effect, first demonstrated by Peter Slamecka and Peter Graf, and replicated hundreds of times since, shows that information people generate themselves, even imperfectly and with effort, is remembered far better than information they simply receive.

When an organization systematically replaces human work with AI-generated content, it does more than just obstruct the learning process—it erodes the very standards by which individuals evaluate quality. Over time, professionals risk losing the capacity to discern whether their own thinking is sound. They grow dependent on external validation (AI outputs, peer consensus, and managerial approval) as their own judgment atrophies from disuse. This is not a deficiency that shows up in a performance review; it reveals itself when the stakes are highest.

What Zimmerman’s Research Actually Predicts

Barry Zimmerman spent decades studying one question: how do humans actually get better at things? His answer was the Cyclical Phases Model, which is a three-phase cycle that every meaningful learning attempt moves through, whether we're aware of it or not.

Forethought. Before you begin a task, you set your goals, decide on a strategy, and activate what you already know about the problem. This isn’t just administrative preparation. It’s the moment you take genuine ownership of the challenge and determine how you will approach it.

Performance. During the task, you actively monitor progress, detect when you are drifting off course, and make real-time adjustments. This internal feedback loop (observing, evaluating, and adapting) is what distinguishes improving, adaptive performance from routine execution.

Self-Reflection. After the task is complete, you compare the outcome to your original plan, identify the factors that influenced the result, and decide what you will do differently next time. This is where learning compounds, turning a single experience into improved capability the next time you face a similar challenge.

Now consider what happens when AI is the primary cognitive resource:

The Forethought phase is outsourced. AI proposes the strategy. You review and approve, but the cognitive act of committing to a direction, the moment where you wrestle with the problem and decide how to approach it, never fully occurs.

The Performance phase is bypassed. Since AI monitors and adjusts the process in real time, your own internal feedback loop is never activated. Consequently, you lose the ability to detect errors, recalibrate your approach, and strengthen the mental pathways that foster adaptive performance.

The Self-Reflection phase has no content. When AI produces the output, there’s no human performance to evaluate. As a result, the attribution cycle, the process of assessing what works and what doesn’t, never begins, and improvement grinds to a halt.

Zimmerman’s model precisely predicts that this pattern hinders genuine capability development. Instead, it fosters a dependency on tools that merely mimics competence. This illusion holds until the tool is no longer available at a critical moment requiring independent judgment. In that instant, the underlying capability that should have been cultivated is absent, unable to function without AI support.

The Capability Regression Nobody Is Measuring

Zimmerman’s Multi-Level Model describes how self-regulatory skill develops through four stages: Observation → Emulation → Self-Control → Self-Regulation.

In practical terms, every professional moves through the same developmental sequence on the path to genuine expertise:

• Observation: You watch experts perform the work and see how decisions are made.

• Emulation: You begin practicing with guidance, feedback, and clear instructions.

• Self-Control: You execute the task independently while applying internalized standards.

• Self-Regulation: You adapt those standards fluidly when the situation changes.

The first two stages, Observation and Emulation, are the guided phases of learning. In these initial steps, you acquire skills by watching others, following established models, and practicing with dedicated support. This foundational scaffolding is essential, as it is the starting point for every future expert.

However, for development to occur, this scaffold must gradually be removed. The challenge of working without this support is what guides a learner toward Self-Control (the internalization of standards) and, eventually, Self-Regulation (the ability to adapt those standards to new situations).

Now consider what happens when AI becomes a permanent scaffold.

The developmental pressure that normally pushes someone from Level 2 (Emulation) to Level 3 (Self-Control) never arrives. Because the AI system is always faster, always polished, and always one prompt away, there is little incentive to internalize its thinking process.

As a result, the learner remains stuck at Level 2 indefinitely, capable when the tool is present, but unable to perform at the same level without it.

From the outside, this distinction is invisible.

Measured by productivity dashboards, output volume, or tool usage, a Level 2 professional using AI can look identical to a Level 4 professional with genuine adaptive expertise. The difference only surfaces when true capability is required, such as in:

• Novel problems with no obvious template.

• High-stakes judgment calls where trade-offs matter.

• Complex situations that demand a nuanced interpretation of context.

By the time that moment arrives, the gap has been building for months.

This is what sits underneath the McKinsey finding that only 6% of organizations are achieving meaningful AI impact. It is not that 94% have the wrong tools. It’s that 94% have not designed the human capability system that makes the tools compound rather than substitute human capability.

This dynamic is why so few organizations achieve a meaningful impact from AI. The problem is rarely the technology; when studies show only 6% of organizations see significant results, it’s not because the other 94% chose the wrong tools. It’s because most companies haven’t built the human systems needed to amplify judgment, learning, and skill development.

The Two AI Usage Patterns

There is a precise distinction worth making, and one that most AI rollout strategies have never made explicit.

The Replacement Pattern: In this pattern, you encounter a problem → Query AI → Review output → Submit. The human’s cognitive contribution is reduced to an editorial role. As productivity increases, individual capability can atrophy.

The Sharpening Pattern: With this pattern, you encounter a problem → Form your own analysis first → Engage AI to challenge, surface what you missed, and pressure-test your reasoning → Compare AI’s position to your own → Update your thinking → Submit. This process not only increases productivity but also compounds your capability over time.

While the usage metrics for both patterns are identical, their long-term effects over 24 months differ significantly. And here is the operational problem: no adoption dashboard can tell them apart. Organizations are measuring tool engagement when they should be measuring capability development. They are asking, “Are our people using AI?” when the question that determines their future is, “Are our people getting better because of how they are using AI?”

These are not the same question. And the one you ask will determine the one you answer.

What This Actually Looks Like

Organizations that are building genuine AI-augmented capability, not just AI-dependent productivity, design their AI integration around five principles drawn directly from the SRL research:

Forethought first. Before turning to AI for any significant task, it’s best to first establish your own perspective. This involves conducting your own analysis, forming your own hypotheses, and outlining your own recommended approach. Only then should you use AI to challenge, refine, and stress-test your work, not to create it from scratch.

Generation before consumption. For any task where true understanding is the goal, resist the urge to turn to AI right away. First, formulate your own response. This initial effort, even if flawed, activates the cognitive pathways essential for deep learning. Engaging with the material yourself first primes your mind to better absorb and integrate the insights AI can provide.

Reflection as a designed practice. After working with AI, it’s crucial to pause and reflect. Ask yourself: Where were my initial ideas strong? What new perspectives did the AI reveal? How can I bridge that gap in my own knowledge or skills? This self-reflection, a key phase of Zimmerman’s model of learning, isn’t automatic; it must be intentionally designed into the workflow.

Difficulty Is the Mechanism, Not the Obstacle. When using AI to sharpen your thinking, you might struggle. The AI’s challenge may reveal gaps in your knowledge, the forethought-first approach might feel slow, and generating ideas before consuming content might expose what you don’t yet know. This difficulty is not a problem; it’s proof that the learning process is working.

Measure Capability, Not Just Output. Don't just measure what people can produce with AI; measure what they can produce without it. Track this gap and work to close it. The leaders of tomorrow won't be the organizations with the highest AI adoption today. They will be the ones in which human and AI capabilities grow in tandem, creating a powerful synergy in which each amplifies the other.

These five principles map directly onto the Outlearn Loop, the same learning architecture that powers The Performance Playbook series, applied here to AI-augmented work.

The Question That Actually Matters

Let’s return to Sarah. According to her firm’s adoption dashboard, she’s a high performer with exemplary AI usage. But here’s what the dashboard would have recommended for her: query first, review the output, and then submit. It defaults to the Replacement Pattern simply because no one has ever made the distinction explicit.

What Sarah needed wasn’t less AI, but a different approach to it. She needed a firm that structured its workflow to prioritize human analysis, requiring her to develop her own insights before consulting artificial intelligence. That single change in process is what separates a professional who becomes sharper with every AI interaction from one who becomes increasingly dependent on it.

The capability Sarah lost wasn’t visible on any dashboard. It eroded quietly, one outsourced interaction at a time.

The question worth asking is not: Are our people using AI?

It is: Are we building capability or just building dependency?

If this reframes how you’re thinking about your AI investment, share it with the leader in your organization who owns the adoption dashboard and ask them which question they’re actually trying to answer.

Next in this series: Article 2 — The New Leadership Competency Nobody Is Developing. Knowing how to use AI is a tool skill. Knowing how to design environments where using AI makes your people better is a leadership skill, and almost nobody is building it.

References

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.

Bjork, R. A., & Bjork, E. L. (1992). A new theory of disuse and an old theory of stimulus fluctuation. In A. Healy, S. Kosslyn, & R. Shiffrin (Eds.), From learning processes to cognitive processes: Essays in honor of William K. Estes (Vol. 2, pp. 35–67). Lawrence Erlbaum Associates.

Gartner. (2025). CIO and technology executive survey. Gartner Research.

Kirkpatrick, D. L., & Kirkpatrick, J. D. (2006). Evaluating training programs: The four levels (3rd ed.). Berrett-Koehler.

McKinsey & Company. (2025). The state of AI: How organizations are rewiring to capture value. McKinsey Global Institute.

Mollick, E. (2024). Co-intelligence: Living and working with AI. Portfolio/Penguin.

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604.

Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778.

Zimmerman, B. J. (1989). A social cognitive view of self-regulated academic learning. Journal of Educational Psychology, 81(3), 329–339.

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–39). Academic Press.

Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70.

Zimmerman, B. J., & Schunk, D. H. (Eds.). (2011). Handbook of self-regulation of learning and performance. Routledge.

Charles Good works with organizations that have already solved the adoption problem and are now asking the harder question. As President of the Institute for Management Studies, he reaches over 20,000 professionals annually. His Outlearn Loop framework, built on behavioral and learning science, is the architecture organizations use to design AI integration that builds human capability rather than substituting for it. He writes The AI Capability Playbook and The Performance Playbook on Substack and hosts The Good Leadership Podcast (Apple / Spotify / YouTube).

The AI tools are deployed. The licenses are active. The usage metrics are moving in the right direction. The productivity gains are real: tasks that used to take hours are now completed in minutes, outputs that required senior expertise are now produced by junior staff, and the pace of execution has measurably accelerated.

And in a growing number of organizations, the humans are getting less capable every week.

The erosion of capability is neither dramatic nor visible. It does not trigger alerts in any monitored system. Instead, it happens gradually and quietly; the specific and consequential decay occurs when the cognitive work that builds expertise is systematically handed to technology. As a result, people stop developing the very judgment, pattern recognition, and adaptive reasoning that these tools were intended to enhance.

Many organizations will discover this the hard way: at a critical moment when it’s needed most, under pressure, or in a high-stakes situation. Or they might realize it when facing a genuinely novel problem that defies existing patterns and demands independent, adaptive, and deeply grounded human expertise.

At that moment, they will find that the people who were supposed to possess that expertise no longer do. Instead, they have spent the last eighteen months practicing something else entirely.

If your organization is in the middle of an AI transformation, a moment of truth is approaching. We know this because learning and behavioral science predicts it with startling accuracy. So why is this prediction being ignored? It’s not because the science is questionable. It’s because most companies are asking the wrong question about their AI transformation, making the real problem invisible.

The Wrong Question

Every boardroom conversation about AI transformation is organized around some version of the same question:

How do we get our people to use AI more?

Higher adoption rates. More licenses are opened regularly. Better integration into daily workflows. More consistent engagement with the tools the organization has already paid for.

While these are all reasonable questions, they’re also the wrong ones, not because adoption does not matter, it does, but because it optimizes for a metric that can’t distinguish between two completely different futures.

In the first future, human capability and AI capability are growing together. People are using AI to think harder, to pressure-test their own reasoning, to surface what they missed, to challenge assumptions they did not know they were making, and to produce better work than either a human or a technology could produce alone. The organization’s collective intelligence is compounding. The judgment, pattern recognition, and adaptive expertise distributed across its people are getting stronger with every AI interaction, because every AI interaction requires the human to generate, compare, and update their own thinking.

In the second future, AI capability is replacing human capability in ways that are invisible until they are consequential. People are using AI to generate the thinking they should be doing themselves, receiving outputs they consume without encoding, building tool dependency instead of genuine expertise, and skipping the cognitive work that produces the neural architecture of genuine skill.

While individual productivity may be on the rise, the organization’s collective intelligence is quietly eroding. This paradox leaves employees both more productive and more fragile, a paradox that usage dashboards fail to capture.

Both futures have identical adoption metrics. The divergence only becomes visible when conditions change.

The question, “How do we get our people to use AI more?” fails to see this distinction, as it optimizes for usage in both potential futures equally. An organization that asks only this question has no way of knowing which future it’s building until the difference becomes apparent. By then, the gap has been months or years in the making.

The Right Question

The question the science demands is different.

How can we create an organization where AI encourages deeper thinking rather than replacing it, and where uniquely human abilities like judgment, pattern recognition, and decision-making are strengthened daily?

This question changes everything downstream. It points to different investments, metrics, and design principles for integrating AI into how people work. It requires organizations to think not just about tool deployment but about the human capability architecture that tool deployment is either building or eroding.

And it requires understanding what the science actually says about how humans develop genuine capability, because you can’t build something you don’t understand. Most people designing AI transformation programs have never been trained in learning science, cognitive psychology, or behavioral science. They have been trained in change management, project management, and technology deployment. Those are valuable disciplines, but they are not the disciplines that explain why capability develops or degrades.

The scientific principles behind this are well-established, with roots stretching back over a century. The foundational research includes Ebbinghaus’s forgetting curve, first documented in 1885, and Bjork’s work on desirable difficulties, spanning the last 40 years. Further contributions include the generation effect, identified in 1978, and Zimmerman’s models of self-regulated learning, published in 1989 and 2000. Additionally, Gollwitzer’s findings on implementation intentions have been replicated across hundreds of studies, while Hadwin and Järvelä’s research on socially shared regulated learning has been evolving since 2011.

Although none of this was developed with AI in mind, all of it predicts with uncomfortable precision what AI as a cognitive replacement does to human capability. It also shows what conditions are necessary for genuine expertise to develop alongside powerful tools, rather than be replaced by them.

Until now, this body of science has remained largely separate from discussions about AI’s transformative potential. This series aims to bridge that gap.

What This Series Is

The AI Capability Playbook is six articles built around a single argument: the capability crisis most AI transformation efforts are quietly building is predictable, preventable, and already underway in organizations that are measuring the wrong things and asking the wrong questions.

The series is not for everyone.

It’s not for the individual professional building their personal learning system, that work lives in The Performance Playbook, the operational guide for running The Outlearn Loop on yourself, phase by phase. If that is what you are looking for, start there, in a seperate category of this channel.

This series is for the organisational decision-maker.

The CIO who bought the licences and is now watching the adoption metrics, wondering why the promised capability gains haven’t materialised.

The CHRO who senses a shift in the workforce’s independent judgment but can’t yet name or measure it.

The L&D leader who knows their training programs aren’t keeping pace with tool deployment but struggles to articulate this in terms that the business can act on.

The people leader responsible for building a team that performs under real-world conditions is now watching AI reshape what “performance” truly means.

If you are responsible for an AI investment or advising someone who is, this is the series that connects the science to what you are watching happen.

How the Argument Builds

The series is organized as a complete argument. Each article builds on the last while standing on its own.

The series opens with the reframe. Article 1 (Beyond AI Adoption: Is Your Organization Getting Smarter?) argues that the wrong question is driving the wrong strategy. Organizations optimizing for usage metrics are heading toward a capability crisis; their dashboards will not detect it until it is too late or very expensive to fix. It establishes the foundational distinction between two patterns of AI use: one that compounds human capability, and one that quietly substitutes for it. The science behind why this happens (the generation effect, Zimmerman’s self-regulated learning cycle, cognitive load theory, and scaffold dependency) is laid out precisely.

Article 2 establishes the leader’s responsibility. (The New Leadership Competency Nobody Is Developing) It makes the case that the most consequential leadership skill in the AI era is deliberately shaping how work gets done and how AI gets used, so that human capability compounds rather than atrophies alongside rising productivity. It identifies the three failure modes most organizations are currently in, draws on Deci and Ryan’s Self-Determination Theory to explain why AI-optimized workflows undermine intrinsic motivation and mastery development, and gives four concrete moves that distinguish leaders who are building genuine capability from those who are not.

Article 3 brings the argument to the individual. (How to Use AI Without Losing Your Edge) It is for the professional navigating this in real time, with real deadlines and real pressure to produce. It introduces the two-category framework at the personal level, the capabilities worth protecting in your own development versus those that can be freely automated, and gives five specific behavioral habits that determine whether AI use is compounding your expertise or quietly eroding it.

Article 4 addresses the measurement gap. The capability crisis described in this series is invisible to standard adoption dashboards — by design, not by accident. Article 4 describes what it actually takes to measure whether human capability is growing or atrophying alongside rising AI productivity, and gives organizational leaders the metrics that tell the truth about which future they are building.

Article 5 moves from the individual to the team. The individual-level failure modes are serious. The team-level failure modes — the aggregation illusion, the dominance trap, regulatory loafing, and scaffold dependency at scale — are potentially more consequential for organizational performance, and they are entirely invisible to standard adoption metrics. Article 5 names them precisely, explains the science behind each, and describes what collective capability development looks like when it is being built versus when it is being eroded.

The series closes with the architecture. Article 6 is constructive. It describes what genuine AI-augmented capability development looks like as a designed system — the elements that address every failure mode the series identifies, the Outlearn Loop applied at organizational scale, and the metrics that actually tell the truth about whether an organization is getting smarter or becoming more dependent.

Read in sequence, the six articles take you from the reframe that makes the problem visible to the architecture that makes it solvable. Each article can also be read independently — follow whatever thread the title stops you on.

A Note on the Science

The learning and behavioral science featured in this series isn’t included merely for credibility; it’s here because it provides the only precise explanation for what is actually happening.

Every consulting framework, every change management methodology, every technology adoption model currently being applied to AI transformation was developed before AI existed at scale. They describe what has historically been true about how organizations adopt new technologies.

However, they do not describe what is specifically true about how AI, as a cognitive tool rather than a productivity tool, interacts with the human capability development process.

This is precisely what learning science describes. The principles of how humans encode, retrieve, develop, and sustain expertise are universal, explaining how we learned to use spreadsheets, how surgeons mastered laparoscopic techniques, and how pilots adapted to digital cockpits. With the same force and specificity, learning science can be used to show how professionals can learn to work alongside AI.

Herein lies the critical difference: none of those earlier technologies could do the cognitive work the human was supposed to be developing. A spreadsheet doesn’t think for you, but an AI tool can, if you allow it. The choice of whether to adopt this capability, and how you integrate it into your organization, will ultimately determine the future you build.

This series will break down the science behind this choice, explaining the costs and outcomes in an easy-to-understand, practical way.

The Question Worth Asking

Before you read the six articles, there is one question worth carrying through all of them.

Not, are our people using AI?

Are you building capability or just building dependency?

Those are not the same question. Everything in this series is an attempt to explain why, and to give you the architecture for ensuring the answer to the second question is yes, regardless of what your adoption dashboard currently says about the first.

If this reframes the question you're asking about your AI investment, the six articles that follow are built to take you from the problem to the architecture.

The first article in this series: The Beyond AI Adoption: Is Your Organization Getting Smarter? The real question isn’t adoption. It’s whether AI is making your people smarter

References

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.

Deci, E. L., & Ryan, R. M. (2000). The “what” and “why” of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268. https://doi.org/10.1207/S15327965PLI1104_01

Ebbinghaus, H. (1885/1913). Memory: A contribution to experimental psychology (H. A. Ruger & C. E. Bussenius, Trans.). Teachers College, Columbia University. (Original work published 1885)

Gartner. (2025). CIO and technology executive survey. Gartner Research.

Gollwitzer, P. M. (1999). Implementation intentions: Strong effects of simple plans. American Psychologist, 54(7), 493–503. https://doi.org/10.1037/0003-066X.54.7.493

Hadwin, A. F., Järvelä, S., & Miller, M. (2011). Self-regulated, co-regulated, and socially shared regulation of learning. In B. J. Zimmerman & D. H. Schunk (Eds.), Handbook of self-regulation of learning and performance (pp. 65–84). Routledge.

McKinsey & Company. (2025). The state of AI: How organizations are rewiring to capture value. McKinsey Global Institute.

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604. https://doi.org/10.1037/0278-7393.4.6.592

Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745

Zimmerman, B. J. (1989). A social cognitive view of self-regulated academic learning. Journal of Educational Psychology, 81(3), 329–339. https://doi.org/10.1037/0022-0663.81.3.329

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–40). Academic Press. https://doi.org/10.1016/B978-012109890-2/50031-7

Zimmerman, B. J., & Moylan, A. R. (2009). Self-regulation: Where metacognition and motivation intersect. In D. J. Hacker, J. Dunlosky, & A. C. Graesser (Eds.), Handbook of metacognition in education (pp. 299–315). Routledge.

Charles Good works with organizations that have already solved the adoption problem and are now asking the harder question. As President of the Institute for Management Studies, he reaches over 20,000 professionals annually. His Outlearn Loop framework, built on behavioral and learning science, is the architecture organizations use to design AI integration that builds human capability rather than substituting for it. He writes The AI Capability Playbook and The Performance Playbook on Substack and hosts The Good Leadership Podcast (Apple / Spotify / YouTube).