Stop Trying to Replicate a Single Star Performer

In 2012, journalist Kurt Eichenwald published a devastating exposé in Vanity Fair. Microsoft, he reported, had become a place where talented people fought each other instead of the competition. The culprit was stack ranking. Every review cycle, managers sorted employees into a fixed curve of winners and losers. Rankings were updated once a year. Between reviews, everyone knew who the previous cycle’s winners were and copied them. Here was the paradox. Microsoft had no shortage of diverse talent. World-class engineers from dozens of countries, with wildly different technical backgrounds, filled its teams. The cognitive diversity was already in the room. Yet the organization was converging on mediocrity, because employees were all chasing the same role models. That’s the conclusion we drew from our latest research. We developed a computational model to study the performance implications of a stack ranking system like Microsoft’s. Our findings imply that the problem may not be who a company hires. The bigger problem, we found, is who everyone copies, and how long that information stays frozen. Our model captures the logic of annual stack-ranking system found in many organizations: A single top performer is identified, others copy that person, and the ranking does not refresh until the next cycle. In such systems, two failures can arise. First, “overshooting”: employees keep copying the same top-ranked star long after absorbing what is useful, overwriting their own distinctive practices in the process—a net loss in learning. Second, “frozen targets”: the designated star is chosen at the start of the cycle and never updated. By the time the cycle ends, other employees have improved through different paths, but nobody notices because the rankings have not been refreshed. The organization keeps learning from yesterday’s best. Its diverse knowledge never gets recombined into better solutions. That is the real cost. What Our Research Found Knowledge recombination, defined as combining correct practices held by different people into superior solutions, is the central challenge of organizational learning, innovation, and strategy. In 1991, James March published what became one of the most-cited papers in management research, shaping thinking across organizational theory, cultural evolution, and complexity science. Using a computational model, he showed that organizations perform better when people learn slowly from each other. The received wisdom ever since: Slow learning preserves diverse practices and prevents premature convergence. Our research, forthcoming in the Strategic Management Journal, revisits this result using the same method. In this context, a computational model simulates an organization of learners who observe one another and copy the practices of those who perform best. Each simulated employee holds a partial, possibly correct answer to a complex problem whose full solution no single person knows. Across thousands of simulation runs, we varied two things: How fast people copy others, and how often the system refreshes information about who is actually performing well. What we found surprised us: slow learning works, but by accident. March’s model, like most annual performance systems (such as stack ranking), refreshes information about who is performing best only at the end of a review cycle. When learning is slow, only a few copying moves occur before the next update, so the system barely changes between updates. The two failure modes our model identifies, overshooting and frozen targets, rarely arise, simply because little changes before the next update cycle. It looks like friction is helping. In reality, the information just happens to stay fresh. From Fixed Targets to Live Targets Once you see this, a better approach becomes obvious: keep the information fresh on purpose. Most organizations operate fixed-target systems: They identify top performers at the start of a review cycle and hold that information constant. What we propose instead is a live-target system, where the identity of high performers is updated continuously. Imagine a live leaderboard, not a frozen ranking. In our simulations, when performance information refreshes after each learning event, fast learning no longer leads to premature convergence. Two mechanisms explain why. First, people stop copying once they see they’ve caught up. With live performance information, an employee who has matched the current best performer knows it immediately and stops; no more overshooting, no more overwriting practices that were already correct. Second, and more important, underdogs who reach the frontier through different paths become visible immediately. Others can now learn from them. Different correct practices, held by different people, get combined into superior solutions. We call the resulting mechanism target diversity: a rotating set of high performers whose distinct knowledge gets recombined. The Mind-the-Gap Rule Our most actionable finding is what we call the mind-the-gap rule. When people far from the frontier learn fast and those near the frontier learn selectively, simulated organizations reach near-optimal performance on complex tasks. The advantage over uniform slow learning grows with problem complexity. When problems have many aspects to get right, slow learning stalls because underdogs take too long to catch up, and their unique knowledge never enters the recombination process. Under the mind-the-gap rule, those same underdogs reach the frontier quickly through distinct paths and immediately become fresh imitation targets. To see this dynamic in action, consider a sales team. Junior reps far from the frontier learn fastest by copying current top performers, accelerating their catch-up. Once a junior rep’s numbers approach the frontier, the system shifts: they slow down, protect the unconventional moves that got them there, and become a new imitation target for the rest of the team. The frontier expands rather than collapses around a single style. Population Diversity Is Not Target Diversity Scott Page’s influential work on the diversity bonus showed that groups of diverse problem-solvers can outperform groups composed of the highest-ability individuals. But capturing that bonus requires that diverse knowledge actually spreads. Spreading, in turn, requires target diversity: a rotating set of role models whose distinct practices get copied and combined. Most organizations invest heavily in population diversity through hiring, only to undermine target diversity through their performance systems. When everyone copies the same established star, the organization converges on one combination of practices. Other employees hold different correct practices the star lacks. Because these employees never become imitation targets, their knowledge never enters the recombination process. The Underdog Premium The most counterintuitive prescription: The underdog who just caught up is more valuable to copy than the established star. The star’s practices have already been widely copied, so little new information remains to extract. The underdog arrived through a distinct combination of practices that has not yet spread. That person is the key to knowledge recombination, precisely because they lack a long track record. This cuts against managerial instinct. We trust track records. We promote people with proven histories. We build best-practice libraries around established winners. But track records reflect where knowledge has already converged. The real gains lie in the knowledge that has not yet spread, and it tends to sit with those who just reached the top through unexpected paths. The same logic reframes talent development. Most organizations treat it as one-directional: bring people in, train them, align them with established best practices. That approach erases the distinct knowledge that justified the hire in the first place. A better design is two-sided. New hires learn from current top performers to accelerate catch-up. At the same time, their emerging strengths are made visible early, before they converge to organizational norms. The effect is double-sided: development accelerates the new hire’s growth, and it raises the bar for incumbents, who now compete against a moving frontier rather than a fixed one. Boundary Conditions and Limitations The prescription is not universal. For routine tasks where best practice is well established, rapid convergence is desirable: copy the known answer fast and move on. The mind-the-gap logic pays off when problems are complex, when multiple correct paths exist, and when people copy high performers rather than conforming to majority practice. Our evidence comes from a computational model. It identifies the mechanism behind one of the most influential results in organizational learning and offers a concrete design principle. Our prescriptions have been tested in simulations, not yet in field settings. Testing the mind-the-gap rule directly in firms, research teams, or online platforms is therefore the natural next step. Microsoft, Revisited Return to Microsoft. When Satya Nadella became CEO in 2014, his most important internal move was to change who got copied and how quickly information about them flowed. He killed stack ranking and replaced it with continuous feedback through a system called Perspectives, where collagues flag what someone should “keep doing” and what they should “rethink.” Performance signals became live rather than annual. The company-wide Hackathon, which grew to more than 18,000 participants, made anyone’s breakthrough visible immediately, regardless of rank. Underdogs could reach the frontier and become role models overnight. The cultural shift from a “know-it-all” to a “learn-it-all” organization was, at its core, a shift from fixed targets to live targets. Microsoft’s market capitalization grew from roughly $300 billion to over $3 trillion in the decade that followed. We do not attribute that growth to a learning-system redesign alone. Azure, the OpenAI partnership, and a string of acquisitions were all part of the story. What we can say is that the cultural and operational changes described here were among the preconditions for the creative recombination those bets required. What’s Next: AI and Knowledge Recombination Knowledge recombination is the central problem not just of organizational learning, but also of innovation, science, and strategy. AI now makes the mind-the-gap rule actionable at scale. Systems that continuously track contributions can measure the gap between each employee and the current frontier and adjust how much each person learns from others. When the gap is large, AI can accelerate learning by surfacing relevant practices from top performers. When the gap is small, AI can do something more important: help frontier performers explore selectively, protecting the unique knowledge that makes them valuable as future imitation targets. The risk is that organizations deploy AI the wrong way, training a single corporate model on last year’s best practices and pointing everyone at it. That is premature convergence at machine speed. The opportunity is the opposite: AI that keeps the target set rotating. The hardest part of implementing this is not technical. It is cultural. Organizations that have built status hierarchies around established stars will resist surfacing the underdog. The question is no longer whether your organization is diverse enough. It is whether your systems can spot who just caught up to the frontier, and surface their knowledge before it gets trained away. Knowledge recombination does not happen automatically. It happens by design.

Stop Trying to Replicate a Single Star Performer

Summary.

Partner Center