Microsoft Stack Ranking 2013: Keep the Curve or Scrap Employee Ratings?

The room

The psychology: relative evaluation

People constantly measure themselves against others, and a ranking turns that habit into the rule of the game. Under an absolute standard, a colleague’s success costs you nothing. Under a forced distribution, it does: if only two of ten can be rated great, every strong teammate makes your own top rating less likely. Economists call this structure a tournament. It can raise effort, because people work harder to beat a rival, but it also changes how they treat the rivals sitting next to them.

That is the trade-off at the center of the Microsoft case. The curve was designed to reward the best and remove the weakest, and the research suggests it can do some of that. But the same structure gives employees a reason to withhold help and to treat calibration as a contest. Experiments find that when people can affect each other’s results, ranking can turn from a spur into a source of sabotage.

One Microsoft made the tension sharper. Ballmer’s plan defined success as coordination within and among teams, while the review system still paid people for finishing ahead of the colleagues they were meant to coordinate with. Brummel’s memo addressed this directly by making contributions to others part of the definition of impact. The psychology is simple: people tend to do what they are compared on.

Why smart people fall for it

Three forces made the curve work against the teamwork Microsoft was asking for.

1. 1

   A colleague’s gain becomes your loss

   With a fixed number of top ratings, helping a teammate shine can cost you your own place in the distribution.

2. 2

   Ranking feedback invites sabotage

   When people can affect each other’s results, competition for rank can push some to undermine rivals rather than outperform them.

3. 3

   People do what they are compared on

   A review that measures individual standing teaches people to optimize standing, whatever the strategy memo says about collaboration.

Did Microsoft stack ranking really hurt cooperation?

The strongest evidence against the curve is testimony, not data. Vanity Fair’s case rested on interviews with dozens of current and former employees, who described colleagues competing for position instead of against rival companies. That is powerful, but it is also a sample of people willing to talk to a journalist about a company in trouble. GeekWire’s Todd Bishop, who covered Microsoft closely, called the article a caricature that highlighted the company’s worst qualities.

Nor can stack ranking carry the whole blame for Microsoft’s difficult decade. Vanity Fair itself also pointed to bureaucracy and an over-reliance on Windows, and Ballmer replaced the company’s divisional structure in 2013. The curve may have made those problems harder to solve by discouraging cooperation across groups, but no one has measured how much it contributed, and a fair account should say so.

What the controlled research adds is a mechanism. In the experiment by Johannes Berger, Christine Harbring and Dirk Sliwka, a forced distribution made workers more productive when they worked alone, and harmful when they could sabotage each other. Gary Charness, David Masclet and Marie Claire Villeval found something similar even with flat pay: telling people their rank raised average effort, but some participants accepted a personal cost to sabotage others or inflate their own scores, and the sabotage dragged overall performance down. Group identity reduced sabotage among peers but intensified rivalry between groups.

Large software projects sit at the interdependent end of that spectrum, because one group’s work usually depends on another’s. If the lab findings hold anywhere, they should hold in a setting where helping or withholding help is easy and hard to observe. That does not prove the curve damaged Microsoft, but it makes the employees’ complaints consistent with what experiments predict.

What did the research say about forced ranking?

The best-known study in favor of forced distribution is a 2005 simulation in Personnel Psychology by Steve Scullen of Drake University and his co-authors Bergey and Aiman-Smith. They modeled a workforce over 30 years in which the lowest-rated employees were fired each year, varying how reliable the ratings were, how well new hires were selected and how many people left voluntarily. In every scenario, workforce potential ended higher than it began, and the best case, with aggressive firing and reliable ratings, raised it by 48 percent.

The details matter more than the headline. Almost all of the gain came early, about 16 percent a year in the first two years, falling to around 2 percent by year six and about 1 percent by year ten, with no improvement after year twenty. Once the weakest have gone, each new round removes people who are closer to average. The researchers were also clear that they did not measure effects on company performance, so the model leaves out exactly the costs critics worried about, such as lost cooperation and morale.

Later work extended the model. A 2011 simulation by Jamie Mulligan and Rebecca Bull Schaefer found that a system used for three years and then dropped did not produce lasting gains, while a probationary version, which warned employees before firing them, was associated with lower voluntary turnover. Read together, the studies suggest a curve can lift a workforce that has many weak performers, but its value shrinks quickly and its hidden costs grow as the remaining people become stronger.

That pattern matters for Microsoft’s situation in 2013. The company had ranked its staff for years and had already spent time trying to soften the system. If the simulations are right, a curve that old would mostly be sorting good people from other good people, which is when its benefits are smallest and the pressure to compete with colleagues is greatest.

How would a modern team decide?

A modern team would start by separating the jobs a review system does. It sets pay, it identifies people who are not working out, and it helps people improve. A forced curve combines all three into one annual number, and the competition it creates for that number can undermine the development conversation. Microsoft’s 2013 design pulled them apart: Connects for feedback through the year, managers’ budgets for pay.

It would also take the strongest objection seriously. When the change was announced, Ben Thompson of Stratechery argued that the deeper problem had been strong performers paid like average ones, and that without forced differentiation managers might spread rewards evenly to avoid hard conversations. That risk is real in any system that relies on manager judgment, and Microsoft’s later practice, which by 2025 included a 0 to 200 rewards scale and a formal process for low performers, shows the company kept looking for ways to differentiate pay without a quota.

Third, a modern team would match the system to how work is done. Where jobs are independent and output is easy to count, relative comparison does less harm, and the Berger experiment suggests it can raise effort. Where people depend on each other, as in software teams, the same pressure invites withholding and sabotage. Measuring contributions to others, as Microsoft’s new definition of impact did, is one way to point the competition at the right target.

Finally, it would plan the transition as carefully as the design. Business Insider’s 2014 report showed how quickly rumors filled the gap when one division kept part of the old process for a few months. Employees who spent years under a curve will look for signs that it survives in disguise, so clear communication and consistent practice across divisions matter as much as the policy itself.

The manager's playbook

1. Name what each review is forDecide which parts of the process set pay, which flag poor performance and which support development, and design each on its own terms.
2. Map how interdependent the work isWhere people rely on each other’s output, avoid rules that make a colleague’s success reduce your own rating.
3. Measure contributions to othersInclude how people use colleagues’ input and help others succeed in the definition of impact, so cooperation counts.
4. Keep pay differentiation without quotasGive managers a budget and clear criteria for exceptional work, and check that rewards are not simply spread evenly.
5. Audit the rollout across divisionsConfirm every unit has dropped the old mechanics, and correct rumors quickly with plain facts about how decisions are made.

Questions people ask