In May 2013, four employees at Haier, the Chinese appliance maker, wanted to build a laptop for gamers. They had read through 30,000 online reviews of gaming PCs and boiled them down to 13 complaints. Then they wrote to Zhou Zhaolin, head of the Haier platform that included the laptop business, begging for a meeting.
“My first instinct was to kill the project,” Zhou later told Gary Hamel and Michele Zanini of Harvard Business Review.
Then he realized it wasn’t his call. “In making decisions,” he said, “we have to let users and entrepreneurs, not managers, speak.” He gave the team 1.8 million RMB of seed money, about $270,000, with more to come only if the market liked the product. In December 2013 the first 500 laptops went on sale online. They sold out in five days. A second batch of 3,000 went in 20 minutes.
That wasn’t only one manager’s generosity. Haier was turning itself into small units, more than 4,000 of them by 2018, each with three formal rights: to set its own strategy, hire its own people and set its own pay. The case has the rest, including the cost: in recent years, HBR noted, Haier had dismissed more than 10,000 employees.
A question of proportion
Henri Fayol, who ran a French mining and metals company and set out his principles of management more than a century ago, would have recognized Zhou’s dilemma. Centralization, he wrote, is “not a system of administration good or bad in itself.” Some of it always exists. The question of centralizing or decentralizing “is a simple question of proportion. It is a matter of finding the limit that suits the enterprise.”
He gave a definition that still works. “Everything that increases the importance of the subordinates’ role is decentralization; everything that reduces it is centralization.”
How far to go, Fayol said, depends on the boss, the people, and the conditions. A chief with great ability and energy can centralize a lot and reduce subordinates to “simple agents of execution.” One who prefers to draw on their experience and advice can decentralize widely. Because those abilities keep changing, the right proportion is “constantly variable.” And it isn’t only a question for the top. There is no manager at any level, he wrote, who cannot widen or narrow the initiative of the people below.
His test was practical: “the best possible use of the faculties of all the personnel.”
What decentralizing costs
Pushing decisions down is not free. In 2012, Henrik Kniberg and Anders Ivarsson described how Spotify organized its engineers into small “squads,” each run like a mini startup. They named the price in the same paper: “the potential downside to full autonomy is a loss of economies of scale. The tester in squad A may be wrestling with a problem that the tester in squad B solved last week.” Spotify’s answer was groups across squads, called chapters and guilds, meant to win back “some economies of scale without sacrificing too much autonomy.” A former Spotify product manager, Jeremiah Lee, later wrote that the squad model “was only ever aspirational and never fully implemented.” The case covers that.
Haier had the same problem and a similar fix. It grouped its units into platforms, each with an owner whose job was to bring units together to spot chances to collaborate. No one reported to the platform owner, who had no staff.
So the real choice is never “central or not.” It is which decisions, at which level, with what glue between the pieces.
The salons
Imagine someone who owns three hair salons and approves every discount personally. A regular customer asks for a little off the price because a colour came out uneven. The stylist texts the owner, who is busy with a client across town. The customer waits, then leaves unhappy. Letting each salon’s manager give up to a set amount, while keeping the price list central, might fix the queue without letting prices drift. If one manager starts discounting far more than the others, the owner might pull that limit back for a while. That is Fayol’s “constantly variable” in practice.
How much to let the agent decide
People working with AI agents are setting the same dial, and moving it. In a study of Claude Code, Anthropic found that newer users, with fewer than 50 sessions, let the agent run with full auto-approval in about 20 percent of sessions. By 750 sessions, that rose to over 40 percent.
Experienced users didn’t stop paying attention. They interrupted the agent more often, in about 9 percent of turns against 5 percent for new users. Anthropic read this as a change of strategy, from approving each action to watching and stepping in when needed. The agent also adjusts: on the most complex tasks, Claude Code stopped to ask for clarification more than twice as often as on the simplest ones.
Anthropic’s conclusion sounds a lot like Fayol’s. The autonomy an agent exercises, it wrote, “is co-constructed by the model, the user, and the product.” It depends on the boss, the worker and the conditions, and it shifts as each of them changes.
Zhou kept his job as platform head. What he gave away was one decision, the one he concluded belonged to the team and its customers.