Page 159 of Haidilao’s Hong Kong listing prospectus, published in September 2018, sets out how the hot pot chain pays the people who run its restaurants. There are two options, and each manager gets whichever comes out higher.
Option A: 2.8% of the profits of the restaurant you manage.
Option B: 0.4% of the profits of the restaurant you manage, plus 2.8% to 3.1% of the profits of the restaurants run by your apprentices, plus 1.5% of the profits of the restaurants run by their apprentices.
Do the arithmetic. Put in one apprentice whose restaurant earns as much as yours, and Option B already beats Option A. In the prospectus’s own words, managers “may receive a higher percentage of profits from their mentees’ restaurants than their own.” The restaurant you run every day carries the smallest percentage on the page.
What Grove counted
In High Output Management, published in 1983, Andy Grove, then Intel’s president, recalled asking a group of middle managers what a manager’s output was. They told him: judgments and opinions, direction, allocation of resources, mistakes detected, subordinates developed, commitments negotiated.
“Do these things really constitute the output of a manager?” he wrote. “I don’t think so.” Those were activities. The output of a manager running a wafer plant was finished wafers; of a high school principal, educated students; of a surgeon, a healed patient. Then he wrote it as an equation: a manager’s output equals the output of his organization plus the output of the neighboring organizations under his influence.
Your output as a manager is what your team, and the teams you influence, produce.
A busy calendar proves nothing either way. Grove’s own analogy was the quarterback: he doesn’t win games alone, and “league standings are kept by team, not by individual.” A manager can do their own job well, Grove wrote, “but that does not constitute his output.”
Haidilao wrote the second half of Grove’s equation into its payroll. Apprentices’ restaurants are the neighboring organizations, and the formula puts a price on them. The bet runs both ways: if a manager fails the company’s assessments and is removed, the prospectus says, “his/her mentor and the mentor’s mentor will be subject to financial penalties.”
Google’s Project Oxygen looked at the same idea through data. By the company’s account, it found that in 2008 teams under high-scoring managers had less turnover.
When the count goes up
Grove’s equation tells you what to count. It doesn’t tell you when more of it is a mistake.
In 2018 Haidilao had 363 restaurants, and headquarters, the prospectus said, had no fixed target for openings. By the end of 2021 it had 1,443. That November it announced it would shut or suspend about 300 restaurants with low traffic; by year end, 260 had closed for good. Among the reasons it listed: a high density of Haidilao restaurants nearby. The company pointed to the pandemic and to “our rapid expansion strategy over the past two years.” It did not blame the pay formula, and nothing in its filings shows how much the formula contributed. But its restructuring added a line: a “family head” who benefits from the family’s expansion “shall be accountable for the performance of the restaurants under his or her management.”
Influence is also hard to see. Grove extended the idea to people with no reports at all: an engineer on a body that sets procedures for every Intel wafer plant can raise the output of all of them, and individual contributors who spread know-how “should also be seen as middle managers.” A formula like Haidilao’s can only price the influence it can trace.
The agents who trained the machine
That gap showed up in one of the first large studies of generative AI at work. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,172 customer support agents at a Fortune 500 software company as an AI assistant that suggests replies in real time was rolled out.
Issues resolved per hour rose 15 percent on average. The least skilled fifth of agents gained 36 percent. The most skilled saw no productivity gain and small declines in quality. The authors read the pattern as the AI “capturing and disseminating” the behaviors of the most productive agents, though that is their interpretation, not something the study measured directly.
In Grove’s terms, the top agents’ know-how was now raising output across the floor. Their own numbers, resolutions per hour, couldn’t show it. The authors flagged the risk: top agents’ conversations feed the model’s ongoing training, and failing to reward them for continuing to produce good answers could make the model worse.
If your team’s assistant runs on examples, open the prompt library or example set and note whose work is in it. Then look at that person’s last review and see whether any of it appears there.
The Friday no one counted
Imagine you’re the senior engineer who knows the payments system best. Every Friday afternoon, three other teams bring you their questions about it. Your review lists your own tickets.
This quarter you’re behind on tickets, so you stop the Fridays. Your numbers recover. Over the next two months, the other three teams each ship a payments change late, and none of their reviews says why.