5D Library › 一人公司自动化架构、数字人直播漏斗与任务级人工智能套利 · English

第一章 科斯边界的逆转:团队规模、协调成本与一人公司的上限

企业为什么有规模

科斯问:既然市场协调生产如此有效,企业为什么存在?他的回答是:使用市场有成本(找价格、谈判、订约),所以企业会扩张,直到在内部多组织一笔交易的成本等于到外部购买的成本 (Coase 1937; Williamson 1975)。企业内部同样有成本:每两个必须就某件事达成一致的人就是一条沟通渠道,nn 人的团队有 n(n−1)/2n(n-1)/2 条 (Brooks 1975)。AI 智能体同时改变了两边:它降低了不雇人也能完成工作的成本,又不增加人与人之间的渠道。本章要问的是:这会把最优团队规模压低到什么程度,以及「一个人加一群智能体」在哪里失效。

能交付的最小团队

设一家年营收 RM 1,000,000 的企业,需要 25,000 个任务小时。AI 智能体能以每小时 RM 8 完成其中的比例 ss,但每个 AI 小时需要 0.05 小时的人工审核。每人每年工作 1,800 小时,每两个人之间每年因协调损失 60 小时。

推导

# Chapter 1. p01: How large a team does a RM 1,000,000 business need?
# The business needs H = 25,000 task-hours a year (RM 40 of revenue per
# task-hour). A share s of the hours can be done by AI agents at RM 8
# per task-hour, but each AI hour needs r = 0.05 hours of human review.
# Each person works 1,800 hours; every pair of people loses k = 60 hours
# a year to coordination, so a team of n has n * 1800 - k n(n-1)/2
# useful hours (Brooks's n(n-1)/2 channels). Salary RM 60,000 a person.
# For each s the smallest team that delivers, and its profit, is found.
H, PRICE, HOURS, K, WAGE, AI, R = 25000, 40, 1800, 60, 60000, 8, 0.05

def useful(n):
    return n * HOURS - K * n * (n - 1) / 2

def team(s):
    need = (1 - s) * H + R * s * H
    for n in range(1, 200):
        if useful(n) >= need:
            return n, need
        if useful(n + 1) <= useful(n):
            return None, need
    return None, need

if __name__ == "__main__":
    print("   AI share  human hours  team  coordination  profit (RM)")
    for s in (0.0, 0.5, 0.8, 0.9, 0.95, 0.98):
        n, need = team(s)
        cost = n * WAGE + s * H * AI
        print("   {:7.2f}  {:11,.0f}  {:4d}  {:9,.0f} h  {:12,.0f}".format(
            s, need, n, K * n * (n - 1) / 2, H * PRICE - cost))
    best = max(range(1, 40), key=useful)
    print("   most useful hours any team can give: n = {}, {:,.0f} h"
          .format(best, useful(best)))
#    AI share  human hours  team  coordination  profit (RM)
#       0.00       25,000    21     12,600 h      -260,000
#       0.50       13,125     9      2,160 h       360,000
#       0.80        6,000     4        360 h       600,000
#       0.90        3,625     3        180 h       640,000
#       0.95        2,438     2         60 h       690,000
#       0.98        1,725     1          0 h       744,000
#    most useful hours any team can give: n = 30, 27,900 h

没有 AI 时,企业需要 21 人,每年在协调上损失 12,600 小时,亏损 RM 260,000。一半工作自动化后,需要 9 人,利润 RM 360,000;自动化比例 0.80 时 4 人,利润 RM 600,000;0.95 时 2 人。只有在 0.98 时团队才缩成一个人:创始人自己的工作加审核共 1,725 小时,利润 RM 744,000。协调项也给任何人类团队设了上限:超过 30 人后,再加人不再增加任何有效工时(最多 27,900 小时)。

一人公司的上限

决定性的参数不只是 ss,还有审核负担 rr。一位有 1,800 小时的创始人,最多能运转 1800/((1−s)+rs)1800/((1-s) + rs) 个任务小时。

推导

# Chapter 1. p02: The revenue ceiling of a one-person firm. A founder
# with 1,800 hours a year does the (1 - s) share of task-hours that AI
# cannot do and reviews AI work at r hours per AI task-hour, so the most
# task-hours she can run is 1800 / ((1 - s) + r s); at RM 40 a
# task-hour that sets a revenue ceiling. RM 1,000,000 needs 25,000
# task-hours, i.e. (1 - s) + r s at most 0.072.
HOURS, PRICE = 1800, 40

def ceiling(s, r):
    return HOURS / ((1 - s) + r * s) * PRICE

if __name__ == "__main__":
    rs = (0.02, 0.05, 0.10, 0.20)
    print("   revenue ceiling, RM a year")
    print(("   AI share" + "".join("     r={:.2f} ".format(r) for r in rs))
          .rstrip())
    for s in (0.80, 0.90, 0.95, 0.98, 0.99):
        print("   {:8.2f} ".format(s) +
              "".join(" {:10,.0f} ".format(ceiling(s, r))
                      for r in rs).rstrip())
    print("   RM 1,000,000 needs (1 - s) + r s at most {:.3f}".format(
        HOURS / 25000))
    for r in rs:
        s_min = (1 - HOURS / 25000) / (1 - r)
        txt = ("AI share at least {:.3f}".format(s_min) if s_min <= 1
               else "out of reach even at s = 1")
        print("     review r={:.2f}: {}".format(r, txt))
#    revenue ceiling, RM a year
#    AI share     r=0.02      r=0.05      r=0.10      r=0.20
#        0.80     333,333     300,000     257,143     200,000
#        0.90     610,169     496,552     378,947     257,143
#        0.95   1,043,478     738,462     496,552     300,000
#        0.98   1,818,182   1,043,478     610,169     333,333
#        0.99   2,416,107   1,210,084     660,550     346,154
#    RM 1,000,000 needs (1 - s) + r s at most 0.072
#      review r=0.02: AI share at least 0.947
#      review r=0.05: AI share at least 0.977
#      review r=0.10: out of reach even at s = 1
#      review r=0.20: out of reach even at s = 1

审核负担为每个 AI 小时 0.05 小时时,AI 完成 0.90 的工作,营收上限是 RM 496,552;0.98 时是 RM 1,043,478。营收 RM 1,000,000 要求 (1−s)+rs(1-s) + rs 不超过 0.072:r=0.02r = 0.02 时 AI 比例至少要 0.947,r=0.05r = 0.05 时至少 0.977,r=0.10r = 0.10 或更高时,即使 AI 包揽一切也做不到。所以「零员工、百万营收」是关于审核的命题,而不只是关于自动化的:谁不能压缩核查 AI 产出的时间,谁就无法独自达到。

一人公司的营收上限与 AI 比例的关系,四种审核负担(p02)。

假说 · 尚未验证

作者的诠释(尚未验证): 在使用 AI 智能体的一人企业中,「每个创始人小时的营收」更多由创始人对每件 AI 交付物的审核时间解释,而不是由任务自动化比例解释。什么会证明它错: 在记录了工时与营收的数据中,自动化比例对每个创始人小时营收的预测力至少与审核时间一样好。

小结

协调渠道按 n(n−1)/2n(n-1)/2 增长,所以 AI 接手任务小时后,最优团队迅速缩小:AI 比例 0.80 时从 21 人降到 4 人。营收 RM 1,000,000 的一人公司,既需要极高的 AI 比例,也需要很低的审核负担;每个 AI 小时审核 0.10 小时时,它就无法实现。

参考文献

Brooks, Frederick P. 1975. The Mythical Man-Month: Essays on Software Engineering. Addison-Wesley.
Coase, Ronald H. 1937. “The Nature of the Firm.” Economica 4 (16): 386 to 405.
Williamson, Oliver E. 1975. Markets and Hierarchies: Analysis and Antitrust Implications. Free Press.

Fan, Y. (2026). One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage: Coasean Transaction-Cost Collapse, Cyclic Script Retention and High-Margin Solopreneur Portfolios [一人公司自动化架构、数字人直播漏斗与任务级人工智能套利] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090235 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).