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
第八章 全年账本与一项预注册检验
一门生意,三种经营方式
把各部分合在一起:一门每年需要 5,500 个任务小时的生意(直播运营、B2B 交付、产品与内容),营收来自第四章的数字人直播、两份顾问月费和一条产品线。方案 A 有团队,不用 AI 智能体;方案 B 是创始人一个人,AI 完成 0.8 的任务小时;方案 C 再加一位兼职编辑,从创始人手中接过 400 小时的审核与零碎工作。
推导
# Chapter 8. p15: A year's ledger for three ways to run the same business.
# Work: 5,500 task-hours a year (livestream operations 1,500, B2B
# delivery 2,400, product and content 1,600). Revenue: avatar livestream
# RM 1,120 a day (p08), two retainers at RM 5,000 a month, a product at
# RM 6,000 a month. All three run the same avatar stream; they differ in
# who does the task-hours. A: founder plus a team, no AI agents (staff
# at RM 60,000 and 1,800 hours a year, 60 hours a year lost per pair of
# people). B: founder alone, AI doing 0.8 of the task-hours at RM 8 an
# hour, 0.05 review hours per AI hour. C: as B, plus a part-time editor
# (RM 24,000 a year) who takes 400 of the founder's hours.
H, HOURS, WAGE, K = 5500, 1800, 60000, 60
REVENUE = 1120 * 365 + 2 * 5000 * 12 + 6000 * 12
def team_for(need):
n = 1
while n * HOURS - K * n * (n - 1) / 2 < need:
n += 1
return n
if __name__ == "__main__":
print(" revenue RM {:,.0f} a year".format(REVENUE))
n = team_for(H)
a_cost = (n - 1) * WAGE
print(" A by hand : {} people incl. founder, cost RM {:,.0f}, profit"
" RM {:,.0f}".format(n, a_cost, REVENUE - a_cost))
s, r = 0.8, 0.05
founder = (1 - s) * H + r * s * H
b_cost = s * H * 8
print(" B solo+AI : founder {:,.0f} h, cost RM {:,.0f}, profit"
" RM {:,.0f}".format(founder, b_cost, REVENUE - b_cost))
c_cost = b_cost + 24000
print(" C +editor : founder {:,.0f} h, cost RM {:,.0f}, profit"
" RM {:,.0f}".format(founder - 400, c_cost, REVENUE - c_cost))
print(" founder hours spare in C: {:,.0f} of {:,}".format(
HOURS - (founder - 400), HOURS))
# revenue RM 600,800 a year
# A by hand : 4 people incl. founder, cost RM 180,000, profit RM 420,800
# B solo+AI : founder 1,320 h, cost RM 35,200, profit RM 565,600
# C +editor : founder 920 h, cost RM 59,200, profit RM 541,600
# founder hours spare in C: 880 of 1,800三种方案的年营收都是 RM 600,800。A 需要包括创始人在内 4 个人,成本 RM 180,000,剩下 RM 420,800。B 的 AI 费用是 RM 35,200,剩下 RM 565,600,但创始人要在 1,800 小时中工作 1,320 小时。C 多付 RM 24,000,剩下 RM 541,600,创始人只需 920 小时,还有 880 小时空余。C 是更稳健的选择:B 留给上新、生病或一队上交事项(第二章)的余量太少,而 C 的空余小时正是新产品线或第三个顾问客户所需要的。
循环话术的预注册检验
第三、四章的直播数字在实测之前都只是假设。检验以直播的一小时为单位随机分组,而不是以单个观众为单位,因为同一小时的观众处在同样的条件下;每个时段约 200 名观众、时段内相关系数 0.005 时,设计效应为 1.995 (Kohavi et al. 2020)。
推导
# Chapter 8. p16: How many viewers a fair test of the run-of-show needs.
# Two-proportion test, two-sided alpha 0.05, power 0.80. Tests: a 4- vs
# a 6-minute cycle (32.0 vs 22.9 orders per 1,000 viewers, p06); avatar
# vs human host (0.6 x 0.032 vs 0.032, p08); and a 10% relative lift on
# 0.032. Viewers arrive in hour-long blocks of about 200 that share
# conditions, so blocks, not viewers, are randomised; with an intra-block
# correlation of 0.005 the design effect is 1 + (200 - 1) x 0.005.
import math
from statistics import NormalDist
Z = NormalDist().inv_cdf
def n_per_arm(p1, p2, alpha=0.05, power=0.80):
pbar = (p1 + p2) / 2
num = (Z(1 - alpha / 2) * math.sqrt(2 * pbar * (1 - pbar)) +
Z(power) * math.sqrt(p1 * (1 - p1) + p2 * (1 - p2)))
return math.ceil(num ** 2 / (p1 - p2) ** 2)
if __name__ == "__main__":
deff = 1 + (200 - 1) * 0.005
print(" design effect {:.3f}".format(deff))
for name, p1, p2 in (("4 vs 6 minute cycle", 0.0320, 0.0229),
("avatar vs human", 0.6 * 0.032, 0.032),
("10% lift on 0.032", 0.032, 0.0352)):
n = n_per_arm(p1, p2)
blocks = math.ceil(n * deff / 200)
print(" {:20s}: {:7,d} viewers per arm, {:4d} hour-blocks".format(
name, n, blocks))
# design effect 1.995
# 4 vs 6 minute cycle : 5,060 viewers per arm, 51 hour-blocks
# avatar vs human : 2,389 viewers per arm, 24 hour-blocks
# 10% lift on 0.032 : 49,777 viewers per arm, 497 hour-blocks比较 4 分钟与 6 分钟的周期(p06 中每 1,000 名观众 32.0 单对 22.9 单),每组需要 5,060 名观众,约 51 个小时时段。比较数字人与真人主播(转化率比 0.6),每组需要 2,389 名,24 个时段。要检出转化率 0.032 上 10% 的相对提升,每组需要 49,777 名观众、497 个时段:小幅改进无法由一家小店在短时间内证实,也不应该宣称。
注
一人企业的四步核查。 第一,每周测量每件 AI 交付物的审核时间与上交比例,它们决定了上限(p02、p04)。第二,在每个面对客户的智能体步骤之后设检查(p03)。第三,用店铺自己的退货损失计算盈亏平衡 ROAS,不用简式(p07)。第四,等副业连续几个月覆盖生活费之后再辞职,并保留至少一个长期客户(p11、p12)。
假说 · 尚未验证
预注册预测(尚未验证): 在按小时时段随机分组的检验中,4 分钟周期每 1,000 名观众的订单数比 6 分钟周期至少多 15%,数字人主播的转化率在真人的 0.5 到 0.8 之间。什么会证明它错: 较短周期的订单数没有显著更多,或者数字人的转化率比落在这个区间之外。
全书小结
随着协调渠道减少,AI 智能体缩小了最优团队规模:AI 比例 0.80 时从 21 人降到 4 人,只有审核负担很轻时才能缩到一个人(p01、p02)。误差沿智能体链条累积,处理上交的创始人是一个会饱和的排队系统(p03、p04)。循环话术有最佳周期长度,在示意性的离开率下约 4.0 分钟,演示中途进场的观众买得最少(p05、p06)。简式盈亏平衡 ROAS 可能往两个方向误导;数字人在高峰胜过真人,除非它的转化率低于真人的 0.551(p07、p08)。暴露度是任务的属性;最安全的收入在难以自动化、又被自动化任务供给的任务里(p09、p10)。等市场给出答案再辞职,并在组合中保留长期客户(p11、p12)。白圭的规则需要会回归的市场,《考工记》的百工说明了成文标准能省下什么(p13、p14)。
参考文献
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).