5D Library › One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage · 中文
Chapter 1. Coase’s Boundary Reversed: Team Size, Coordination and the One-Person Ceiling
Why firms have a size at all
Coase asked why, if markets coordinate production so well, firms exist at all, and answered that using the market has costs (finding prices, negotiating, contracting), so a firm grows until organising one more transaction inside costs as much as buying it outside (Coase 1937; Williamson 1975). Inside the firm there is a cost too: every pair of people who must agree on something is a channel of communication, and a team of has of them (Brooks 1975). AI agents change both sides. They lower the cost of doing work without hiring, and they do not add human channels. The question of this chapter is how far that pushes the best team size down, and where “one person and a cluster of agents” stops working.
The smallest team that delivers
Take a business with RM 1,000,000 of revenue a year, which needs 25,000 task-hours. AI agents can do a share of the hours at RM 8 an hour, but each AI hour needs 0.05 hours of human review. Each person works 1,800 hours, and each pair of people loses 60 hours a year to coordination.
Calculation
# 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 hWithout AI the business needs 21 people, loses 12,600 hours a year to coordination and makes a loss of RM 260,000. With half the hours automated, 9 people and RM 360,000 of profit; with 0.80, 4 people and RM 600,000; with 0.95, 2 people. The team shrinks to one person only at 0.98, where the founder’s own work and review come to 1,725 hours and profit is RM 744,000. The coordination term also caps any human team: past 30 people, adding a person adds no useful hours at all (27,900 hours at most).
The one-person ceiling
The decisive parameter is not only but the review burden . A founder with 1,800 hours can run at most task-hours.
Calculation
# 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 = 1At a review burden of 0.05 hours per AI hour, the revenue ceiling is RM 496,552 when AI does 0.90 of the work and RM 1,043,478 at 0.98. RM 1,000,000 needs to be at most 0.072: with that means an AI share of at least 0.947, with at least 0.977, and with or more it is out of reach even if AI did everything. “Zero employees, a million in revenue” is therefore a statement about review, not only about automation: whoever cannot cut the time spent checking AI output cannot reach it alone.
Hypothesis · not established
The author’s interpretation (not yet tested): among solo firms using AI agents, revenue per founder hour is explained better by the founder’s review hours per AI-produced deliverable than by the share of tasks automated. What would show it wrong: in logged time and revenue data, the automated share predicts revenue per founder hour at least as well as review hours do.
Summary
Coordination channels grow as , so AI that takes over task-hours shrinks the best team quickly: from 21 people to 4 at an AI share of 0.80. A one-person firm with RM 1,000,000 of revenue needs both a very high AI share and a low review burden; at 0.10 review hours per AI hour it is out of reach.
References
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).