5D Library › One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage · 中文

Chapter 4. Break-Even ROAS and the 24-Hour Avatar

The short formula and the exact one

Paid traffic is judged by ROAS, return on ad spend: GMV divided by ad spend. A common rule of thumb puts the break-even ROAS at 1/(m−r−c)1/(m – r – c), with gross margin mm, refund rate rr and platform commission cc, all as shares of GMV. The rule charges every refunded order as a full loss of revenue on top of the margin, and still charges commission on it. Counting instead per RM 1 of GMV that a refunded order returns its revenue and its commission but leaves a loss LL (return shipping, goods that cannot be resold), the contribution before ads is (1−r)(m−c)−rL(1-r)(m-c) – rL, and break-even ROAS is its reciprocal.

Calculation

# Chapter 4. p07: Break-even ROAS, the short formula and the exact one.
# Per RM 1 of GMV: gross margin m (price less cost of goods), refund
# rate r (the order is refunded), platform commission c on GMV (refunded
# with the order), and a loss L per refunded RM 1 (return shipping and
# goods that cannot be resold). The exact contribution before ads is
# (1 - r)(m - c) - r L, so ROAS_BE = 1 / ((1 - r)(m - c) - r L).
# The short rule ROAS_BE = 1 / (m - r - c) charges every refund as a
# full loss of revenue and still charges commission on it.

def short(m, r, c):
    x = m - r - c
    return 1 / x if x > 0 else None

def exact(m, r, c, loss):
    x = (1 - r) * (m - c) - r * loss
    return 1 / x if x > 0 else None

def fmt(v):
    return "{:6.2f}".format(v) if v else "  none"

if __name__ == "__main__":
    print("   margin refund commission loss   short rule   exact")
    cases = ((0.45, 0.15, 0.08, 0.08), (0.60, 0.10, 0.10, 0.10),
             (0.30, 0.25, 0.10, 0.10), (0.30, 0.25, 0.10, 0.40),
             (0.70, 0.30, 0.05, 0.20))
    for m, r, c, loss in cases:
        print("   {:5.2f}  {:5.2f}    {:5.2f}    {:4.2f}     {}      {}"
              .format(m, r, c, loss, fmt(short(m, r, c)),
                      fmt(exact(m, r, c, loss))))
    print("   ('none': no ROAS breaks even)")
#    margin refund commission loss   short rule   exact
#     0.45   0.15     0.08    0.08       4.55        3.31
#     0.60   0.10     0.10    0.10       2.50        2.27
#     0.30   0.25     0.10    0.10       none        8.00
#     0.30   0.25     0.10    0.40       none       20.00
#     0.70   0.30     0.05    0.20       2.86        2.53
#    ('none': no ROAS breaks even)

For a typical case (margin 0.45, refunds 0.15, commission 0.08, loss 0.08 per refunded RM 1), the short rule asks for a ROAS of 4.55 and the exact one for 3.31: the short rule is too strict, and would stop campaigns that pay. For a low-margin, high-refund product (0.30 and 0.25) the short rule says no ROAS can break even, while the exact formula gives 8.00, or 20.00 if refunded goods are mostly lost. The direction of the error depends on LL; the exact formula needs one number more, and the shop’s own return records supply it.

A digital avatar for 24 hours, or a human at peak?

An AI avatar host can stream around the clock at a small hourly cost, but converts fewer viewers than a skilled human. Below, an avatar converts 0.6 of a human’s rate and costs RM 12 an hour against RM 150 for a human host; viewers arrive along an illustrative daily curve peaking at 21:00.

Calculation

# Chapter 4. p08: A digital avatar for 24 hours, a human host at peak,
# or both. Viewers arriving each hour of the day follow the illustrative
# curve below (peak at 21:00). Orders per viewer are 0.032 for a human
# host (the best cycle of p06) and 0.6 of that for an avatar. Each order
# contributes RM 26.7 after cost of goods, commission and refunds. A
# human host costs RM 150 an hour, an avatar RM 12. A host is worth an
# hour only if viewers * orders per viewer * 26.7 exceeds its cost.
ARRIVE = (60, 35, 20, 12, 8, 8, 15, 30, 55, 70, 80, 90,
          110, 120, 100, 95, 105, 130, 170, 240, 320, 360, 300, 160)
CONV, AVATAR, PER_ORDER = 0.032, 0.6, 26.7
COST = {"human": 150.0, "avatar": 12.0}

def hour_profit(h, who):
    mult = 1.0 if who == "human" else AVATAR
    return ARRIVE[h] * CONV * mult * PER_ORDER - COST[who]

def plan(assign):
    return sum(hour_profit(h, w) for h, w in assign.items())

if __name__ == "__main__":
    for who in ("human", "avatar"):
        need = COST[who] / (CONV * (1.0 if who == "human" else AVATAR)
                            * PER_ORDER)
        print("   {:6s} pays its way above {:.1f} viewers an hour".format(
            who, need))
    peak = range(19, 23)
    mixed = {h: "avatar" for h in range(24)
             if hour_profit(h, "avatar") > 0}
    mixed.update({h: "human" for h in peak})
    plans = {
        "human 19:00-23:00": {h: "human" for h in peak},
        "avatar 24 hours": {h: "avatar" for h in range(24)},
        "avatar where it pays": {h: "avatar" for h in range(24)
                                 if hour_profit(h, "avatar") > 0},
        "human peak + avatar": mixed,
    }
    for name, a in plans.items():
        print("   {:22s}: {:2d} hours, RM {:7,.0f} a day".format(
            name, len(a), plan(a)))
    v = max(ARRIVE)
    a_star = 1 - (COST["human"] - COST["avatar"]) / (v * CONV * PER_ORDER)
    print("   at the peak hour ({} viewers) a human beats the avatar only"
          .format(v))
    print("     if the avatar converts below {:.3f} of a human".format(
        a_star))
#    human  pays its way above 175.6 viewers an hour
#    avatar pays its way above 23.4 viewers an hour
#    human 19:00-23:00     :  4 hours, RM     442 a day
#    avatar 24 hours       : 24 hours, RM   1,093 a day
#    avatar where it pays  : 19 hours, RM   1,120 a day
#    human peak + avatar   : 19 hours, RM     985 a day
#    at the peak hour (360 viewers) a human beats the avatar only
#      if the avatar converts below 0.551 of a human

A human host pays her way only above 175.6 viewers an hour, an avatar above 23.4. Four hours of human hosting at the peak earn RM 442 a day; an avatar for 24 hours, RM 1,093; an avatar only in the 19 hours where it pays, RM 1,120. Adding a human at the peak to the avatar lowers the day to RM 985, because with these numbers the avatar earns more than a human even at 21:00. That reverses only if the avatar converts below 0.551 of a human at the peak. The result depends on the conversion ratio, which no model can supply: it has to be measured (Chapter 8).

Contribution per hour of a human host and an avatar across the day (p08).

Note

Disclose the avatar. Several platforms and jurisdictions require that AI-generated hosts be labelled as such. A stream that lets viewers think a synthetic host is a real person invites refunds and complaints that no break-even formula includes.

Summary

The short break-even formula 1/(m−r−c)1/(m – r – c) can be wrong in either direction; the exact contribution (1−r)(m−c)−rL(1-r)(m-c) – rL gave 3.31 where the rule gave 4.55, and 8.00 where the rule gave none. An avatar that converts 0.6 as well as a human at a twelfth of the cost beats a human even at the peak; below a ratio of 0.551 it does not.

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).