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

第四章 投流盈亏平衡 ROAS 与 24 小时数字人

简式与精确式

付费流量用 ROAS(广告支出回报率)衡量:GMV 除以广告支出。一个常见的经验公式把盈亏平衡 ROAS 写成 1/(m−r−c)1/(m – r – c),其中 mm 是毛利率,rr 是退款率,cc 是平台佣金率,都以占 GMV 的比例计。这个公式在毛利之外把每笔退款都当作全额营收损失,又照样对它收佣金。若按每 RM 1 GMV 计算:退款的订单退回营收与佣金,但留下损失 LL(退货运费、无法再卖的货品),那么投流前的贡献是 (1−r)(m−c)−rL(1-r)(m-c) – rL,盈亏平衡 ROAS 是它的倒数。

推导

# 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)

典型情形(毛利 0.45,退款 0.15,佣金 0.08,每退款 RM 1 损失 0.08)下,简式要求 ROAS 4.55,精确式只要 3.31:简式过严,会叫停本来赚钱的投放。低毛利、高退款的产品(0.30 与 0.25),简式说任何 ROAS 都无法打平,精确式给出 8.00;若退回的货大多报废,则是 20.00。误差的方向取决于 LL;精确式多需要一个数字,店铺自己的退货记录就能提供。

24 小时数字人,还是高峰时段的真人

AI 数字人主播可以昼夜不停地直播,每小时成本很低,但转化率低于熟练的真人。下面设数字人的转化率是真人的 0.6,每小时成本 RM 12,真人主播 RM 150;观众按一条示意性的日曲线进场,晚上 21:00 达到高峰。

推导

# 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

真人主播只有在每小时 175.6 名观众以上才划算,数字人只要 23.4 名。高峰时段真人播 4 小时,每天赚 RM 442;数字人播 24 小时,RM 1,093;数字人只在划算的 19 个小时播,RM 1,120。在数字人之外再在高峰加真人,一天反而降到 RM 985,因为按这组数字,数字人即使在 21:00 也比真人赚得多。只有当数字人在高峰的转化率低于真人的 0.551 时,结论才会反过来。结果取决于转化率之比,而这是任何模型都给不了的:它必须实测(第八章)。

一天各时段真人主播与数字人每小时的贡献(p08)。

注

标明数字人。 一些平台与司法辖区要求把 AI 生成的主播标注出来。让观众以为合成主播是真人的直播,会招来盈亏平衡公式没有计入的退款与投诉。

小结

简式盈亏平衡公式 1/(m−r−c)1/(m – r – c) 可能往任一方向出错;精确贡献 (1−r)(m−c)−rL(1-r)(m-c) – rL 在简式给出 4.55 时给出 3.31,在简式认为无解时给出 8.00。转化率为真人 0.6、成本只有十二分之一的数字人,即使在高峰也胜过真人;比值低于 0.551 时则不然。

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).