5D Library › 对话式成交自动机、异议消解信息增益与高客单销售动力学 · English

第六章 信任阈值停时规则与人工坐席规划

继续聊、请求付款,还是交给人

在每一轮对话之后,AI 面对三个选择:继续聊;直接请求付款;交给人工成交顾问。这是一个最优停时问题:选择何时停止收集信息、采取不可逆的行动。AI 的信念是「买家已经准备好」的概率 pp。每多聊一轮,有 10% 的买家离开,还没准备好的买家有 0.04 的概率被说服而准备好;准备好的买家有 0.6 的概率表现出购买信号,没准备好的只有 0.15。请求付款时,准备好的买家有 0.7 的概率成交(RM 9,800),没准备好的会拒绝,其中一半就此离开。人工接手的成本按每次计,人工对准备好的买家成交率 0.9,对没准备好的 0.2。

推导

# Chapter 6. p11: When should the AI ask for payment, hand over to a human,
# or keep talking? The buyer is ready to buy or not; the AI holds the
# belief p = P(ready). Each further turn: 10% of buyers drift away, a
# buyer who is not ready becomes ready with probability 0.04 (questions
# and proof do their work), and the buyer shows a buying signal with
# probability 0.6 if ready, 0.15 if not. Asking for payment closes a ready
# buyer with probability 0.7 (RM 9,800); a buyer who is not ready refuses,
# and half of them leave. A human closer's time costs c per hand-over;
# the human closes a ready buyer with probability 0.9 and one not ready
# with 0.2. Value iteration on a grid of 401 beliefs.
V, N = 9800.0, 400
DRIFT, NURT, SIG, ASK, LEAVE, HUMAN = 0.10, 0.04, (0.6, 0.15), 0.7, 0.5, \
    (0.9, 0.2)

def interp(f, p):
    x = p * N
    i = min(int(x), N - 1)
    return f[i] + (f[i + 1] - f[i]) * (x - i)

def solve(cost):
    f = [0.0] * (N + 1)
    for _ in range(5000):
        g, act = [], []
        for i in range(N + 1):
            p = i / N
            q = p + (1 - p) * NURT
            cont = 0.0
            for sy, sn in ((SIG[0], SIG[1]), (1 - SIG[0], 1 - SIG[1])):
                pz = q * sy + (1 - q) * sn
                cont += pz * interp(f, q * sy / pz)
            cont *= 1 - DRIFT
            stay = p * (1 - ASK) + (1 - p) * (1 - LEAVE)
            close = p * ASK * V + stay * interp(f, p * (1 - ASK) / stay)
            esc = (p * HUMAN[0] + (1 - p) * HUMAN[1]) * V - cost
            best = max((cont, "keep talking"), (close, "ask for payment"),
                       (esc, "hand to a human"))
            g.append(best[0])
            act.append(best[1])
        done = max(abs(a - b) for a, b in zip(f, g)) < 1e-7
        f = g
        if done:
            break
    return f, act

if __name__ == "__main__":
    for cost in (250.0, 500.0):
        f, act = solve(cost)
        print("   human time worth RM {:.0f} a hand-over:".format(cost))
        prev = None
        for i in range(N + 1):
            if act[i] != prev:
                print("     from P(ready) {:.3f}: {}".format(i / N, act[i]))
                prev = act[i]
        if cost == 250.0:
            for p in (0.1, 0.3, 0.6):
                i = int(p * N)
                print("     value at {:.1f}: RM {:,.0f} (ask now RM {:,.0f})"
                      .format(p, f[i], p * ASK * V))
#    human time worth RM 250 a hand-over:
#      from P(ready) 0.000: keep talking
#      from P(ready) 0.138: hand to a human
#      from P(ready) 0.500: ask for payment
#      value at 0.1: RM 2,454 (ask now RM 686)
#      value at 0.3: RM 3,768 (ask now RM 2,058)
#      value at 0.6: RM 6,001 (ask now RM 4,116)
#    human time worth RM 500 a hand-over:
#      from P(ready) 0.000: keep talking
#      from P(ready) 0.295: hand to a human
#      from P(ready) 0.415: ask for payment

人工时间按每次 RM 250 计时,最优规则分成三段:信念低于 0.138 时继续聊;0.138 到 0.500 之间交给人;0.500 以上由 AI 直接请求付款。这就是信任阈值停时规则:AI 不应在信念很低时就要钱,也不必把有把握的成交交给人。信念 0.3 时,按最优规则这段对话的价值是 RM 3,768,而立刻请求付款只值 RM 2,058。人工时间更贵(每次 RM 500)时,交给人的区间缩小到 0.295 至 0.415。

对话价值随信念的变化与三段最优行动(人工每次 RM 250,p11)。

注

阈值依赖人工时间的影子价格。 人手越紧,交给人的区间越窄;第五章的吉廷斯指数与增量排序,决定的是区间之内先交哪一段。

后面要站几位人工顾问

交给人的对话以随机的节奏到来,每次人工通话平均 25 分钟。如果买家要等,就会流失:等 ww 分钟后仍在的概率是 e−w/10e^{-w/10}。爱尔朗 C 公式给出需要等待的概率 (Gans et al. 2003)。

推导

# Chapter 6. p12: How many human closers stand behind the AI? Hand-overs
# arrive at random (Poisson) at lam per hour; a human call lasts 25
# minutes on average (exponential). Erlang C gives the chance a buyer
# must wait and P(wait > t) = Pw * exp(-(c mu - lam) t). A waiting buyer
# is still there after w minutes with probability exp(-w / 10), so the
# share lost while waiting is Pw * (1/10) / (gap + 1/10), where
# gap = (c mu - lam) per minute.
from math import exp, factorial

MU = 60 / 25.0                       # calls per closer per hour

def erlang_c(lam, c):
    a = lam / MU
    if a >= c:
        return 1.0
    s = sum(a ** k / factorial(k) for k in range(c))
    top = a ** c / factorial(c) * c / (c - a)
    return top / (s + top)

def row(lam, c):
    pw = erlang_c(lam, c)
    gap = (c * MU - lam) / 60          # per minute
    late = pw * exp(-gap * 5)
    lost = pw * (1 / 10) / (gap + 1 / 10)
    return pw, late, lost

if __name__ == "__main__":
    print("   hand-overs/h  closers  P(wait)  P(wait > 5 min)  lost waiting")
    for lam in (3.0, 6.0, 9.0):
        c0 = int(lam / MU) + 1
        for c in range(c0, c0 + 3):
            pw, late, lost = row(lam, c)
            print("   {:12.0f}  {:7d}  {:7.3f}  {:15.3f}  {:12.3f}".format(
                lam, c, pw, late, lost))
#    hand-overs/h  closers  P(wait)  P(wait > 5 min)  lost waiting
#               3        2    0.481            0.414         0.370
#               3        3    0.155            0.110         0.091
#               3        4    0.042            0.024         0.020
#               6        3    0.702            0.635         0.585
#               6        4    0.320            0.237         0.200
#               6        5    0.130            0.079         0.065
#               9        4    0.865            0.823         0.786
#               9        5    0.462            0.360         0.308
#               9        6    0.227            0.145         0.120

每小时 3 次交接时,2 位顾问让 0.481 的买家需要等待,0.370 在等待中流失;3 位降到 0.155 与 0.091;4 位降到 0.042 与 0.020。每小时 6 次交接时,4 位顾问仍有 0.200 在等待中流失,5 位降到 0.065。每小时 9 次时,5 位顾问流失 0.308,6 位降到 0.120。人工坐席按平均负荷配置,会在等待中丢掉大量最有价值的买家;应当比负荷多配一到两位。

假说 · 尚未验证

作者的假说(尚未验证): 在 AI 与人工协同的私域销售中,按信念阈值交接(而不是按买家要求或按热度交接)的团队,每次人工通话带来的成交额至少高 20%。什么会证明它错: 在按周随机切换交接规则的实验中,每次人工通话的成交额差距小于 20%。

小结

最优的停时规则是三段式:信念低时继续聊,居中时交给人,高时由 AI 请求付款;人工越贵,交给人的区间越窄。人工坐席要按排队来配置:按平均负荷配人,等待中的流失可达四成以上。

参考文献

Gans, Noah, Ger Koole, and Avishai Mandelbaum. 2003. “Telephone Call Centers: Tutorial, Review, and Research Prospects.” Manufacturing & Service Operations Management 5 (2): 79–141.

Fan, Y. (2026). Conversational Closing Automata, Objection-Handling Information Gain and High-Ticket Sales Kinetics: Diagnostic Question Trees, Trust-Threshold Stopping Rules and AI Closer Escalation Protocols [对话式成交自动机、异议消解信息增益与高客单销售动力学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090221 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).