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

第二章 诊断式提问树:问哪一题、问几题

比特多不等于价值高

一个问题值得问,不是因为它的回答有很多比特,而是因为回答可能改变下一步的行动:报哪一档价,或者先不报价。决策理论把这称为信息的价值;它与信息量相关,但并不相同。下面用动态规划精确求解一棵自适应的提问树:在每个节点,要么停下来按当前信念选最好的方案,要么问一个还没问过的问题,选择使期望价值最高的那一个。

推导

# Chapter 2. p03: Which question next? An adaptive question tree solved
# exactly by dynamic programming over the questions not yet asked: at each
# node either stop and take the best plan of p02, or ask the question
# whose answers lead to the most expected value (each question loses 2%
# of buyers). Compared with asking in the fixed S-P-I-N order and with
# always asking the question that carries the most bits.
import importlib.util

def load(name):
    spec = importlib.util.spec_from_file_location(name, name + ".py")
    m = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(m)
    return m

p01 = load("p01_belief")
p02 = load("p02_pomdp")
QS = list(p01.Q)

def best(belief, left, depth=0, log=None):
    stop = p02.value(belief)[1]
    choice, val = None, stop
    for q in left:
        v = 0.0
        for yes in (True, False):
            post, pz = p01.posterior(belief, p01.Q[q], yes)
            v += pz * best(post, [x for x in left if x != q], depth + 1)[1]
        v *= 1 - p02.DROP
        if v > val + 1e-9:
            choice, val = q, v
    return choice, val

def expected_questions(belief, left):
    q, _ = best(belief, left)
    if q is None:
        return 0.0
    n = 1.0
    for yes in (True, False):
        post, pz = p01.posterior(belief, p01.Q[q], yes)
        n += pz * (1 - p02.DROP) * expected_questions(
            post, [x for x in left if x != q])
    return n

def by_bits(belief, left, k):
    if k == 0 or not left:
        return p02.value(belief)[1]
    q = max(left, key=lambda x: p01.info(belief, p01.Q[x]))
    v = 0.0
    for yes in (True, False):
        post, pz = p01.posterior(belief, p01.Q[q], yes)
        v += pz * by_bits(post, [x for x in left if x != q], k - 1)
    return v * (1 - p02.DROP)

if __name__ == "__main__":
    prior = dict(p01.PRIOR)
    q, v = best(prior, QS)
    print("   pitch at once            RM {:,.0f}".format(
        p02.value(prior)[1]))
    fixed = max(p02.ask_then_offer(k)[0] for k in range(5))
    print("   fixed S-P-I-N, best k    RM {:,.0f}".format(fixed))
    bb = max(by_bits(prior, QS, k) for k in range(5))
    print("   most bits first, best k  RM {:,.0f}".format(bb))
    print("   optimal tree             RM {:,.0f}".format(v))
    print("   first question: {}".format(q.split(" (")[0]))
    print("   expected questions asked: {:.2f}".format(
        expected_questions(prior, QS)))
    for yes in (True, False):
        post, pz = p01.posterior(prior, p01.Q[q], yes)
        q2, _ = best(post, [x for x in QS if x != q])
        print("   if {}: next {}".format("yes" if yes else "no",
                                       q2.split(" (")[0] if q2 else "stop"))
#    pitch at once            RM 2,522
#    fixed S-P-I-N, best k    RM 2,671
#    most bits first, best k  RM 2,677
#    optimal tree             RM 2,748
#    first question: S situation
#    expected questions asked: 1.73
#    if yes: next N need-payoff
#    if no: next stop

直接报价的期望价值是 RM 2,522。按固定的 S、P、I、N 顺序问、选最好的题数,价值 RM 2,671;每次都问比特最多的问题,价值 RM 2,677;最优提问树的价值 RM 2,748,比直接报价高 9.0%,而且平均只问 1.73 题。最优树先问情境问题:回答「是」(团队超过五人、讲得出营收),再问需求回报问题;回答「否」,就停下来,不再追问。

五种做法的期望价值:直接报价、固定顺序、比特优先、最优提问树与完全信息上限(p03)。

注

问到能改变决定为止。 对一个已经回答「团队很小、说不出营收」的买家,再问痛点与影响,也不会让卖方改报高档方案;这些问题的比特仍然存在,价值却是零。

该问几题

每多问一题,就有一部分买家失去耐心。下面在不同的流失率下重新求解最优提问树。

推导

# Chapter 2. p04: How many questions before the offer? The optimal tree of
# p03 is re-solved for different shares of buyers lost per question, from
# none (a patient buyer) to 15% (a buyer who came for a price). Reported:
# the value per lead, its gain over pitching at once, and the expected
# number of questions the optimal tree asks.
import importlib.util

def load(name):
    spec = importlib.util.spec_from_file_location(name, name + ".py")
    m = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(m)
    return m

p03 = load("p03_question_tree")
p01, p02 = p03.p01, p03.p02          # the same module objects p03 uses

if __name__ == "__main__":
    prior = dict(p01.PRIOR)
    base = p02.value(prior)[1]
    print("   lost per question  RM per lead  gain   questions asked")
    for d in (0.0, 0.02, 0.05, 0.10, 0.15):
        p02.DROP = d
        _, v = p03.best(prior, p03.QS)
        n = p03.expected_questions(prior, p03.QS)
        print("   {:17.0%}  {:11,.0f}  {:+5.1%}  {:15.2f}".format(
            d, v, v / base - 1, n))
#    lost per question  RM per lead  gain   questions asked
#                   0%        2,894  +14.8%             3.26
#                   2%        2,748  +9.0%             1.73
#                   5%        2,584  +2.5%             1.00
#                  10%        2,522  +0.0%             0.00
#                  15%        2,522  +0.0%             0.00

买家很有耐心(不流失)时,最优树平均问 3.26 题,价值 RM 2,894,比直接报价高 14.8%;每题流失 2% 时,平均 1.73 题,高 9.0%;流失 5% 时只问 1.00 题,高 2.5%;流失 10% 以上时,最优做法就是直接报价。诊断式提问的价值,取决于买家愿意被问多少。

假说 · 尚未验证

作者的假说(尚未验证): 对主动来询价的买家,第一句话先问一个情境问题(团队与营收),再决定报价档位,比第一句话就报价,每个询价者的六个月收入至少高 5%。什么会证明它错: 在按询价者随机分配两种开场的实验中,收入差距小于 5%,或前者更低。

小结

最优提问树先问能改变决定的问题,问到不再能改变决定就停。在示意模型中,它比直接报价高 9.0%,平均只问 1.73 题;买家越没有耐心,值得问的题越少,流失率达到每题 10% 时就不该再问。

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).