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
第一章 买家隐藏的状态与每个问题的信息量
把一场高客单对话写成部分可观测的决策问题
一位潜在客户在私域里询问一项 RM 9,800 的企业牌照或 RM 29,800 的进阶方案。卖方看不到两件最要紧的事:他现在付得起多少(预算),以及他的问题有多急(紧迫度)。卖方能做的,是提问、观察回答、更新判断,最后决定报哪一档价,或者先不报价。这是一个部分可观测的马尔可夫决策过程(POMDP)(Kaelbling et al. 1998):真实状态隐藏,卖方根据信念(对状态的后验分布)行动。
在示意模型中,预算有三档:付不起 RM 9,800(先验 0.45)、付得起 RM 9,800(0.35)、付得起 RM 29,800(0.20);紧迫度高的先验是 0.4,与预算独立。拉克姆的 SPIN 法把诊断提问分成四类:情境(Situation)、问题(Problem)、影响(Implication)与需求回报(Need-payoff)(Rackham 1988)。每个问题的回答(是或否)在不同状态下出现的概率不同,它的信息量就是状态与回答之间的互信息 (Lindley 1956)。
推导
# Chapter 1. p01: What a buyer is hiding, and how much each question tells.
# Hidden state = budget x urgency. Budget: B0 cannot pay RM 9,800 now,
# B1 can pay 9800, B2 can pay 29800 (prior 0.45 / 0.35 / 0.20); urgency
# high with prior 0.4, independent. Each SPIN-style question has a yes/no
# answer whose probability depends on the state. A price pitched straight
# away is also a "question": the buyer's accept/decline answer is evidence.
# Information = mutual information I(state; answer), in bits.
import math
from math import log2
STATES = [(b, u) for b in (0, 1, 2) for u in (0, 1)]
PB, PU = (0.45, 0.35, 0.20), 0.4
PRIOR = {s: PB[s[0]] * (PU if s[1] else 1 - PU) for s in STATES}
# P(yes | budget, urgency) for each question
Q = {
"S situation (team > 5, revenue known)": lambda b, u: (0.1, 0.6, 0.9)[b],
"P problem (names a costly problem)": lambda b, u: 0.8 if u else 0.2,
"I implication (puts a monthly RM cost)":
lambda b, u: (0.1, 0.45, 0.75)[b] * (1.0 if u else 0.5),
"N need-payoff (says solving is worth it)":
lambda b, u: (0.2, 0.6, 0.85)[b] * (1.0 if u else 0.5),
}
# accept probability of a pitch, by price
def accept(price, b, u):
if price == 29800:
return (0.0, 0.0, 0.55 if u else 0.08)[b]
return (0.02, 0.6 if u else 0.12, 0.6 if u else 0.12)[b]
def entropy(p):
return -math.fsum(v * log2(v) for v in p.values() if v > 0)
def posterior(prior, like, yes):
post = {s: prior[s] * (like(*s) if yes else 1 - like(*s))
for s in prior}
z = math.fsum(post.values())
return {s: v / z for s, v in post.items()}, z
def info(prior, like):
h = entropy(prior)
out = 0.0
for yes in (True, False):
post, pz = posterior(prior, like, yes)
out += pz * (h - entropy(post))
return out
if __name__ == "__main__":
print(" uncertainty before talking: {:.3f} bits".format(entropy(PRIOR)))
for name, like in Q.items():
print(" {:<42} {:.3f} bits".format(name, info(PRIOR, like)))
for price in (9800, 29800):
like = lambda b, u, p=price: accept(p, b, u)
print(" pitch RM {:,} at once (accept or decline) {:.3f} bits"
.format(price, info(PRIOR, like)))
# uncertainty before talking: 2.484 bits
# S situation (team > 5, revenue known) 0.343 bits
# P problem (names a costly problem) 0.268 bits
# I implication (puts a monthly RM cost) 0.164 bits
# N need-payoff (says solving is worth it) 0.169 bits
# pitch RM 9,800 at once (accept or decline) 0.229 bits
# pitch RM 29,800 at once (accept or decline) 0.174 bits对话开始前,卖方对买家状态的不确定性是 2.484 比特。情境问题提供 0.343 比特,问题类 0.268 比特,影响类 0.164 比特,需求回报类 0.169 比特。
一个要改正的直觉:过早报价并非没有信息
常见的说法是,过早报价「什么也没问出来」。核对之后并非如此:直接报 RM 9,800,买家的接受或拒绝本身就提供 0.229 比特,比多数诊断问题还多;直接报 RM 29,800 提供 0.174 比特。过早报价的代价不在信息,而在于被拒绝之后的后果:一部分买家就此离开,留下来的人对第二个报价的接受率也打了折。
先问还是先报价
推导
# Chapter 1. p02: Ask first or pitch first? Exact expected value per lead,
# by enumerating every answer path. Ask the first k questions of p01 in
# the order S, P, I, N (each question loses 2% of buyers), then pick the
# plan with the most expected value under the posterior:
# (a) RM 29,800, and if declined a RM 9,800 downsell, (b) RM 9,800 only,
# (c) no pitch now: nurture. A nurtured buyer becomes urgent within six
# months with probability 0.35 and is then worth the best plan for the
# urgent version of his state, discounted by 0.9. After a declined pitch
# a buyer leaves for good (50% if not urgent, 20% if urgent: a pitch that
# comes too early feels pushy), and the rest accept a second offer (and
# any later one) only half as often (lost trust).
# The bound is a seller who knows the state (full information).
import math
from itertools import product
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")
DROP, TRUST, NURT, DISC = 0.02, 0.5, 0.35, 0.9
LEAVE = (0.5, 0.2) # by urgency: low, high
ORDER = list(p01.Q)
def pitch(s):
"""Immediate revenue and close probability of plans (a) and (b)."""
a29, a98 = p01.accept(29800, *s), p01.accept(9800, *s)
second = (1 - a29) * (1 - LEAVE[s[1]]) * TRUST * a98
return {"29800+downsell": (29800 * a29 + 9800 * second, a29 + second),
"9800 only": (9800 * a98, a98)}
def later(s):
b = s[0]
return NURT * DISC * max(v[0] for v in pitch((b, 1)).values())
def plans(s):
out = {}
for name, (r, c) in pitch(s).items():
out[name] = (r + (1 - c) * (1 - LEAVE[s[1]]) * TRUST * later(s), c)
out["nurture"] = (later(s), 0.0)
return out
def value(belief):
best = None
for name in ("29800+downsell", "9800 only", "nurture"):
r = math.fsum(belief[s] * plans(s)[name][0] for s in belief)
c = math.fsum(belief[s] * plans(s)[name][1] for s in belief)
if best is None or r > best[1]:
best = (name, r, c)
return best
def ask_then_offer(k):
rev = close = 0.0
used = {}
for answers in product((True, False), repeat=k):
belief, pr = dict(p01.PRIOR), 1.0
for q, yes in zip(ORDER[:k], answers):
belief, pz = p01.posterior(belief, p01.Q[q], yes)
pr *= pz
name, r, c = value(belief)
keep = (1 - DROP) ** k
rev += pr * keep * r
close += pr * keep * c
used[name[:5]] = used.get(name[:5], 0.0) + pr
return rev, close, used
if __name__ == "__main__":
print(" questions RM per lead close now plans used (share)")
for k in range(0, 5):
rev, close, used = ask_then_offer(k)
mix = ", ".join("{} {:.2f}".format(n, v)
for n, v in sorted(used.items()))
print(" {:9d} {:11,.0f} {:9.3f} {}".format(k, rev, close, mix))
orev = math.fsum(p01.PRIOR[s] * max(v[0] for v in plans(s).values())
for s in p01.STATES)
print(" full-information bound: RM {:,.0f} per lead".format(orev))
# questions RM per lead close now plans used (share)
# 0 2,522 0.108 29800 1.00
# 1 2,666 0.138 29800 0.43, 9800 0.57
# 2 2,661 0.111 29800 0.19, 9800 0.57, nurtu 0.24
# 3 2,671 0.121 29800 0.27, 9800 0.57, nurtu 0.17
# 4 2,669 0.115 29800 0.20, 9800 0.59, nurtu 0.21
# full-information bound: RM 3,466 per lead在示意模型中,被拒绝的报价会让一部分买家离开(不急的买家 50%,急的 20%),其余人对之后报价的接受率减半;每问一个问题,有 2% 的买家流失;暂不报价、转入培育的买家,六个月内有 0.35 的概率变急。
直接报价,每个潜在客户的期望价值是 RM 2,522,当下成交率 0.108。先问一个情境问题再报价,价值升到 RM 2,666,成交率升到 0.138。再多问一个问题,价值几乎不变(RM 2,661),当下成交率反而降到 0.111,因为有 0.24 的路径选择了「先培育、不报价」:诊断让卖方认出了还没准备好的买家,不去消耗他们。如果卖方确切知道每个买家的状态,价值的上限是 RM 3,466。
注
成交率不是唯一的目标。 诊断提问的价值有时表现为更高的当下成交率,有时表现为少报一次会被拒绝的价:后者让当下成交率下降,却保留了买家的未来价值。只看当下成交率,会低估诊断。
小结
在示意模型中,过早报价并非没有信息,它的代价是被拒绝之后的流失与信任折损。先问一个诊断问题,价值从 RM 2,522 升到 RM 2,666;完全信息的上限是 RM 3,466。这条上限给出诊断提问可能带来的最大收益。
参考文献
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).