Dynamic Effects of Agent Language on Customer Satisfaction & Purchase

5.0

One might wonder whether affective or cognitive language is more important overall. To address this question, we compare the proportions of positive versus negative areas of the beta curve for each functional feature.

For both customer satisfaction and purchase, the majority of both affective and cognitive language contributions are positive (Table 1). However, the relatively larger negative area for employee affective language suggests it is particularly important to know when to speak to customers more affectively (i.e., start and end, but not middle).

 

낮은 심각도에서 A-C-A-C 전략. 높은 심각도에서 A-C-C-A 전략.

 

5.1 Dynamic Effects of Agent Language on Customer Satisfaction
5.1 고객 만족에 대한 상담원 언어의 동적 효과

 

The ultimate result is represented by the βl(t) curves estimated from the sparse functional regression in (8). Predictors have a positive (negative) relationship with the outcome of interest when a given βl(t) curve and its confidence interval lie above (below) zero. We examine the relationship between agent affective and cognitive language and both customer satisfaction and purchase.

최종 결과는 (8)의 희소 기능 회귀에서 추정된 βl(t) 곡선으로 표시됩니다. 예측 변수는 주어진 βl(t) 곡선과 그 신뢰 구간이 0보다 위(아래)에 있을 때 관심 결과와 양(음) 관계를 갖습니다. 상담원의 정서적 언어와 인지적 언어, 그리고 고객 만족도와 구매 간의 관계를 조사합니다.

 

 As predicted, the model estimates for agent affective language (Figure 3a) reveal a positive relationship (pointwise 95% confidence interval above zero) between agent affective language and customer satisfaction at the conversation’s beginning and end (48.75% of the conversation). In contrast, customer satisfaction is higher when agents avoid affective language (pointwise 95% confidence interval below zero) during the middle of the call (51.25% of the conversation). Approximately two-thirds (63.59%) of the positive conversational contribution for agent affective language occurs at the start of the conversation, with the remainder (36.41%) at the conversation’s end.

 예측한 대로 상담원 감정 언어에 대한 모델 추정값(그림 3a)은 대화 시작과 끝(대화의 48.75%)에서 상담원 감정 언어와 고객 만족도 사이에 긍정적인 관계(0보다 높은 지점별 95% 신뢰 구간)를 나타냅니다. 대조적으로 상담원이 통화 중(대화의 51.25%) 동안 감정적인 언어(포인트와이즈 95% 신뢰 구간 0 미만)를 피하면 고객 만족도가 더 높아집니다. 상담원 감정 언어에 대한 긍정적인 대화 기여도의 약 3분의 2(63.59%)는 대화가 시작될 때 발생하고 나머지(36.41%)는 대화가 끝날 때 발생합니다.

 

The beta curve for agent cognitive language is quite different (Figure 3b). While customer satisfaction is higher by using affective language at the start of the call, speaking more rationally during this time appears to be costly. Cognitive language’s positive conversational impact (94.33% of its positive contribution) instead arises in the middle of the conversation.

상담원 인지 언어의 베타 곡선은 상당히 다릅니다(그림 3b). 통화를 시작할 때 감정적인 언어를 사용하면 고객 만족도가 더 높아지지만 이 시간에 더 합리적으로 말하는 것은 비용이 많이 드는 것처럼 보입니다. 대신 인지 언어의 긍정적인 대화 효과(긍정적인 기여의 94.33%)는 대화 중간에 발생합니다.

 

Parameter estimates for the customer satisfaction model are provided in Table A2 in the Web Appendix. The beta curves remain similar when we exclude the control variables (see Web Appendix Figure A2). Note that the average employee does not seem to follow the beta curves revealed. Instead, affective language is at its lowest point at the start of the call (Figure 1a), when it is particularly important, while cognitive language was near its lowest point between 10% and 40% into the conversation (Figure 1b), which is when customer satisfaction seems most likely to benefit from such language.

고객 만족도 모델에 대한 매개변수 추정치는 웹 부록의 표 A2에 나와 있습니다. 통제 변수를 제외해도 베타 곡선은 유사하게 유지됩니다(웹 부록 그림 A2 참조). 평균 직원은 공개된 베타 곡선을 따르지 않는 것 같습니다. 대신, 정서적 언어는 통화가 시작될 때 가장 낮은 지점(그림 1a)에 있는 반면, 인지 언어는 대화 중 10%에서 40% 사이의 가장 낮은 지점에 가까웠습니다(그림 1b). 고객 만족이 그러한 언어의 혜택을 받을 가능성이 가장 높을 때.

Figure 3: Beta Curves for Agent Affective (a) and Cognitive (b) Language in Relation to Customer Satisfaction (dotted lines: pointwise 95% confidence intervals) 고객 만족도와 관련된 상담원 감정(a) 및 인지(b) 언어에 대한 베타 곡선(점선: 포인트별 95% 신뢰 구간)

 

5.2 Dynamic Effects of Agent Language on Customer Purchase 
5.2 고객 구매에 대한 에이전트 언어의 동적 효과

While the customer satisfaction measure is useful given its occurrence immediately after the service interaction, more satisfied customers should also make more purchases (Zeithaml et al. 1996). Are the dynamic effects of agent affective and cognitive language sufficiently robust that they might be linked to post-call purchases over a longer period of time?
고객 만족도 측정은 서비스 상호 작용 직후에 발생한다는 점에서 유용하지만 더 만족한 고객은 더 많은 구매를 해야 합니다(Zeithaml et al. 1996). 상담원의 정서적 및 인지적 언어의 동적 효과가 장기간에 걸쳐 통화 후 구매와 연결될 수 있을 만큼 충분히 강력합니까?

We apply a functional Poisson regression model to estimate the relationship between agent affective and cognitive language and downstream customer purchase behavior. That is, we use a Log link function in (8) to relate the predictors with the mean of the order count. The Poisson model includes the same sets of functional and scalar variables as in the functional linear regression, and adds a control for each customer’s baseline buying behavior using the number of orders they placed up to 30 days prior to the conversation (Orders 30 Pre).

우리는 기능적 푸아송 회귀 모델을 적용하여 상담원의 정서적 언어 및 인지적 언어와 다운스트림 고객 구매 행동 간의 관계를 추정합니다. 즉, 예측 변수를 주문 수의 평균과 연결하기 위해 (8)에서 로그 링크 기능을 사용합니다. 푸아송 모델은 기능 선형 회귀에서와 동일한 기능 및 스칼라 변수 세트를 포함하고 대화(Orders 30 Pre) 전에 주문한 수를 사용하여 각 고객의 기준 구매 행동에 대한 제어를 추가합니다.

Results are similar to those observed for customer satisfaction (Figure 4; parameter estimates are provided in Web Appendix Table A3).8 Replication with purchase is valuable due to not only its behavioral (rather than self-reported) nature and its direct financial impact, but also its stronger inference of causality thanks to the greater time lag between it and the dynamic language predictors (i.e., up to 30 days).

결과는 고객 만족도에 대해 관찰된 결과와 유사합니다(그림 4, 매개변수 추정치는 웹 부록 표 A3에 제공됨). 구매 시 복제는 행동(자가 보고하기 보다는) 특성과 직접적인 재정적 영향으로 인해 가치가 있습니다. 뿐만 아니라 동적 언어 예측 변수(즉, 최대 30일) 사이의 더 큰 시차 덕분에 인과 관계에 대한 더 강력한 추론이 가능합니다.

 

Figure 4: Beta Curves for Agent Affective (a) and Cognitive (b) Language in Relation to Customer Order Count within 30 Days Post Interaction