Improving customer routing in contact centers: An automated triage design based on text analytics

• Author: Paulo Goes,Guangzhi Shang,Noyan Ilk
• Date: 01 July 2020
• Published date: 01 July 2020
• DOI: http://doi.org/10.1002/joom.1084

RESEARCH ARTICLE

Improving customer routing in contact centers: An

automated triage design based on text analytics

Noyan Ilk

1

| Guangzhi Shang

1

| Paulo Goes

2

1

Department of Business Analytics,

Information Systems, and Supply Chain,

College of Business, Florida State

University, Tallahassee, Florida

2

Department of Management Information

Systems, Eller College of Management,

University of Arizona, Tucson, Arizona

Correspondence

Guangzhi Shang, Department of Business

Analytics, Information Systems, and

Supply Chain, College of Business,

Florida State University, Tallahassee, FL.

Email: gshang@business.fsu.edu

Handling Editor: Aravind

Chandrasekaran

Abstract

We propose an automated triage design for intelligent customer routing in

live-chat contact centers and demonstrate its implementation using a real-

world data set from an S&P 500 firm. The proposed design emerges as a syn-

thesis of text analytics and predictive machine learning methods. Using

numerical experiments based on the simulation of the firm's contact center,

we demonstrate the service level, time, and labor cost benefits of the auto-

mated design over two other triage designs (i.e., customer choice triage and

human expert triage) that are commonly employed in the real world. Through

additional analyses, we explore the generalizability of the automated design

for creating solutions for different types of communication channels. Our work

has implications for managing customer relations under emerging communi-

cation technologies (e.g., live-chat, e-mail, and social media) and more broadly

for demonstrating the use of text analytics and machine learning to improve

Operations Management practice.

KEYWORDS

contact center, customer routing, design science, machine learning, service quality management,

text analytics

1|INTRODUCTION

Contact centers of medium-to-large sized firms com-

monly receive a large volume of service inquiries on a

broad range of topics (Aks¸in, Armony, & Mehrotra,

2007). At the same time, their service agents can resolve

these inquiries faster and better when they are special-

ized in a narrow set of skills (Gans, Koole, & Man-

delbaum, 2003). These two characteristics of contact

centers give rise to the need for a routing mechanism that

can help match service inquiries with agent specialties

and direct customers to the correct service departments.

The extant literature on customer routing is rich with

findings derived from theoretical models, but relatively

sparse on implementation examples that can illustrate the

real-world value of such findings.

1

Furthermore, the theo-

retical customer routing literature, known as multiskill

routing (see Aks¸in, Karaesmen, & Lerzan Örmeci, 2007;

Koole & Pot, 2006 for reviews), focuses on the assignment

optimization problem, assuming the service requirements

of all incoming customers are perfectly observable. Unfor-

tunately, this assumption is not sound in practice. While

the typical contact center interface allows customers to

indicate their inquiry type, this indication is often highly

inaccurate due to a lack of understanding of the root cau-

ses of problems and agents' expertise (NAQC, 2010).

Customer–agent mismatches lead to inflated service times

that waste agent effort, as well as to prolonged customer

wait times that diminish service quality levels. According

to industry reports, these negative consequences could

cost contact centers millions of dollars each year (Gupta,

McMahon, Jain, & Kanagasabai, 2008).

In this study, we explore how combining machine

learning with text mining can improve customer routing

Received: 4 February 2019 Revised: 30 January 2020 Accepted: 5 February 2020

DOI: 10.1002/joom.1084

J Oper Manag. 2020;66:553–577. wileyonlinelibrary.com/journal/joom © 2020 Association for Supply Chain Management, Inc. 553

accuracy (i.e., decrease transfer rates) in live-chat contact

centers. While originated in the telephone contact center

context, the issue of customer–agent mismatch looms

even larger as companies expand their online customer

service channels—an industry trend in the past decade

(Leggett & Schoeller, 2015). Live-chat contact centers, in

which agents interact with customers through a text-

based instant-messaging format, are particularly popular

among firms with online presence. According to industry

reports, live-chat is rated by the majority of customers

(41%) as the most preferred service channel (Kayako,

2017) and in terms of adoption, it is predicted to surpass

traditional channels (e.g., telephone) by 2021 (Young,

2020). In this online context, the counterpart to a tele-

phone's menu-based selection is a web-based drop-down

menu interface, which, similar to the keys pressed in a

touch-tone system, is commonly used as the input to

route customers to service departments. Accordingly, the

customer routing process in the online environment

is prone to the same root causes that lead to customer–

agent mismatches, hence service transfers.

Transfer rate is a key performance indicator (KPI) that

measures the percentage of customer contacts that could

not be resolved by the initial service agents and hence are

transferred (a.k.a., rerouted) to other agents with the cor-

rect skill set. Given its importance, it would be misleading

to purely attribute a high transfer rate to operational over-

sight or firm incompetence. In fact, our discussions with

industry executives revealed that transfer rates, along

with a number of other fundamental KPIs, are among the

most closely monitored and studied operational metrics

in contact centers. The question, then, is why do firms

have difficulty curtailing this type of errors? The answer

to this question lies within the ill structured nature

(Simon, 1973) of the current routing design. In many con-

tact centers, customer routing decisions are carried out

through a technology known as the interactive voice

response (IVR) system (Irani, Shajahan, & Kemke, 2006).

This time-worn technology relies on customers' ability to

match the root cause of their calls with the list of options

presented in the IVR menus (e.g., touch-tone menu for

telephone and drop-down menu for live-chat). In other

words, customers are put in charge of the routing deci-

sions, and firms have limited opportunity to identify and

correct matching mistakes before the agent–customer

interaction commences. In their pioneering work,

Holmström, Ketokivi, and Hameri (2009) succinctly

describe such circumstances as “ill structured decision

situations,”in which decision makers (e.g., firms) are

aware of the goals (e.g., addressing the mismatching prob-

lem), yet are limited in their ability to achieve these goals

with existing solution designs and technologies. Accord-

ingly, the authors advocate for the use of an academic,

yet practical, problem solving approach—that is, the

design science approach, to help practitioners address

such field problems.

Empowered by this approach, we investigate if an auto-

mated triage design can improve the routing accuracy by uti-

lizing customer-provided problem description texts—an

input that is often available in live-chat but not in telephone

contact centers. At the core of the design lies a text-analytic

machine learning model that helps improve customer–agent

matching and hence reduce transfer rates. The model utilizes

customers' open-ended responses to a problem description

request (e.g., “please briefly describe your problem”)toauto-

matically predict the correct service department to handle

the problem. In this aspect, the proposed design resembles

an automated version of human expert triage (Freeman,

Savva, & Scholtes, 2017; Lee, Pinker, & Shumsky, 2012;

Shumsky & Pinker, 2003) in the contact center environment.

An important building block of design science

research is pragmatic validity (Denyer, Tranfield, & Van

Aken, 2008; Groop, Ketokivi, Gupta, & Holmström, 2017;

Kaipia, Holmström, Småros, & Rajala, 2017). To demon-

strate its validity and performance, we implement the

proposed design and conduct simulation experiments

based on a real-world data set from the live-chat contact

center of an S&P 500 firm. The data are used for both

implementing the proposed design (i.e., building and

evaluating the text-analytic machine learning model) and

tuning the simulation parameters. The simulation experi-

ments allow us to investigate the performance of the pro-

posed automated triage design over two other solution s

commonly employed in the real world: (a) customer

choice triage, in which customers self-select the routing

destinations using a menu-based system and (b) human

expert triage, in which the rout ing destinations are deter-

mined by human experts (i.e., gatekeepers) employed

by the firm. Our findings highlight the benefits of the

proposed design in terms of reduc ing queue abandon-

ment and waiting times and improving the overall ser-

vice level.

Itisimportanttopointoutthatwhilethemotivating

force that initiated our research is practice-driven, we

pay attention to the generalizability of the proposed

design as well as the theoretical novelty of our findings.

While our solution design is motivated and tested for a

live-chat contact center, its applicability can be general-

ized into contact centers with different types of commu-

nication channels such as telephone and social media

(e.g., by utilizing speech-to-text technologies and ana-

lyzing posts on social media platforms). However,

customer service requests arriving via telephone and

social media channels might have different characteris-

tics such as shorter problem description lengths and

more variability in arrival rates. We conduct robustness

554 ILK ET AL.