Using Text Analytics to Derive Customer Service Management Benefits from Unstructured Data

Abstract

Estimates suggest that about 80% of today's enterprise data is unstructured. 2 Unlike structured data, which is tidy and mostly numeric, unstructured data is often textual and, therefore, messy. Unstructured data comprises documents, emails, instant messages or user posts and comments on social media, and presents a challenge to data miners; analyzing unstructured data is more complex, more ambivalent and more time consuming. Extracting knowledge from unstructured text, also known as text mining or text analytics, has been, and still is, limited by the ability of computers to understand the meaning of human language. 3 The written word, however, can provide valuable insights about the inner workings and environment of an organization; it has the potential to improve an organization's productivity while generating value for customers. Some organizations are successfully leveraging textual data as part of their analytics efforts. For example, pharmaceutical companies mine patents and scientific literature to improve product development processes. Doctors use text analytics to aid medical diagnosis by mining electronic health records. Insurance companies analyze claims and damage reports to mitigate risk or detect fraud.

Full text

December 2016 (15:4) | MIS Quarterly Executive 243
The Growth of Text Analytics1
Estimates suggest that about 80% of today’s enterprise data is unstructured.2 Unlike
structured data, which is tidy and mostly numeric, unstructured data is often textual and,
therefore, messy. Unstructured data comprises documents, emails, instant messages or user
posts and comments on social media, and presents a challenge to data miners; analyzing
unstructured data is more complex, more ambivalent and more time consuming. Extracting
knowledge from unstructured text, also known as text mining or text analytics, has been, and
still is, limited by the ability of computers to understand the meaning of human language.3
The written word, however, can provide valuable insights about the inner workings and
environment of an organization; it has the potential to improve an organization’s productivity
while generating value for customers.
Some organizations are successfully leveraging textual data as part of their analytics efforts.
          
product development processes. Doctors use text analytics to aid medical diagnosis by mining
electronic health records. Insurance companies analyze claims and damage reports to mitigate
risk or detect fraud.
1 Federico Pigni is the accepting senior editor for this article..
2 See, for example, “Discover the Digital Universe of Opportunities: Rich Data and the Increasing Value of the Internet of Things,”
EMC Digital Universe Study, EMC (with IDC), 2014, available at http://www.emc.com/leadership/digital-universe/index.htm#2014;
and Dhar, V. “Data Science and Prediction,” Communications of the ACM (56:12), 2013, pp. 64-73.
3 See Jurafsky, D. and James, H. Speech and Language Processing: An Introduction to Natural Language Processing, Computa-
tional Linguistics and Speech Recognition, Pearson, 2009.
Using Text Analytics to Derive Customer
Service Management Benefits from
Unstructured Data
Deriving value from structured data is now commonplace. The value of unstructured
textual data, however, remains mostly untapped and often unrecognized. This article
describes the text analytics journeys of three organizations in the customer service
management area. Based on their experiences, we provide four lessons that can guide
other organizations as they embark on their text analytics journeys.1
Oliver Müller
IT University of Copenhagen
(Denmark)
Stefan Debortoli
University of Liechtenstein
(Liechtenstein)
Iris Junglas
Florida State University
(U.S.)
Jan vom Brocke
University of Liechtenstein
(Liechtenstein)

244 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
Market forecasters have estimated that, by
2020, the market for text analytics will reach
$5.93 billion.4 Today, 30% of business analytics
projects include the delivery of insights based
on textual data. In some industries (e.g., banking,
retail, healthcare), growth rates for text analytics
of up to 50% are not uncommon.5 Within
organizations, text analytics is predominantly
used by marketing and customer service
units. Analyzing customer-related data in the
form, for example, of customer call notes or
service requests, allows organizations to better
understand their customers’ problems, react to
service issues in a timely fashion and proactively
plan process improvements that can alleviate
problems in the long term. For instance, by
      
into and within an organization, enterprises
can identify the issues of greatest concern to
their customers, monitor how issues evolve over
time, “slice and dice” issues by customer groups,
product categories or other dimensions and
recommend solutions for recurring issues.
This article examines three organizations
that aggressively use text analytics to improve
customer service management. All three use
text mining to analyze the content streams of
incoming service requests so they can better
understand their customers’ problems and, as
a consequence, improve their customer service
processes.
The three organizations come from different
geographies (U.S. and Europe) and industries
(research and education, manufacturing and
      
services, and maintenance and repair of physical
goods) and serve different types of customers
(internal and external). All three are faced with
similar challenges in harnessing the enormous
amounts of unstructured textual customer
feedback they receive during service interactions.
4 Text Analytics Market by Applications (Marketing & Customer
Experience Management, Data Analysis & Forecasting, Enterprise
Information Management, & Other Industry Specic Applications),
Deployment, Vertical, & by Region - Global Forecast to 2020, mar-
ketsandmarkets.com, 2015, available at http://www.marketsandmar-
kets.com/PressReleases/text-analytics.asp.
5 Duncan, A. D., Linden, A., Koehler-Kruener, H., Zaidi, E. and
Vashisth, S. Market Guide to Text Analytics, Gartner, 2015, available
at https://www.gartner.com/doc/3178917/market-guide-text-analytics.
A New Generation of Text
Analytics Solutions
The goal of enabling computers to understand
natural language is as old as commercial
computing itself.6 Since the early 1950s,
researchers have been trying to automate
the analysis of human language—with mixed
success. It wasn’t until the 2000s when the rise
of statistical machine learning approaches,
driven by the increasing availability of digitized
texts from the World Wide Web and by increases
in computing speed and memory, made text
analytics feasible for commercial applications.
Early natural language processing systems
were developed by and for computational
linguists and were well beyond the understanding
of the average business analyst. In contrast,
the latest generation of text analytics solutions
is more usable. These solutions possess four
distinctive features.
First, advanced text analytics tools are data-
driven rather than rule based. While early
natural language processing systems were
based on hand-written grammatical rules
for extracting meaning from texts, today’s
systems rely on statistical machine learning
techniques to automatically discover patterns
in large collections of texts. As a result, they can
process a wide variety of texts from different
domains without needing laborious updates to
an underlying rule base. They are also able to
handle ungrammatical texts—for example, posts
taken from online forums or messages from
social networking sites. In addition, the inductive
nature of today’s solutions ensures that they
learn over time—the more data they crunch, the
better the results. For example, one of the key
success factors of IBM’s Watson, a supercomputer
that beat two former (human) grand champions
on the game show “Jeopardy!,” is its ability to
integrate new data sources with minimal efforts
and to increase its performance with each new
data source added (e.g., books, dictionaries,
webpages, Wikipedia).
Second, the latest text analytics systems
are able to process textual data in real time
rather than in periodic batch runs. Increases in
6 See Jurafsky, D. and James, H., op. cit., 2009; and Manning, C.
D. and Schütze, H. Foundations of Statistical Natural Language
Processing, MIT Press,1999.

December 2016 (15:4) | MIS Quarterly Executive 245
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
computing power and storage capacities make
it feasible to analyze textual data as it streams
into an organization. For example, the Twitter
application programming interface (API) provides
access to more than 6,000 tweets a second, or 500
million tweets per day—a feature that Starbucks
uses to “listen to the voice of its customers.”7
During new product launches, Starbucks
performs sentiment analyses of Twitter and other
blogs and discussion forums to determine the
success of a new product in near real time and
to react to potential issues on the spot (e.g., by
reducing prices or changing the coffee blend).
Third, today’s text analytics solutions make
probabilistic inferences rather than drawing
deterministic conclusions. While rule-based
systems by nature provide a single answer,
probabilistic systems provide alternative
answers to any given query, each weighted with
a likelihood of being correct or relevant. For
example, IBM’s Watson generates hundreds of
potential answers (hypotheses) to any given
       
       
       
machine-generated solution or ignore it. This can
greatly increase the level of trust in situations
7 Watson, H. J. “Tutorial: Big Data Analytics: Concepts, Tech-
nologies, and Applications,” Communication of the Association for
Information Systems (34), 2014.
where the stakes are high—for example,
when trying to spot fake insurance claims or
performing medical diagnoses.
Fourth, until recently the outputs produced
by text analytics systems largely consisted of
cryptic codes, attached as annotations to the
original text. These outputs were primarily
meant to be read and interpreted by a computer
scientist or to serve as input for other algorithms.
The latest generation of text analytics systems,
in contrast, aggressively uses visual displays
that communicate results in an intuitive and
effective way. Twitter, for example, offers various
interactive graphics (http://interactive.twitter.
com) that enable non-specialists to explore and
interpret the contents of the stream of tweets
produced in reaction to various social, economic
or political events. The visual outputs of text
analytics tools, combined with the increased
overall usability of the tools, enable ordinary and
non-technical line-of-business people to analyze
written texts themselves.
These four features of modern text analytics
solutions are summarized in Table 1.
Each of the organizations discussed below
used the same tool, which had these four features
of advanced text analytics solutions. The tool was
developed as part of the research effort on which
this article is based and then adapted and tailored
Table 1: Features of the Latest Generation of Text Analytics Solutions
Feature Business Value
Foundaon
Data-driven instead of rule-based • Requires less manual eort
• Provides self-learning systems
Speed
Real-me instead of periodic
batch runs
• Enables connuous “listening” to textual
data streams
• Reduces reacon mes
Logic
Probabilisc inferences instead of
determinisc decisions
• Explores dierent answers to a given
problem
• Gains trust of users
Output
Intuive visualizaons instead of
crypc annotaons
• Provides fast and eecve communicaon
to end users
• Enables self-service text analycs for lines
of business communicaon to end users

246 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
to the needs of each organization. The core of the
tool is the ability to analyze large volumes of text
in a probabilistic, data-driven fashion; it does
the analysis in near real time and presents the
results in an easy-to-understand visual form. The
tool is based on the probabilistic topic-modeling
algorithm Latent Dirichlet Allocation (LDA),
which is described in the box. A base version of
the tool is freely available at MineMyText.com.8
8 For a more detailed description of LDA, see Blei, D. M. “Proba-
bilistic topic models,” Communications of the ACM (55:4), 2012, pp.
77-84.
Case Organizations
The three service organizations that
participated in our research varied in the type of
service offered, their audience and their industry
as well as their geography. Overall, we conducted
more than 25 interviews and workshops with
representatives of these organizations. The range
of informants included business and IT people
ranging from executive to operational levels.
Box: Text Analytics with Probabilistic Topic Modeling using
the Latent Dirichlet Allocation (LDA) Algorithm
The idea behind probabilisc topic modeling algorithms like Latent Dirichlet Allocaon (LDA) is that
texts exhibit mulple topics in dierent proporons.8 For example, the service cket displayed in the
top le corner of the gure below is reporng a failed pickup of a machine tool that should be sent to a
repair center. On the right hand side of the gure are words associated with various topics. For example,
there are words about return logiscs (e.g., “return,” “failed,” “pick-up”) and about maintenance (e.g.,
“maintenance,” “required,” “tool”). The bar chart at the boom le corner shows how the given service
ckets blends two topics (Topics 2 and 5) by providing probabilies of topic occurrences (i.e., 60% of the
cket talks about maintenance and 40% about return logiscs).
The primary advantage of LDA is its ability to automacally discover topics and their word distribuons
from large collecons of texts and to annotate individual texts with topic labels. Grouping and
aggregang the probabilisc topic assignments for thousands of service ckets enables, for example, a
service manager to quickly gain an overview of current issues and to track their development over me.

December 2016 (15:4) | MIS Quarterly Executive 247
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
Information Technology Services at
Florida State University
Information Technology Services (ITS) is
the central IT organization for Florida State
University (FSU). It provides IT support to 16
colleges and more than 110 centers, facilities,
labs and institutes, serving 11,000 faculty and
staff along with more than 40,000 students.
It offers more than 100 IT services, including
     
licensing, classroom and learning management
system support and online training. ITS is also
responsible for FSU’s overall IT strategy, setting
and enforcing standards, and managing and
ensuring compliance—a vital need for a state
institution.
As part of an enterprise-wide CRM installation
in 2011, ITS decided to add a component for
capturing and managing IT service requests
across business units. This component had
multiple goals. Its primary purpose was to
      
into one system and to manage their execution.
But it also built a knowledge base of “best
practices” and thus supported IT personnel by
providing an easy-to-use and transparent way of
solving IT service problems.
Today, ITS’s system captures, among other
things, textual descriptions and manual
     
email, network), when they were reported, who
reported them and which channels were used to
report them (e.g., email, phone). The system also
allows agents to add updates in the form of notes
and a description of the solution once a case is
resolved. Requests are routed to different work
groups depending on their category, type or detail
        
queue manager who triages requests and assigns
them to agents based on workload. At present,
more than 100,000 service requests and their
resolutions have been logged by the system since
its inception in 2011.
To better understand the issues and challenges
of the service delivery process, ITS prepares a
weekly report as part of its analytical efforts.
This report details, for example, the number of
requests issued in the past week by the various
business units, the number of open and closed
requests, along with the type of channel used to
send a request and the customer that requested
it. Overall, the report provides a snapshot of the
IT support group’s case load as well as how well
the group is meeting response time guidelines. It
thus provides a base for ITS to educate support
agents on proper case management.
The report represents a valuable summary
of the major elements of the stream of incoming
service requests, namely the who, when, where
and how of a service event. In contrast, insights
into the what and why of a service request are
captured in natural language within the textual
9 FSU uses text mining
        
why of service requests, thereby supplementing
     
report.
ITS applied “topic modeling” to more than
100,000 service tickets that had been received
        
seven most pressing topics to which ITS had
to respond. Each topic is represented by a list
of words that are associated with the topic
(displayed visually as a word cloud), a sample
sentence from a service ticket representative of
the topic, a descriptive label for the topic and the
prevalence of the topic plotted over time.
     
topics and their evolution over time provided
ITS with a detailed picture of the history and
evolution of service requests. For example, the
analysis showed that requests to register new
devices on the network (Topic 1) rose steadily
in frequency from October 2011 (2.2% of all
tickets) to October 2013 (4% of all tickets). This
period coincided with an initiative to register
all managed network devices by their unique
MAC address for accountability and ease of
management. In winter 2013, requests suddenly
declined again, returning to previously observed
levels (to less than 2%).
Other topics, like email forwarding (Topic
3), suddenly rose, which can be explained by
an email platform migration that triggered
      
time. Likewise, the rapid rise in phone services
requests (Topic 4) can be explained by an ongoing
9 For more information on the elements of digital data streams, see
Pigni, F., Piccoli, G. and Watson, R. “Digital Data Streams: Creating
Value from the Real-time Flow of Big Data,” California Management
Review (58:3), 2016.

248 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
Figure 1: ITS’s Seven Most Pressing Service Requests Topics
Words describing the topic Representative
sentence from
associated
service tickets
Descriptive
label
Timeline of topic prevalence
Topic 1 Good morning
hostmaster.
Please setup
MAC assigned
DHCP and
update DNS
as per the
following
informaon:
host name:
XXX*
Registering
new
devices
on the
network
Topic 2 Requesng
XXX to set up
an alternate
log in on the
Mac Computer
located in
Room XXX at
Alumni Center
Facilies.
Computer
belonged
to former
employee XXX.
It will need to
be backed up
and re-imaged
for a new
employee
whose
potenal start
date is June
21st. Please
contact me or
XXX for further
informaon.
Oce move

December 2016 (15:4) | MIS Quarterly Executive 249
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
Topic 3 I have two email
addresses at
FSU—a student
address and
an employee
address. My
student email is
being forwarded
to my gmail
address. Aer
the migraon,
however, both
my school and
work email are
being forwarded
to my gmail
address. Any
help would
be greatly
appreciated.
Forwarding
email
Topic 4 When user
makes a call on
his VOIP phone,
the “Hold”
opon does not
always display on
his screen. When
he called the
service desk, he
was able to put
me on hold and
it worked just
ne but when
he makes other
calls, the opon
will not appear;
it will be blank
above the buon
where Hold
should appear.
Phone
service
Topic 5 I am having
trouble prinng
to the HP 9050
(the sta B&W
printer). It was
working ne
yesterday and
it will print
from my oce
but it will not
print from any
of the 3 sta
machines. Any
help you can give
will be greatly
appreciated.
Thanks!
Printer

250 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
campus-wide initiative to move regular phone
lines to a VOIP service.
The process of spotting and explaining
historical service requests patterns raised
ITS’s interest in analyzing more recent topics.
For example, ITS was aware that installing
and updating software (Topic 7) was a current
     
prevalent this topic was and had been over the
years, providing additional empirical evidence
for continuing an existing project aimed at
streamlining software updates.
ITS also realized that the textual analysis
     
of its service request taxonomy that had been
underway since 2013. The existing taxonomy
used a three-tier architecture to classify service
      
a user was asked to classify the nature of the
      
        
between classroom and computer lab support,
email, or logins, passwords and system access.
The second and third tier (each comprising three
Topic 6 Need Building
access and
access to suite
3100 for two
Graduate Assts.
1. XXX FSU CARD
#: XXX 2. XXX FSU
CARD #: XXX
Building
access
cards
Topic 7 We have a
computer
that cannot
access certain
SharePoint
features (data
sheet view)
due to having a
64-bit system.
According
to Microso
support, this
machine needs
to download and
install the 2007
Oce System
Driver: Data
Connecvity
Components.
This installaon
requires an
admin password.
Please advise.
Soware
updates
*Due to the sensive nature of the data, we have redacted porons of these entries with “XXX”

December 2016 (15:4) | MIS Quarterly Executive 251
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
      
user to specify their request even further. The text
  

ITS’s use of text analytics provided a glimpse
into the possibility of extracting insights from
   
elements of digital data streams. By using text
analytics, ITS was not only able to supplement
insights provided by the traditional quantitative
report but also to validate efforts already
underway.
Customer Service at Hilti
Hilti is a leading-edge technology provider to
construction professionals, located in Schaan,
Liechtenstein. Its many products, including
power drills and demolition hammers, are
the preferred tools used by the construction
and building maintenance industry around
the globe. Hilti markets its high-tech products
through a direct sales model, with two-thirds
of its 20,000 employees having daily customer
contact involving more than 200,000 customer
interactions. A large part of these interactions are
handled by Hilti’s multi-channel contact center,
where agents capture each of the one million
incoming service requests per year as a so-called
customer care note in a CRM system and manage
its resolution.
The global contact center is known for its fast
response in handling individual customer issues.
However, extracting strategic knowledge from the
steadily expanding database of transcribed phone
calls and incoming emails requires enormous
efforts. Once a month, data is extracted from
the CRM system and imported into a custom
database for further analysis. The objective of this
analysis is to unearth the root causes of recurring
product failures, monitor persistent complaints
about customer service and identify potential
business process improvements. Findings from
the analysis are discussed as part of a quarterly
meeting of executives from all business units.
The analysis for the quarterly meetings is
carried out manually, using a combination of
     
     
to oversights. If, for example, the analyst
uses a search term different from the terms
used by the originator of a service request to
      
become distorted. This well-known “vocabulary
problem”10 of text-based human-computer
interaction is especially severe in large and
heterogeneous domains where many different
types of users interact with a system. Hilti’s
service requests, for example, include many
technical terms and abbreviations, and are
written by a heterogeneous group of users,
ranging from service agents with their own
varying backgrounds to customers from different
industries and geographies.
Another problem that has made the manual
analysis cumbersome is rooted in Hilti’s
service process. Agents are required to classify
requests into just one of 20 categories, even if
a request matches more than one category or
an appropriate proper category does not exist.
       

data, might lead to a biased picture of customer
feedback and even prevent newly emerging issues
from being detected in a timely manner.
The time lag between the occurrence of
events and their analysis, combined with the
largely manual analysis of patterns, meant that
Hilti was missing out on many opportunities for
capturing business value from the digital streams
of customer feedback. As part of our research
  
be addressed by a probabilistic and data-driven
analysis of customer care notes through the use
of topic modeling. A team of six, including two
from Hilti’s CRM team, two from Hilti’s IT team
and two researchers, was formed to explore the
      
was tested in Hilti’s German market during spring
2014 and captured about 50,000 service requests
over a six-month period. Later that year, a second
prototype with a total of 30,000 service requests
from Hilti’s French market and spanning a period
of four months, was built.
These prototypes were built around a
   
service requests to be summarized and explored
      
the overall distribution of topics, provided the
ability to drill down into individual topics, traced
10 Furnas, G. W., Landauer, T. K., Gomez, L. M. and Dumais, S. T.
“The Vocabulary Problem in Human-System Communication,” Com-
munications of the ACM (30:11), 1987, pp. 964-971.

252 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
each topic’s evolution over time and “sliced and
diced” topics by geography or customer segments.
Figure 2 depicts the entry point into Hilti’s
dashboard. It shows a tree map that visualizes
the prevalence of a particular topic by the size
of its rectangle and its semantic similarity with
other topics by its position and color. This kind
of visualization is based on the fact that some
topics are similar to each other in meaning and
sometimes even overlap. For example, Topics 1
and 13, displayed as light green in Figure 2, are
both logistics related: one represents delivery
failures, the other delayed deliveries. Likewise,
    
maintenance and repair of machine tools), and

By clicking on one of the topics in the tree
map, users are led to a details page that provides
further information about the topic. Among
other things, this page shows the words that are
associated with the topic, as well as its evolution
over time. Figure 3, for instance, shows the details
for Topic 5 (“Pick-up failed because of wrong
address”). It includes a word cloud containing the
most probable words associated with the topic,
along with a stacked line chart representing the
development of the topic over time, broken down
by customer segments.
For example, the graph documents a rise in
failed pick-ups for customers in the “E” segment
(brown area). Drilling further into the data for
the “E” segment and displaying affected customer
names and locations (not shown in Figure 3)
allows Hilti to identify potential root causes of the
failures.
For the future, Hilti envisions enhancing
the dashboard by providing automated alerts
triggered when a topic varies by more than two
standard deviations from the norm. These alerts
will help Hilti to address emerging customer
issues in near real time.
A major part of the text analytics project at
Hilti was dedicated to evaluating and improving
the accuracy of the topic-modeling and
visualization algorithms. The solution passed
through several iterative development cycles
to identify the best possible set of parameters.
These included determining the ideal granularity
of topic modeling by increasing or decreasing
the number of topics to be extracted or varying
the natural language preprocessing pipeline
to remove “noise” from the input data. Output
from the solution was also repeatedly compared
to categorizations made by human experts.
Experiments showed that the solution agreed
with expert judgments in 77% of cases, which
Figure 2: The Tree Map of Topics is Part of Hilti’s Text Analytics Dashboard

December 2016 (15:4) | MIS Quarterly Executive 253
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
Hilti deemed to be more than satisfactory,
given the time savings that could be realized by
automating the text analysis process.
Hilti’s main goals in mining unstructured
digital data streams were related to achieving
     
 
Hilti was interested in streamlining internal
operations to save costs, while, at the same time,
improving customer service. By leveraging text
mining and visual analytics, Hilti increased its
ability to interpret unstructured textual streams,
which enabled it to monitor the feedback of
thousands of customers in real time and provided
the possibility of responding to patterns and
trends in an ad hoc fashion.
IT Service Outsourcing at Inventx
Inventx AG is a Swiss IT company that
specializes in providing IT services to leading
      
service management solutions, its portfolio
comprises the planning, implementation,
integration, hosting and maintenance of core
banking solutions. In addition, Inventx runs one
of the most modern data centers in Switzerland.
A central part of its business model is providing
outsourcing services to banks for back-, middle-
    
continuous availability and high performance of
these applications is vital both for Inventx and for
its customers. To ensure superior service, Inventx
operates its own helpdesk for managing and
resolving IT service requests.
A major part of a service agent’s daily routine
is driven by scripts that capture how a particular
IT problem can be solved. In the past, these
scripts were written by agents for agents, using
Microsoft OneNote. OneNote is a collaboration
tool that allows users to capture typed as well
as handwritten notes, screenshots and audio
snippets, and to share them with others. These
scripts are written in a free format and edited
by hand; their existence, accuracy and currency
depend highly on the intrinsic motivation
of agents, who volunteer to contribute their
knowledge to OneNote.
Figure 3: Drilling Down into One Topic

254 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
Inventx decided to use text analytics to
improve its IT service management process
with two goals in mind. As well as automating
the process of categorizing and monitoring the
      
its helpdesk, Inventx wanted to embed text
analytics directly into the process of solving
individual customer issues. Currently, about
60% of incoming service requests are resolved
by the helpdesk immediately. However, 40%
     
Identifying the appropriate expert and routing
requests in a reliable and timely fashion is a
non-trivial task, given the complexity of the IT
systems Inventx is operating and the importance
of these systems for its banking clients. Inventx
has implemented a “recommender” system that
cannot only suggest the most appropriate expert
for solving a problem, but also possible solutions.
This system is based on a set of more than
100,000 historical service tickets that include
data about who solved which problem, how the
problem was solved and how much time the
resolution required. Using case-based reasoning,
the system derives its intelligence from mining
the textual descriptions of these past problems
and solutions.
When a new service request arrives, the

similar prior requests and then identifying
experts and solutions that have proven to be
       
4). To achieve this, the outputs of the topic
modeling analysis (i.e., the statistical assignment
of topics to requests) are used to match new and
past service requests—an approach that has
been shown to work well for electronic health
records.11
However, Inventx learned early on (as did
ITS and Hilti) that the inadequate quality of
data in its historical service requests hinders
the implementation of this approach. While
analyzing historical requests, Inventx became
aware that the solution description for individual
requests was often too short and sometimes even

“task completed,” without specifying exactly how
the task had been completed. As a consequence,
Inventx started promoting usage of the solution
       
increasing the quality of its knowledge base.
11 Miotto, R. and Weng, C. “Case-based reasoning using electronic
health records eciently identies eligible patients for clinical trials,”
Journal of the American Medical Informatics Association (22:e1),
2015, pp. e141-e150.
Figure 4: Architecture of a Topic-based Solution Recommender

December 2016 (15:4) | MIS Quarterly Executive 255
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
Comparing Inventx’s use of text analytics
with the other two cases reveals one major
difference. Both Hilti and ITS focused on
extracting novel insights from their digital data
streams, whereas Inventx went beyond that.
Based on real-time analysis of its textual data
streams, Inventx implemented a set of automated
processes, such as the automated routing of
requests to the most suitable agent or the data-
driven recommendation engine to aid agents in
providing the best answers possible.
Interestingly, Inventx was not interested
in visualization. Even though its text analytics
system is based on the same tool used by ITS and
Hilti, the tool is hidden in the back end and not
used by human analysts. This approach of using
digital data streams to drive, and even automate,
decision making has been referred to as “process-
to-action.” In contrast, ITS’s and Hilti’s approach

new and unprecedented insights are gained into
the inner workings and the environment of the
organization.12
Lessons Learned
The goal of all three organizations that we
accompanied on their text analytics journeys
was to extract insights from the contents of
unstructured digital data streams to better
understand their customers’ needs and improve
their service processes. Despite their differences
in the type of service offered, their customers,
their industry and their geography, as well as the
initiatives that were triggered as part of the text
analysis, four general lessons can be drawn from
their experiences.
Lesson 1: Position Text Analytics in
Business Units, Not IT
      
whether their analytics initiatives should be
positioned in the central IT department or
embedded in a business unit. Prior research
      
capabilities that can help organizations to make
this decision.13 The dimensions are dataset,
toolset, skillset and mindset, and it has been
12 Pigni, F., Piccoli, G. and Watson, R., op. cit., 2016.
13 See Barlow, M. “The Culture of Big Data,” O’Reilly Media,
Inc., 2013; and Pigni, F., Piccoli, G. and Watson, R., op. cit., 2016.
argued that an organization’s analytics journey
typically goes from dataset to toolset to skillset
to mindset. This perspective would make the
IT department the obvious starting point for
running text analytics because it has two of the
four capability dimensions: toolset (the products
and platforms for analyzing the contents of
digital data streams) and skillset (the coding and
statistical skills for detecting patterns in data
streams). Indeed, until recently, performing text
analytics required extensive technical capabilities.
Early tools were cumbersome to run, lacked
documentation and provided output that lacked
intuitive visualizations. This is no surprise, given
that early implementations were developed by
and for computational linguistics researchers.
Our experience, however, suggests that today’s
text analytics tools can be used successfully
by tech-savvy business people, who typically
have a better knowledge about what datasets
exist (dataset) and how business value can be
generated by harvesting the datasets (mindset).
If they have access to sophisticated and user-
friendly tools (toolset), either provided in the
cloud or in-house, and are able to learn the
necessary skills (skillset), we found they are in a
       
from text analytics. In fact, across all three
organizations we studied, business users had
    
the outputs of topic modeling algorithms when
applied to a dataset they were familiar with.
      
business users often led to an “aha” effect and
triggered lively discussions. They realized that
text analytics can enable them to discover in
a matter of minutes the thematic structure of
more than 100,000 service requests and can
provide tremendous opportunities for service
improvements.
The advantage of positioning text analytics
capabilities in a business unit (e.g., marketing and
sales, customer service) is rooted in the fact that
analytical projects are less about rolling out IT
tools and more about understanding how these
tools might be used for creating business value.14
Business users know best which questions to
investigate, which datasets to explore and how
14 For a more detailed discussion of this topic, see Marchand, D.
A. and Peppard, J. “Why IT Fumbles Analytics,” Harvard Business
Review (91:1), 2013, pp. 104-112.

256 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
to translate insights into actions. Moreover,
introducing text analytics into an organization
this way generates trust among users and
management alike, and ensures that use cases
for text analytics are well understood. Of course,
most business departments will need help in
terms of tools and skills, but the availability of
easy-to-use, cloud-based analytics tools makes
these challenges manageable.
Lesson 2: Learn the Basics of Text
Analytics by Analyzing the Past
      
potentially has high organizational value but has
been impossible to analyze with traditional tools,
it is time to learn the basics of the text analytics
tool. Our research shows that it makes sense
to start by analyzing an historical dataset (e.g.,
last year’s service tickets) before trying to gain
insights from data streams in real time. This
approach provides a way to experiment with the
functionalities of the tool and to learn how to read
       
past, users also learn how to detect patterns and
trends in the data—for example, by identifying
the most prevalent topics in a body of text and
tracking their evolution over time.
To facilitate learning, it is essential to provide
opportunities for sharing individual insights
derived from the tool’s outputs. Whenever text-
mining results were presented at meetings in our
case organizations, typically with representatives
from all levels and business functions present,
lively discussions ensued immediately. Everybody
had a thought to share; they would either throw
in explanations as to why topics spiked or ideas
        
Overall, these initial meetings could be described
as inspired and animated. Even if the discovered
patterns and trends where not necessarily
surprising, having empirical evidence—instead
of intuitions and anecdotes—was perceived as
incredibly useful by all participants and formed
a fertile basis for further discussions and ideas.
While retrospective analysis might be perceived
by some as providing only marginal returns, it is
nevertheless an essential stage for organizations
to pass through because, without it, they will not
be able to fully exploit the text analytics tool.
Moreover, analyzing the past might reveal
some unpleasant IT systemic truths. In all three
case organizations, the initial data-extraction
and cleaning process revealed what the text
       
data quality was sometimes inadequate. As
described above, the written solutions to service
problems were often inadequate, with just “task
completed” rather than a detailed description.
While educating and training users might go
some way toward solving this problem, the
existence of inadequate textual data is indicative
of something more systemic. For example, two of
the three organizations had chosen to implement
the best practices model embedded in their
respective ERP system and had not considered
customizing any major component when they did
this. The consequences of this choice only became
apparent years later—when the data was heading
for text analysis.
Lesson 3: Move Eventually Toward
Real-time Monitoring of Textual
Data Stream
Once organizations have learned the basics
of text analytics and have set up the required
technical infrastructure, they typically begin to
think about extending the analysis to monitoring
textual data streams in real time. Often, they
start by re-running text analytics at shorter and
shorter intervals (e.g., weeks, then days, then
hours) to get a feel for how volatile the data
streams are. Typically, in this period, a lot of ideas
for exploiting the analytics outputs are generated,
and quick wins are often proposed to justify
further investments. Some of the organizations
we studied developed interactive visualizations
for tracking customer issues in real time—for
example, when a new product or service was
introduced or features of existing products were
updated. Others set up alerts triggered when
there was a spike in responses for a particular
topic.
An essential prerequisite for moving toward
real-time monitoring is that the textual data
must be automatically streamed into the
analytics tool. This often involves implementing
adapters to periodically extract data and
accompanying metadata from transactional
systems, implementing streaming APIs to enable
real-time analytics, transforming data structures
and loading data into the text analytics tool. All
three case organizations found that locating

December 2016 (15:4) | MIS Quarterly Executive 257
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
the appropriate data, querying the appropriate
       
set was complete and coherent typically took
more time than expected. They had to identify
programmers with the skills to access the back-
end transactional systems.
Lesson 4: Embed Textual Analytics
into Operational Business Processes
Once organizations have the capability to
automatically analyze the written word, it is
time to translate the insights derived from text
analytics into actions. Increasingly, this means
making existing business processes smarter by
      
For instance, algorithmic analysis of the content
of service requests can enable the automated
routing of tickets to the most appropriate service
agent or the recommendation of solutions
based on how requests were handled in the
past. The objectives an organization decides to
pursue at this point in its text analytics journey
typically have long-term impacts and are aimed
at delivering high business value: changing the
nature of the service process.
This stage of the journey entails more than
“just” extracting insights from data streams
meant to be consumed by a human decision
maker—it means changing the mindset of the
organization to embrace algorithmic decision
making and the automated initiation of actions.
In the projects we accompanied, a somewhat
philosophical dilemma quickly became apparent.
Business managers wondered: “To what extent
are we willing to let technology determine the
way we perform our service operations?”  
this question was meant to delineate between
ideas that are doable in the short term, like
building a dashboard, and those that might be
conceivable in the long term, such as developing
a solution recommender system. After a while,
however, organizations understood that topic
modeling is just the tip of the iceberg. Combining
these algorithms with technologies for speech
recognition, machine-learning algorithms for
making predictions, and avatars used as virtual
service agents has the potential to completely
automate the management of customer service
operations.
Thus, text analytics holds the promise of
     
do as part of a reengineering effort. But unlike
humans, algorithms are able to reengineer in
      
processes has been described as “Analytics 3.0”
or “machine-reengineered processes.”15 But this
approach poses a dilemma for organizations
and seems to be an ideological step that
organizations—and society in general—are
struggling with. After all, what happens if a
machine-recommended solution to a critical
customer problem ends in disaster? Finding the
right middle ground will therefore present a
major challenge in the future.
Concluding Comments
Text analytics is a journey. As organizations
learn the basics of the tools, they also learn
about themselves. Each journey starts by taking a
detailed look at historical textual data, continues
with monitoring data as it comes in and ends with
decisions that change the nature of how work is
carried out. We studied three organizations in
detail that applied text analytics to the contents
of their service requests. While each organization
had a different analytical goal, their experiences
were surprisingly similar. Each set out to achieve
service improvements for its customers, although
      
text analytics. Today’s advanced text analytics
techniques are domainfree—that is, they try to
extract knowledge from data, regardless of its
content.
About the Authors
Oliver Müller
Oliver Müller (oliver.mueller@itu.dk) is
an associate professor in the Information
Management section at the IT University of
Copenhagen. As part of his research, he studies
how organizations can create value with (big)
data and analytics. He particularly focuses on the
application of methods and tools for extracting
knowledge from unstructured data, from both
15 See Davenport, T. H. “Analytics 3.0,“ Harvard Business Review,
2013, available at https://hbr.org/2013/12/analytics-30; and Wilson,
H. J., Alter, A. and Shukla, P “Companies Are Reimagining Business
Processes with Algorithms,” Harvard Business Review, February 8,
2016, available at https://hbr.org/2016/02/companies-are-reimagin-
ing-business-processes-with-algorithms.

258 MIS Quarterly Executive | December 2016 (15:4) misqe.org | © 2016 University of Minnesota
Using Text Analytics to Derive Customer Service Management Benets from Unstructured Data
social media and enterprise-internal data sources.
His research has been published in the Journal of
the Association of Information Systems, European
Journal of Information Systems, Communication of
the Association of Information Systems and various
others.
Iris Junglas
Iris Junglas (ijunglas@fsu.edu) is an associate
professor at Florida State University. Her research
interests cover a broad spectrum of technology
innovations, but most prominent are the areas
of u-commerce (ubiquitous commerce), the
consumerization of IT and business analytics.
Her research has been published in the European
Journal of Information Systems, Information
Systems Journal, Journal of the Association of
Information Systems, MIS Quarterly, Journal
of Strategic Information Systems and various
others. She is a senior associate editor for the
European Journal of Information Systems and an
editorial board member of the Journal of Strategic
Information Systems and MIS Quarterly Executive.
Stefan Debortoli
Stefan Debortoli (stefan.debortoli@uni.li) is
an associate researcher at the University of
Liechtenstein; he is also the CEO and co-founder
of MineMyText.com. His studies focus on big data
analytics as a new inquiry strategy in information
systems research. His work has been published
in the European Journal of Information Systems,
Communications of the Association of Information
Systems and Business & Information Systems
Engineering.
Jan vom Brocke
Jan vom Brocke (jan.vom.brocke@uni.li) is
the Hilti Endowed Chair of Business Process
Management, Director of the Institute of
Information Systems and Vice-President at
the University of Liechtenstein. His research
focuses on digital innovation and transformation,
business process management and data analytics,
and has been published in, among others,
the European Journal of Information Systems,
Journal of Management Information Systems
and MIS Quarterly. He has published seminal
books, including the International Handbook on
Business Process Management as well as BPM -
Driving Innovation in a Digital World. He has held
various editorial roles and leadership positions in
information systems research and education.