Customer-focused service management for contact centers

M. Bhide S. Negi L. V. Subramaniam H. Gupta

Customer-focused service management results when contact center metrics are based on the customer’s expectations. Today, contact centers are a primary touch point between a customer and an enterprise. We identify customer communications, business intelligence, service improvement, and business impact as the four main components of service management. Using real life examples, we show that the voice of the customer is the main driving force for each of these components.

Introduction In order to drive sustainable, profitable organic growth and competitive differentiation, organizations must integrate and align the way they treat customers with their overall service strategy. As a primary customer touch point, the contact center will have to play a more proactive role than what it has traditionally done. Contact center or call center is a general term for help desks, information lines, and customer service centers. It includes dialog-based (both voice and online chat) and e-mail support provided to customers by professional agents. Traditionally, operational attributes such as average handling time for calls, conversion ratio (ratio of successful to total calls), and customer surveys have been the main metrics that drive contact center operations. While focusing on these metrics alone might be sufficient from an operational point of view, these metrics are poor proxies for understanding customer expectations. In this paper, we show that there is a common underlying theme that affects all of these metrics. Understanding this theme can help uncover customer expectations hidden inside customer communications. We believe that by acting without fully understanding the customer’s expectations and focusing mostly on operational attributes, contact centers today have more or less left the customer out of the service management equation. In order to better understand customer expectations, contact centers should adopt a customer-centric approach. Such an approach will enable contact centers to increase customer satisfaction levels while maintaining the operational metrics at optimal levels. To become customer centric, it is important to analyze voice-of-thecustomer (VoC) data (e-mails, surveys, call transcripts)

collected by contact centers to capture both overt and latent customer expectations. The VoC provides a dynamic view of customer needs, problems, opinions, sentiments, inclinations, and propensities that change from time to time. Access to these variables provides an opportunity to dynamically optimize and control the entire business process more effectively. Thus, an analysis of the VoC helps contact centers understand the common theme that affects all the operational attributes. After analyzing the VoC and deriving insights, a customer-centric approach uses these insights to improve the service levels and achieve higher customer satisfaction levels. To do this, it is important to learn what the customers are unhappy about and why they are unhappy. The answers to these questions can vary from one enterprise to another. However, the tools and techniques required to answer these questions can be used across domains. In this paper, we present a framework that makes use of such domain-independent techniques for answering the what-why questions effectively. Our framework forms the backbone of a customer-focused contact center and can be used across domains. In the next section, we provide an overview of our framework for implementing a customer-centric approach at contact centers. The framework has four main stages: customer communications management, insight extraction, service improvement, and measuring impact on business. Each of these stages is explained in its own section later in this paper. We also provide real life experiences gained from implementing this framework at various contact centers. Finally, we conclude the paper.

Copyright 2009 by International Business Machines Corporation. Copying in printed form for private use is permitted without payment of royalty provided that (1) each reproduction is done without alteration and (2) the Journal reference and IBM copyright notice are included on the first page. The title and abstract, but no other portions, of this paper may be copied by any means or distributed royalty free without further permission by computer-based and other information-service systems. Permission to republish any other portion of this paper must be obtained from the Editor. 0018-8646/09/$5.00 ª 2009 IBM

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

M. BHIDE ET AL.

9:1

Measure impact on business • Improvement in customer satisfaction • Improvement in agent productivity • Improvement in conversion/cross-sell/ up-sell rates

Service improvement • Enhance agent skills and efficiency • Implement effective contact patterns

Customer expectation management

Handle customer communications • Interfacing with the customer • Responding and addressing customer queries, requests, and complaints

Business insights extraction • Process voice-of-customer data • Extract reputation for products and services • Discover likes and dislikes related to products and services

Figure 1 Framework for customer-centric service management at contact centers.

specific keywords while talking to the customers or suggestions to cross-sell a product to a specific category of customers. Thus, the customer communications stage is responsible for direct contact with the customer and for the implementation of the suggestions provided by the framework, which ultimately governs the success of the entire customer-focused service management framework. The next stage is the extraction of useful business insights from customer communications. Apart from improving the overall quality of the customer experience through better understanding of customer expectations, analyzing VoC data helps to provide better, timely, and valuable intelligence to other business units within the enterprise. This can include input on customers, products, services, and processes—information that, when captured, identified, assimilated, and turned into usable knowledge, can literally transform the ability of an organization to identify and meet customer expectations. Thus, the business insights extraction stage helps to discover the hidden themes present in the VoC. The business insight extraction process begins with the processing of customer communications obtained from various channels such as e-mails, surveys, and call transcripts. Processing data collected from these channels poses numerous challenges, including:

Overview of service management framework

 Data quality—Data is noisy and contains not only

Figure 1 shows the four main stages of the service management framework, namely:

spelling and grammatical mistakes, but also inconsistent and incomplete sentences. In addition, processing steps such as speech recognition introduce processing noise. Sometimes the content is multilingual where the customers express themselves in two or more languages.  Annotation—Unstructured data must be annotated; that is, the various keywords of interest in the data must be identified. The keywords of interest could include customer name, e-mail ID, phone number, and account number.  Linking—The features extracted and annotated from the unstructured data must be linked with relevant structured elements present in the data warehouse in order to find business-critical associations.

1. 2. 3. 4.

Handle customer communications. Business insights extraction. Service improvement. Measure impact on business.

These stages form a circular workflow where one stage feeds the next. This is a continuous cycle that helps the contact center to continuously optimize and improve customer experience. Customer-centric service management begins at the point of contact with the customer. The contact center agent handles this interaction and addresses a customer’s service- and product-related queries, requests, and complaints. Thus, the main role of this stage is to address the customer’s needs through the contact center agent. Another role played by this stage is that of incorporating the process improvements suggested by the service management framework while communicating with the customer. These process improvements are the end result of using the customer-centric approach in service management; that is, they are generated after the use of all four stages of the framework. These improvements could be in the form of, for example, suggestions to use

9:2

M. BHIDE ET AL.

Once linked, the resulting data offers interesting opportunities to find correlations that would not have been possible by using only structured information. For instance, structured data can help the enterprise detect whether a customer is likely to ‘‘churn’’ (i.e., structured data helps detect what is going to happen). However, the reason why this is going to happen is hidden inside the VoC. Thus, in order to extract full value, the structured and unstructured data needs to be linked in order to provide answers to both the ‘‘what’’ and the ‘‘why’’

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

questions. These answers to the ‘‘why’’ questions, which are hidden in the unstructured text, can be extracted by a variety of unstructured data analysis techniques that make it possible to automatically extract and analyze customer sentiments, complaints, and queries across different channels such as short message service (SMS), emails, call logs, and call transcripts. These insights help in churn prediction [1], campaign management [2], customer satisfaction analysis [3], agent productivity [4], and many other applications. The insights extracted by the business insights extraction process are then used to carry out process transformation for service improvement. Typically, business analysts need to analyze and understand the insights delivered by the business insights extraction process. The analysts are responsible for identifying the correlation of the insights with a set of key performance indicators (KPIs), which could include factors such as customer satisfaction and conversion ratio. Once the correlations have been identified, the analysts suggest changes in the existing processes in order to improve the KPIs. The changes could include change of processes in the enterprise or a change in the training process for the agents. For example, the change could be in the form of a suggestion for using the right words while interacting with the customers. The key business challenge in this stage is to design a transformation process that is feasible, cost effective, and non-disruptive and that incorporates the discovered insights. This needs to be done incrementally, starting first with small test groups, measuring the impact, and then scaling up to the whole team. The next stage in the customer-centric service management framework is that of measuring the impact of the transformation process on a test group. In other words, the analysts need to understand the effect of the transformed process on the KPIs for the test group [4]. These KPIs, which are used for measuring the impact, depend on the process type and business domain. For example, in service calls, improvement in agent performance is measured by corresponding improvement in customer satisfaction scores, whereas for a sales agent, it is an increase in sales revenue. If the process transformation does not produce an expected improvement in the KPIs, then the analysts need to either discard the process transformation or redesign it. On the other hand, if the KPIs do show satisfactory improvement, then the process change is rolled out on a larger scale. If the process change involves retraining of the agents, then it leads to a change in the way the agent interfaces with the customer—which is part of the handling customer communications stage. Thus, a customer-centric approach for service management uses

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

Greeting Pickup and return details Memberships Car options and rates

Customer objection Objection handling

Personal details Mandatory checks

Confirm details

Conclusions

Figure 2 Call flow in a car rental process.

the VoC to identify the hidden insights that in turn are used to improve the services offered by the enterprise. Thus, by focusing on the VoC, a customer-centric approach helps the contact center to address real customer problems, which in turn leads to improved customer satisfaction and better conversion ratio, thereby generating more revenue for the enterprise.

Effective handling of customer communications Customers can communicate with a contact center using either a text-based medium such as e-mail, chat, or SMS or a voice-based medium such as telephone. The customer communication stage has the objective of interfacing the customers with the agents and responding to their product- and service-related queries, complaints, and requests. This stage also has the objective of delivering all of the process improvements suggested by the customer-centric service management framework. In order to effectively carry out their task, agents are trained to use a call flow template [5]. Figure 2 shows the call flow template used by a car rental contact center. For each of these steps, agents are trained on the specific phrases they should use. For example, agents must open the call with a greeting, an introduction and mention of

M. BHIDE ET AL.

9:3

the car rental company’s name. In the mandatory checks part, the agent must find out whether the customer is older than 25 years and has a good driving record and a valid driver’s license. Sometimes agents may be tempted to skip certain sections in order to keep the average call handling time low. However, that may expose the company to various problems including litigation, so during training, agents are specifically trained to follow the script. The purpose of the call flow is threefold. First, it is designed to address the needs of the customer without the agent forgetting anything. Second, it helps the agent to maintain the prescribed average call handling time. Third, it is designed to help close a call successfully. The goal of a car rental contact center agent is to rent out as many cars as possible, so a successful call is one where the customer agrees to rent a car. For an e-mail-based contact center, the call flow is typically captured in the form of canned responses, which are to be sent to the customers. When an agent receives an e-mail, the agent finds the most appropriate canned response for the customer query. The agent then personalizes the canned response and sends it to the customer [6]. As part of the process improvement, the customercentric service management framework might suggest the use of a different set of phrases during the call flow. This is accomplished by changing the call flow template and training the agents for the modified template. In the case of an e-mail-based contact center, a change in the process can be easily achieved by changing the canned response. Thus, the customer communications stage accomplishes the two objectives of responding to customers as well as acting as a delivery channel for the process improvements suggested by the framework.

Business insights extraction The VoC provides a dynamic view of customer needs, problems, opinions, sentiments, inclinations, and propensities that change from time to time. The process of business insights extraction is responsible for discovering this dynamic view, which is hidden inside the transcribed phone calls, e-mail notes, and chats. The insights are typically discovered by finding strong associations within the data. However, before the actual discovery of insights can happen, the data needs to be prepared so that it becomes amenable to association discovery. Hence, the first step in business insights extraction is that of data preparation, as explained in the following subsection. Once the data has been prepared, the actual discovery process starts. The details of the discovery process are explained in the section on extraction of insights.

9:4

M. BHIDE ET AL.

Data preparation Processing of data collected from different channels in a contact center poses numerous challenges. The first key challenge is to improve the data quality, which is generally very poor. Poor data quality leads to difficulty in analyzing the text and makes it difficult to extract insights from the text. On the basis of experiences with real life customer data, we outline simple yet effective techniques to handle noisy data in the section ‘‘Handling noise in customer communications.’’ Once the data is cleaned, it needs to be linked with the structured data present in the enterprise. Linking of the text with the customer profile information present in the data warehouse allows the enterprise to derive useful business insights from the customer communications. Examples of such insights include answers to questions such as, ‘‘What percentage of high-value customers from New York are currently unhappy with the company?’’ Notice that such insights cannot be derived from structured or unstructured data alone. The fact that a customer is unhappy is hidden inside the e-mail note, and the fact that the e-mail note is from a high-value customer is present in the data warehouse. Hence, it is crucial to link the customer communication with the customer profile present in the data warehouse. The first step to finding the link is to capture the customer identification information such as customer ID, credit card number, transaction ID, and shipping ID, which are present in the unstructured text. Due to poor data quality, this can be a very difficult task. We extract these identifiers in the customer communication using annotators as explained in the section on annotating customer data. Once the customer communication has been annotated, the next step is to unambiguously link the text with the customer profile. The details of the linking procedure are given in the section on linking of structured and unstructured data. Handling noise in customer communications Almost all data captured by the contact center is noisy and of poor quality. Agent summaries are noisy because agents want to quickly note the gist of the conversation and proceed to the next call. E-mails written by customers contain not only spelling and grammatical mistakes, but also inconsistent and incomplete sentences. Sometimes the content is multilingual where the customer uses two or more languages; for example, the text is written using the English alphabet, but the word is from a different language. Another form of noise in contact center e-mails is the presence of advertisements, disclaimers, canned greetings, and text of earlier messages repeated as history. This useless additional text makes analysis of the interaction content not only slower, but also less effective since it tends to obscure the actual information contained

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

in the interaction. Also in processing voice data, noise is introduced by an automatic speech recognition system. Techniques for handling poor-quality data are dependent on the type of errors or problems found in the data. These problems that affect the data quality are typically common for a data source; problems found in e-mail data will be common to e-mails in general; similarly, transcribed call records will have specific speech recognition-related issues. Hence, the techniques for handling poor data quality are specific to the data source. We next explain techniques for handling noise in agent notes, e-mail data, and speech transcripts. Customer service agents write down a summary of the call after it ends. Since agents are evaluated by the average call handling time, they speedily make these notes with grammar, spelling, and punctuation being the least of their concerns. This leads to a lot of noise in the transcribed call data. These issues make the task of identifying interesting keywords in the text extremely difficult. One way to address this problem is to focus on words that occur a statistically significant number of times [7]. This helps to eliminate a large number of misspelled words and infrequent proper nouns. Given the absence of structure in the e-mail text, it is hard to devise a perfect procedure for the cleaning task. Simple heuristics can be used to tackle common problems that include removal of stock replies, removal of history text, advertisements, and disclaimers. More complex methods have also been used in other tasks [8]. Automatically transcribed data is noisy and requires a good deal of feature engineering. In addition to recognition noise, spontaneous speech features such as ‘‘um,’’ ‘‘uh,’’ and other fillers need to be removed [9]. We found that the technique of considering only those words that appear a statistically significant number of times combined with the removal of fillers helps to reduce the noise in transcribed data by a significant amount. Annotating customer data In order to make use of the customer communication data, it needs to be linked with the customer profile present in the data warehouse. The first step to identify the link is to annotate the customer identification information such as customer ID, account numbers, and transaction ID present in the communication. This task is far more difficult than merely looking for numeric sequences in the text and then disambiguating these sequences on the basis of the number of digits, prefix sequences, and other patterns. This is due to a variety of reasons [10], some of which are listed below:  The customer and transaction IDs are formatted in a

variety of ways in the e-mail texts. For instance, the bank account and transaction numbers are often

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

stated with hyphens or spaces in between. Hyphens and white space may also appear if the ID is split across two lines in the text.  We might know that the customer ID has six digits, transaction ID has nine digits, and so on. However, sometimes the customer chooses to omit the leading zeroes of the ID (the transaction ID 000321675 appears as 321675); this means that the length of the numeric sequence is not a reasonable hint and it is hard to tell a transaction ID from a customer ID or even a currency value just by looking at the numeric sequence itself. In speech data, some digits may be missed or wrongly recognized.  The first few digits of a numeric sequence can be used as a hint for identifying the type of the number. The first four digits of a credit card number, for instance, are usually unique for a bank and the card type (Visa or MasterCard). The first three digits of a customer identify the branch where the customer first opened an account, and so on. However, these can lead to false positives; the system still cannot distinguish the customer ID 110022 from the postal code 110022. We used unstructured information management architecture (UIMA) annotators [11] to identify customer identification information. At its simplest, an annotator tokenizes the text and applies pattern-based rules on the token sequence obtained to identify the interesting tokens. These rules combine the hints mentioned above (size of the numeric sequence, identifying prefixes) and take the presence of hyphens and white space into account as well. Moreover, they also take hints from the surrounding text to identify the type of the ID identified (e.g., a credit card number could be surrounded by words such as ‘‘visa,’’ ‘‘master-card,’’ and ‘‘expiry’’). Linking of structured and unstructured data The annotation step identifies the set of customeridentifying information mentioned in each customer communication. In the linking step, each annotation is verified by checking whether it corresponds to an entity (of the given type) in the database; if an entity is not found valid, it is discarded. If only one entity remains for the customer communication after the preprocessing, then we do not have a choice and this entity is considered the most relevant. The interesting case occurs when multiple entities remain. A naive procedure would link the customer communication with all the multiple entities present. However, this would not be correct, as typically a customer communication is related only to the problem issues of a single person. We solve this problem by gathering support for each entity mentioned from the remaining information present in the customer

M. BHIDE ET AL.

9:5

communication and eliminating the entities that do not have any support [12]; the details follow. We first build up the context of the given customer communication as the set of valid entities identified as above, along with other information such as the name of the customer obtained from the e-mail header (or the appropriate metadata in case the communication is not an e-mail). We also build up the context of each entity by querying the data warehouse and extracting the name of the customer. The support of an entity in the customer communication is computed as the size of intersection of the context of the entity with the context of the given customer communication. Clearly, the greater the support of an entity, the more relevant it can be assumed to be to the given customer communication [8, 9]. We eliminate the entities with zero support, and among the remaining, we identify those entities with the greatest support as the most relevant to the given customer communication. The discovered links between the customer communication and the customer profiles are populated in a table within the data warehouse. This enables consolidated analysis on both the customer profiles and the customer communication, which can be exploited in a variety of ways. Extraction of insights From a contact center perspective, insights extracted from VoC data can serve two critical purposes. First, the insights related to customer experience, such as customer complaints or difficulties and requests, can be used to better understand and manage customer expectations. Second, the insights that are related to quality issues can be used for rating or improving agent performance. In this section, we provide examples of how insights can be discovered from VoC data in order to improve agent performance and support customer intelligence. Monitoring and improving agent performance In a contact center, how well or badly an agent handles the interaction with a customer is an important factor governing the overall quality of the customer experience. Typically, contact centers employ quality analysts who are responsible for monitoring agent performance. However, due to the large volume of calls serviced daily, it is impossible for quality analysts to analyze individual calls in order to monitor the performance of agents on criteria such as communication skills, problem identification, and problem resolution abilities. Large volumes of data can be handled by analyzing VoC data (transcribed calls between agents and customers) in order to answer some of the key questions faced by quality analysts with the objective of improving agent performance [4, 13]. The reports that are generated at the end of the insight extraction phase can be useful for

9:6

M. BHIDE ET AL.

monitoring the performance of agents or for designing personalized training programs. We evaluated such a technique in a contact center for a car rental company, as illustrated in Figure 3. The calls received by the contact center were recorded and transcribed using an automatic speech recognition system. These calls were then automatically analyzed to find strong associations between events, with a view on improving conversion. Speech recognition systems have an accuracy of approximately 60%. Because of such low accuracy levels, close to half the words were either missing or wrongly recognized. However, it was found that useful insights could still be derived by aggregating the thousands of calls received each day. By analyzing the calls, many key insights were derived. One noticeable feature was that 80% of the customers who began the call by asking for rates (rates customers) did not make a car booking. On the other hand, 80% of the customers who began the call by asking for a reservation (reservation customers) made a booking. Further analysis showed that good agents converted the rates customers to reservation customers by offering discounts. Thus, there was a clear case for a need to improve agent productivity. To make the improvement, it was necessary to train the agents about good practices followed in successful call closure. Several good practices that resulted in a conversion (rates to reservation customers) were searched. Agent utterances that led to conversion were identified and listed [4, 13]. For example, phrases that had a positive influence included mention of special characteristics in vehicles such as brand new vehicle, passenger-side airbag, unlimited mileage, and cruise control. The domain analysts identified that this finding was linked to the conversion KPI. Similar insights can also be derived using agent notes [14]. In order to service calls efficiently and effectively, agents rely on domain-specific knowledge bases. These knowledge bases contain information about common problems, typical customer issues, and their solutions. These knowledge bases help agents to respond quickly and consistently to frequently asked questions and issues. Such models can be created automatically using unsupervised techniques to generate a domain knowledge base automatically from call transcripts [9]. The knowledge base comprises primarily a topic taxonomy in which every node is characterized by topic(s), typical questions-answers (Q&As), typical actions, and call statistics. This knowledge base can help (among other things) to improve the first-call resolution rate, a key parameter in enhancing the customer satisfaction KPI. Customer intelligence Customer intelligence is the process of gathering, analyzing, and exploiting information of a company’s

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

Modified to mention minimum driving age

Call script

Manuals

Changed to link rates directly from database

Procedures

Agent

Structured database Contracts I would like to book a car from New York on…

VoC Terms and conditions

Analysis system

Analyst

Business system

• Time delay in giving rates to customers • Agents not informing customers about minimum driving age

Customer

Figure 3 Contact center process improvement.

customer base. For many enterprises today, contact centers are the only point of contact between them and their customer base. This makes the data generated from contact centers an ideal candidate for gathering customer intelligence. Unfortunately due to the sheer volume involved, contact center representatives can examine only a very small subset of the data to discover insights about the customer, such as reasons for customer dissatisfaction, frequent customer complaints, and problems. One way to avoid this problem is to employ text classification techniques to automate the customer satisfaction analysis task. We have deployed such techniques for many client accounts in large contact centers [3]. This is a first-of-its-kind solution integrating text classification, business intelligence, and interactive document labeling to automate the task of gathering customer intelligence. Another interesting technique related to extracting customer intelligence from unstructured text generated in contact centers involves the analysis of different subsets of documents. The technique uses an algorithm that helps to identify the keywords that are unique to the result of a SQL query [15]. We extended the technique to unstructured text wherein we try to identify the keywords that represent the context of a subset of documents vis-a`vis the entire dataset. This involves finding the keywords that are popular within the subset but are rare outside the subset, that is, in the entire dataset. We use the scoring function of the algorithm to rank the keywords and identify those keywords with a high score; these keywords

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

represent the context of the subset of documents. This technique helps the analysts identify the keywords that are unique to e-mails or calls of a subset of the customers. These keywords might signify the problems or issues that are unique to the subset of customers. Hence, the technique provides a way to identify the problems of, say, high-value customers or customers from a specific region, thereby extracting valuable insights from unstructured text.

Service improvement As mentioned in the introduction, the insights discovered from unstructured text are analyzed by analysts to find their correlation with KPIs. The correlation is then used to suggest changes in the processes of the enterprise. In the following subsection, we present some real life examples of how the insights derived from e-mails were used to improve the processes of one of the largest banks in India. The amount of noise present in email is comparatively less than that present in transcribed phone calls. However, as we show in the subsection ‘‘Service improvement using transcribed calls,’’ in spite of the increased noise content, our techniques allow the contact center to suggest significant service improvements. Service improvement using e-mails We used our customer-centric service management framework at one of the largest banks in India. Our technology helped the bank to integrate the customer e-mails with its data warehouse. Thus, we linked each

M. BHIDE ET AL.

9:7

customer e-mail note with the customer profile present in the data warehouse. Once the e-mails are linked with the data warehouse, it is easy to identify the set of e-mails sent by the platinum customers. Using our subset analysis techniques (mentioned in the section on customer intelligence), the analysts were able to identify the keywords that were unique to the e-mails sent by the platinum customers vis-a`-vis those sent by the rest of the customer base. The keywords returned were ‘‘netbanking,’’ ‘‘overseas,’’ and ‘‘password reset.’’ This pointed to the fact that the platinum customers were facing problems in the password-reset process for their Internetbanking account. An analysis of a few of these e-mails revealed that the platinum customers could not reset their Internet-banking password online. This problem was exacerbated by the fact that the password-reset process of the bank required the customers to visit a branch and fill out a form to reset their password. This was a major problem for the platinum customers who were traveling overseas and could not visit a branch to reset their password. Thus, the customer-centric service management framework helped the analysts discover this crucial piece of insight from the customer e-mails and suggest a change in the password-reset process, which solved a major customer concern. Our framework also helped the analysts to improve the user experience for the platinum customers by suggesting a change in the process of interfacing with the customers. This was enabled as a result of the insights provided by the linking of the customer e-mails with the customer profile present in the data warehouse. Once the unstructured data was linked with the data warehouse, the analysts could use the standard business intelligence tool to gather insights from the structured as well as unstructured data. Business intelligence on this linked information discovered that a significant percentage of the platinum customers were using the e-mail channel to contact the bank. This was surprising because the bank had a team of dedicated relationship managers to cater to the needs of their platinum customers. When a platinum customer used the e-mail channel, the customer was not provided any priority service, which resulted in dissatisfaction among this critical customer base. This insight helped the bank to identify an important loophole in servicing its platinum customers. Using this insight, the analysts suggested a change in the e-mail handling process of the bank. Thus, when linked with the customer profiles, the unstructured data can be a virtual goldmine of information that can be used for a variety of purposes. This is highlighted further by the next example that helped improve the customer profile information present in the data warehouse. The information provided by the bank customers while opening an account or applying for a credit card generally

9:8

M. BHIDE ET AL.

becomes stale with time. This includes the contact information such as phone number, e-mail ID, and address. A major problem faced by the bank was that a large portion of its customer base was deemed noncontactable, meaning that the contact information present in the bank was no longer accurate. Thus, the marketing team of the bank was unable to cater to this large proportion of the bank’s customer base. This resulted in a major loss of revenue for the bank. Business intelligence on the linked information created by our framework enabled the analysts to discover that a significant portion of the customers contacting the bank via e-mail were deemed non-contactable as per bank records. This insight helped the bank to create a new process to extract information such as e-mail ID, phone number, and address from these customer e-mails and update them in the data warehouse. Thus, our framework helped the bank to improve the customer profile information present in its data warehouse, which helped to save revenue for the bank. The crucial fact in all of these examples is that they use the information captured by the contact center and derive valuable insights from it. These insights help the enterprise to provide better service to its customers by addressing the customer problems and concerns and improving business processes. Service improvement using transcribed calls The goal of the service improvement step is to incorporate the findings from the business insights system into the contact center processes. As described in the section on monitoring and improving agent performance, we had deployed our framework at the contact center of a large car rental company. In that deployment, we recorded the customer calls and transcribed them using an automatic speech recognition system. These calls were then automatically analyzed to find business insights with a view to improving conversion. The speech recognition system had an accuracy of approximately 60%, as mentioned earlier. At such low accuracy levels, close to half of the words were missing or wrongly recognized. However, it was found that useful insights could still be derived by aggregating thousands of calls received each day. Once the positive agent practices that resulted in a conversion were identified, the contact center started including these findings in the training program for agents. In this way, positive practices were passed on to all the agents. Specifically, based on the derived insights, agents were deliberately asked to treat customers differently on the basis of whether they were asking for rates or asking for reservations. Further, agents were instructed to include value-selling phrases in their vocabulary. They were also told to use a set of positive phrases that had resulted in

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

bookings. Based on the insights, the call-flow and callscript were modified.

Measuring impact The service improvements suggested by the analysts are first tested on a small set of users. This helps the contact center to validate the usefulness of the suggested changes. The validation of the suggested changes is made by measuring their impact on the selected subset of users. However, measuring the impact on a subset of users introduces its own set of challenges in order to ensure that the impact is measured accurately. These challenges can be mitigated by ensuring that the evaluation process adheres to the following guidelines:  Care should be taken to ensure that the subset of

users is representative of the general population who will finally be targeted.  The baseline (against which the impact will be measured) should be carefully selected to account for seasonal changes. For example, sales during Christmas should not be compared with sales during a lean sales period.  Typically, each process change targets a specific metric. For contact center sales agents, this metric could be the improvement in sales or it could be a reduction in average call handling time. The suggested changes should be such that while improving one metric, other metrics should not be adversely affected. Hence, it is important that the impact be measured on all of the metrics and not just the targeted one. If the process transformation does not produce an expected improvement in the KPIs, then the analysts must either discard the process transformation or redesign it. On the other hand, if the KPIs do show satisfactory improvement, the process change is rolled out on a larger scale. Often, process changes involve considerable cost in terms of retraining of the agents or changes in processes. Retraining of agents leads to a change in the way the agent interfaces with the customer, which is part of the handling customer communications component in the customer-centric service management framework. Based on our experiences with different customers, we now outline the various means of measuring the impact of the process changes. Two types of process changes are suggested by the analysts: 1) changes related to the contact center processes and 2) changes in the processes of the enterprise. In this section, we outline the ways to measure the impact of the changes in the contact center processes and then we provide an overview of the same for changes related to the enterprise processes.

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

Changes in contact center processes We measured the impact due to process improvements suggested by analysts in a contact center for a car rental company. We divided the 90 agents in the car rental reservation center into two groups. One group, consisting of 20 agents, was trained based on the insights derived from the VoC as explained in the section on monitoring and improving agent performance. These 20 agents were told about the findings from the system. They were told that customers can be classified into two types based on the way they began their calls: 1) rates customers and 2) reservation customers. They were also told that rates customers typically did not choose to book a car. The agents were told that offering discounts to rates customers resulted in more bookings from them. They were also asked to use specific phrases more generously in their conversations with the customers. The remaining 70 agents were not told about these findings. By comparing these two groups over a period of two months, we evaluated the effect of the process changes on agent performance. As the evaluation metric, the reservation ratio (the ratio of the number of reservations made to the number of calls resulting in no reservation) was used. Following the training, the reservation ratio of the trained agents was higher by 20% compared to the general population of agents. We also automated the customer satisfaction analysis process and provided the results of the analysis on a dashboard for the contact center management. This allowed the management to quickly focus on the problems and make the necessary changes in the process. It was noted that sometimes agents did not follow mandatory processes. Agent compliance to mandatory processes was measured and fed back. This helped agents to increase compliance by 30%. For contact centers, this is a very important requirement stipulated by law. The dashboard also helped in identifying customer problems quickly and in taking follow-up actions quickly. For example, if many customers were complaining about agents speaking too fast, the contact center relaxed the average call handling time. Similarly, if customers were complaining about the Web site being down, it was quickly fixed so that other customers did not face the same problem. Changes in enterprise processes As mentioned in the section on service improvement, we used the customer-centric service management framework at one of the largest banks in India. One of the findings of our framework was that a large proportion of customers who were deemed non-contactable were using e-mail to communicate with the bank. The analysts suggested the use of information extracted from the e-mail such as e-mail address and phone number in order to improve the

M. BHIDE ET AL.

9:9

Table 1 Impact measurement resulting from process changes in contact center processes. Process

Change

Impact

Car rental

U Call script was changed to incorporate phrases that had positive impact U Customers who started out by finding out rates (‘‘can you tell me the rate for renting a car’’) were treated differently from customers starting out by asking for a reservation (‘‘I would like to rent a car for Sunday’’); e.g., rates customers were offered more discounts or agents created urgency by saying that the deal may not last much longer U Script was changed to include more valueselling phrases (‘‘great brand new car,’’ ‘‘unlimited mileage,’’ ‘‘cruise control,’’ ‘‘passenger-side airbag’’)

U More than 20% increase in rentals U Average call time remained unchanged U Increase in revenue

Agent monitoring

U Individualized score sheets for process compliance, communication skills, and accent

U Compliance on mandatory processes improved by 30% U Personalized training programs for agents to improve CSAT, average handling time, and conversion rates

Customer satisfaction (CSAT)

U Automated the CSAT process U Provided dashboard to view the CSAT analysis to see complaint trends

U Accurate, real-time CSAT analysis U Identification of top issues faced by customers U Speed up in CSAT analysis by at least a week

Improving customer profile

U Improve customer profile of non-contactable customers using customer communication

U Helped reduce non-contactable customers by 23%

Addressing customer pain points

U Enabled Internet-banking password process online U Route mails from platinum customers to their relationship manager

U Improved customer satisfaction for platinum customers

profile information of the customers. This process change resulted in the reduction of non-contactable customers by 23%. The improved customer profiles helped increase the reach of marketing campaigns, thereby providing considerable value to the bank. Another insight offered by the framework was that a significant percentage of the platinum customers were interacting with the bank using e-mail. This led the analysts to suggest a process change wherein the mail sent by platinum customers was automatically routed to the concerned relationship manager. This resulted in personalized service being provided to the platinum customers, thus improving customer satisfaction. The most important process change was that of enabling the reset of the Internet-banking password online without requiring the customer to visit the bank. This was the biggest problem for platinum customers and the changed process acted as a differentiator for the bank, not only to retain existing customers but also to attract new customers. Table 1 summarizes the impact that we

9 : 10

M. BHIDE ET AL.

observed by implementing the process changes in various contact center and enterprise processes.

Summary In this paper, we presented the customer-centric approach to service management. The central concept of our approach is to understand the VoC, which is hidden inside the various forms of customer communications. We outlined various techniques to discover the VoC and use it to increase customer satisfaction levels while maintaining the operational metrics, such as average handling time, at optimal levels. Contact centers today focus only on the operational metrics and fail to understand the needs and problems of customers, which are hidden inside customer communications. Our customer-centric approach is based on understanding and handling hidden customer needs and consists of four stages, which form a circular workflow. The first stage encompasses techniques to capture the customer communications that contain the VoC. The second stage does the important process of data cleaning

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

and linking of the unstructured data with the structured data present in the data warehouse. Using our experience with real customer data, we presented efficient techniques to clean and link information that work very well in practice. We also presented techniques to extract insights from linking e-mails, calls, and customer calls with structured data. We then outlined some examples of how these insights are used to improve the different processes present in the enterprise. Service improvement using the discovered insights forms the third stage of our system. The final stage consists of testing the suggested improvements on a test set. Once validated, the process improvements are then rolled out to the entire customer base. We showed the efficacy of our customer-centric approach by highlighting the benefits accrued by various enterprises following its usage. Thus, our customercentric approach helps a contact center to understand the real issues hidden in the VoC and allows the contact center to use it profitably for improving service levels, addressing customer problems, and achieving high customer satisfaction.

References 1. K. Coussement and D. Van den Poel, ‘‘Integrating the Voice of Customers Through Call Center Emails into a Decision Support System for Churn Prediction,’’ Informat. Management 45, No. 3, 164–174 (2008). 2. M. J. Shaw, C. Subramaniam, G. W. Tan, and M. E. Welge, ‘‘Knowledge Management and Data Mining for Marketing,’’ Decision Support Syst. 31, No. 1, 127–137 (2001). 3. S. Godbole and S. Roy, ‘‘Text to Intelligence: Building and Deploying a Text Mining Solution in the Services Industry for Customer Satisfaction Analysis,’’ IEEE International Conference on Services Computing (SCC), Vol. 2, 2008, pp. 441–448. 4. H. Takeuchi, L. V. Subramaniam, T. Nasukawa, S. Roy, and S. Balakrishnan, ‘‘A Conversation-Mining System for Gathering Insights to Improve Agent Productivity,’’ The IEEE International Conference on E-Commerce Technology and International Conference on Enterprise Computing, E-Commerce and E-Services (CEC-EEE), Tokyo, Japan, 2007, pp. 465–468. 5. K. Kummamuru, P. Deepak, S. Roy, and L. V. Subramaniam, ‘‘Unsupervised Segmentation of Conversational Transcripts,’’ Proceedings of the SIAM Conference on Data Mining (SDM), Atlanta, GA, 2008, pp. 834–845. 6. R. Malik, L. V. Subramaniam, and S. Kaushik, ‘‘Automatically Selecting Answer Templates to Respond to Customer Emails,’’ Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), Hyderabad, India, 2007, pp. 1659–1664. 7. Y. Yang and J. Pedersen, ‘‘A Comparative Study on Feature Selection in Text Categorization,’’ Proceedings of the International Conference on Machine Learning (ICML), Nashville, TN, 1997, pp. 412–420. 8. J. Tang, H. Li, Y. Cao, and Z. Tang, ‘‘Email Data Cleaning,’’ Proceedings of the ACM SIGKDD Conference on Knowledge Discovery in Data Mining (KDD), 2005, pp. 489–498. 9. S. Roy and L. V. Subramaniam, ‘‘Automatic Generation of Domain Models for Call-Centers from Noisy Transcription,’’ Proceedings of the International Conference on Computational Linguistics and Association for Computational Linguistics (COLING-ACL), 2006, pp. 737–744.

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009

10. M. A. Bhide, A. Gupta, R. Gupta, P. Roy, M. Mohania, and Z. Ichhaporia, ‘‘LIPTUS: Associating Structured and Unstructured Information in a Banking Environment,’’ Proceedings of the ACM SIGMOD International Conference on Management of Data, Beijing, China, 2007, pp. 915–924. 11. T. Gotz and O. Suhre, ‘‘Design and Implementation of the UIMA Common Analysis System,’’ IBM Syst. J. 43, No. 3, 476–489 (2004). 12. L. V. Subramaniam, T. A. Faruquie, S. Ikbal, S. Godbole, and M. K. Mohania, ‘‘Business Intelligence from Voice of Customer,’’ Proceedings of the International Conference on Data Engineering (ICDE), 2009, pp. 1391–1402. 13. H. Takeuchi, L. V. Subramaniam, S. Roy, and T. Nasukawa, ‘‘Automatic Identification of Important Segments and Expressions for Mining of Business-Oriented Conversations at Contact Centers,’’ Proceedings of the Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLPCoNLL), Prague, Czech Republic, 2007, pp. 458–467. 14. T. Nasukawa and T. Nagano, ‘‘Text Analysis and Knowledge Mining System,’’ IBM Syst. J. 40, No. 4, 967–984 (2001). 15. P. Roy, M. K. Mohania, B. Bamba, and S. Raman, ‘‘Towards Automatic Association of Relevant Unstructured Content with Structured Query Results,’’ CIKM 2005, Proceedings of the 14th ACM International Conference on Information and Knowledge Management, pp. 405–412.

Received November 21, 2008; accepted for publication June 4, 2009 Manish Bhide IBM Research Division, India Research Laboratory, Plot-4, Block-C, Institutional Area, Vasant Kunj, New Delhi, India 110070 ([email protected]). Mr. Bhide is a Research Staff Member in the Information Management group at the IBM India Research Laboratory. He received a B.Tech. degree in computer science from Visvesvaraya National Institute of Technology (VNIT), Nagpur, in 2000 and an M.Tech. degree also in computer science from India Institute of Technology (IIT), Bombay, in 2002. He is currently pursuing a Ph.D. degree from IIT, Bombay. He joined IBM at the India Research Laboratory in 2002, where he has worked in information integration of XML data. In 2008, he received an IBM Research Division Award for his work on context-oriented information integration. He is author or coauthor of 14 patents and 16 technical papers. Sumit Negi IBM Research Division, India Research Laboratory, Plot-4, Block-C, Institutional Area, Vasant Kunj, New Delhi, India 110070 ([email protected]). Mr. Negi is a Technical Staff Member in the Information Management group at IBM India Research Laboratory. He received a B.Tech. degree in electronics and communication from National Institute of Technology, Durgapur in 2001 and an M.B.A. degree from Faculty of Management Studies, University of Delhi in 2009. He joined the IBM Research Laboratory in 2002. His research interests are in the field of information integration and text mining.

L. Venkata Subramaniam IBM Research Division, India Research Laboratory, Plot-4, Block-C, Institutional Area, Vasant Kunj, New Delhi, India 110070 ([email protected]). Dr. Subramaniam manages the Information Processing and Analytics group at IBM India Research Laboratory, New Delhi. He received a bachelor’s degree in electronics and communication engineering from the Peoples Education Society College of Engineering, Mandya, an M.S. degree in electrical engineering from Washington University, St. Louis, and a Ph.D. degree in electronics from the Indian Institute of Technology, Delhi. His research interests include unstructured information management, statistical natural language processing, noisy text analytics, machine learning, text mining, and the application of these technologies.

M. BHIDE ET AL.

9 : 11

Himanshu Gupta IBM Research Division, India Research Laboratory, Plot-4, Block-C, Institutional Area, Vasant Kunj, New Delhi, India 110070 ([email protected]). Mr. Gupta is a Technical Staff Member in the Information Management group at the IBM India Research Laboratory. He received a B.Tech. degree in computer science from the India Institute of Technology (IIT), Kanpur, in 2004. He is currently pursuing an M.S. degree from IIT, Delhi. He joined IBM at the India Research Laboratory in 2005, where he has worked in context-oriented information integration across structured and unstructured data. In 2008, he received an IBM Research Division Award for his work on context-oriented information integration. He is author or coauthor of four patents and seven technical papers.

9 : 12

M. BHIDE ET AL.

IBM J. RES. & DEV.

VOL. 53 NO. 6 PAPER 9 2009