Knowledge Management in Call Centers: How Routing Rules Influence Expertise and Service Quality Christoph Heitz Institute of Data Analysis and Process Design, Zurich University of Applied Sciences CH-84 Winterthur christoph.heitz@zhaw.ch Geoffrey Ryder, Kevin Ross Baskin School of Engineering, University of California Santa Cruz 56 High Street, Santa Cruz, CA 9564, USA gryder@soe.ucsc.edu, kross@soe.ucsc.edu In a call center, customers are assigned to service agents by routing policies that seek to balance several objectives. Usually, these policies follow myopic rules in order to minimize the waiting time or maximize the quality experienced by the next customer. However, there is a secondary effect of the routing assignment: by learning-on-the-job, the development of the agents expertise depends on the calls they take. In this paper, we address the effect that agent learning has on the service level experienced by customers. A dynamical model of learning and forgetting links the routing policy with knowledge acquisition, treating expertise as an endogenous rather an exogenous variable. Our results indicate that the routing may have a big impact on the knowledge level of a firm, and that classical routing policies may have a negative impact on the distribution of that knowledge.. Introduction A major influence on a customer s satisfaction at a call center is the knowledge level of the agent who takes their call. Knowledge management, in particular maintaining or increasing the cumulative knowledge of the agents, is therefore a key issue for ensuring service quality. This is especially true when the call center operates within dynamic markets, and agents are required to keep pace with trends and advances. For operational management, on the other hand, the knowledge of the employees is usually treated as exogenous to the service delivery process. Knowledge is considered to be a given and fixed resource, and is treated as such for routing and call assignments. This makes sense when all training happens off-line, but does not account for the case where knowledge and expertise are actually gained on-the-job through the service process itself. If
we assume that learning-on-the-job takes place, then the operational rules have an impact on knowledge, and knowledge is therefore an endogenous rather than an exogenous variable. In particular, routing policies determine which agents work on which jobs, and thus may have a major impact on the learning of the agents and their expertise level attained. In our paper, we study how routing might influence the knowledge, how changing knowledge levels will affect customer experience, and how knowledge management and routing can be treated together. We model the expertise of a service agent with simple dynamic equations, reflecting the essential features of gaining expertise through experience (learning) and lowering expertise through absence (forgetting). We find that, in the long run, the expertise level of an agent increases as the arrival rate to this agent increases. That is, a busy agent will maintain a higher level of expertise and therefore give the customers better average service. In a multi-agent environment, the arrival rates to each agent are influenced by the routing policy employed at the call center (Gans et al. 23). Hence different routing policies may lead to different distributions of knowledge/expertise and therefore to different customer quality experiences. We analyze the case of two agents sharing incoming calls, and find that two natural objectives will lead to contrasting policies. One will develop a single expert who handles as many calls as possible, whereas the other will develop two equally trained agents. 2. Dynamics of Agent Expertise We use a straightforward learning-forgetting expression to represent the typical behavior of accepted experience curve models that match a range of empirical human performance measurements. Many such models of varying complexity have been proposed see for example Badiru (992), Shafer et al. (2), Sikström and Jaber (22), and Howick and Eden (27). They are characterized by diminishing returns as the agent approaches the peak of his skill, and forgetting may be modeled as a negative power law or exponential process (Globerson and Levin 987, Nembhard and Osothsilp 2). We assume the quality of the service encounter increases if the worker retains more expertise (Pinker and Shumsky 2, Whitt 26). Consider the evolution of expertise in an agent answering calls to a call center. Let the expertise x(t) of the agent at time t be on a scale x(t), where x(t) = indicates a novice, and x(t) = corresponds to an expert. Define the average time between completed jobs to be T, including the receiving and processing of a job, followed by some time until the next job arrives. The arrival rate of customers to the system is λ =/T. We assume that the agent learns while processing the job (on-the-job), thus increasing its expertise level x(t), and forgets while not processing, leading to a decrease of x(t). In our learning model, the agent s expertise by processing one job increases on the average through x(t) x(t)+α( x(t)) where α is a learning parameter. That is, the experience gain is proportional to ( x(t)), and so becomes geometrically smaller as x(t) approaches expert status. In the absence of 2
forgetting, an agent will move from novice to 63% of the expert level by completing /α jobs. Skills need to be maintained through reinforcement; in the absence of work to occupy an agent, forgetting ultimately reduces the expertise of the agent to zero (novice level). We assume that forgetting occurs at a continuous rate β, so that for a period of length t, the expertise is discounted by e β t. Forgetting occurs only when the agent is idle, so let τ be the expected idle time the time that the agent is not actively helping customers. ( τ = λ ) () µ Learning is designed to be a geometrically decreasing concave function of time, and the forgetting exponential function is convex in time for positive τ. This convex function has been used in previous studies of forgetting see in particular Globerson and Levin (987), and Nembhard and Osothsilp (2) and we adopt it here for its simplicity and tractability. Taking learning events and continuous forgetting together, we get x(t + T )=(x(t)+α( x(t))e βτ (2) Given these simple dynamics, asymptotic behavior of x(t) will tend toward the fixed point x of this equation, with x. The smaller τ becomes (i.e. the more jobs per time unit the agent is handling), the higher the asymptotic expertise level x, and vice versa. Then the steady-state value of expertise becomes: α x = e βτ + α. (3) For an agent who is always busy, τ =, leading to x =. Figure plots the expertise level x for one agent against the arrival rate λ required to achieve it. The service time is fixed at µ =.8. The forgetting rate (β) is.; and medium, fast and slow learning rates (α) are.2,., and.8, respectively. The most important observation of this simple model is that, in the presence of learning and forgetting, there is a relationship between the arrival rate and the asymptotic expertise level of an agent. The more jobs per time unit an agent handles, the higher is his or her resulting expertise. Since routing policies determine how many jobs a specific agent receives, they have an impact on the asymptotic knowledge of the company. 3. Conflicting Goals of the Customer and the Firm The observation of a correlation between arrival rates and expertise leads to the natural question of how a call center should route calls to different agents. A customer will of course prefer to be served by the available agent with the maximum expertise, but this may lead to unintended consequences, as the following example demonstrates. Consider the situation where all incoming jobs are divided between two agents, A and A 2.Takeλ to be the arrival rate of all jobs into the system, and let λ i be the arrival rate of 3
asymptotic expertise achieved, x.8.6.4.2 medium fast slow.2.4.6.8 required arrival rate, λ Figure : The arrival rate λ versus the resulting asymptotic expertise level x for a single agent at different learning rates. jobs routed to agent i, so that λ = λ + λ 2. Parameter p is the fraction of jobs routed to A, and ( p ) is the fraction routed to A 2. We see that the value of p chosen by our decision rule thus determines the two asymptotic expertise levels of the agents and we can introduce the notation x (p ) and x 2 (p ) to denote the dependence of asymptotic expertise on p. In the next section we discuss how one might select the ideal value for p. Customers and firms have different objectives with respect to knowledge of the agents. Customers may prefer to have the maximum available service expertise. Firms on the other hand are interested in the overall knowledge and expertise available within the company. For example, having more than one trained agent mitigates the risk of one agent leaving (and taking their expertise with them). Having agents with similar knowledge level leads to quality assurance whereby each customer receives equivalent service, which might be desirable. Figure 2 illustrates the trade-off between customer and firm perspectives. As a simple example, let the customer s utility be U c (p )=E[x], where E[ ] denotes the expectation value; following the notation used in the figure, this is E[x] =p x (p )+( p )x 2 (p ). Let the firm s utility be U f (p )=x (p )+x 2 (p ), corresponding to the total knowledge of the firm. On the left in Figure 2, the asymptotic expertise attained by each of the two agents under the same learning and forgetting parameters of Figure is shown. The middle plot shows the firm s utility as a function of p, and on the right is the customer s utility. Note that the maximum of the firm s utility U f is a minimum of the customer s utility U c. Further, if the firm chooses solely to increase the utility function for the customer, it destroys its own cumulative expertise. For illustrating these two alternatives, assume that we have two agents with x () =.58, 4
asymptotic expertise achieved, x and x 2.8.6.4.2 x 2 x x 2 (i) medium fast slow.2.4.6.8 proportion of jobs to agent, p x x 2 x firm utility, Uf 2.5.5 (ii).2.4.6.8 p customer utility, Uc 2.5.5 (iii).2.4.6.8 p Figure 2: Plots involving asymptotic expertise levels x i. (i) x i versus the proportion of jobs steered to the first agent, p, (ii) the corresponding firm s utility U f and (iii) the customer s utility U c. and x 2 () =.25 for t=, and consider the medium learning rate. Maximizing the customer s utility is equivalent to routing all jobs to agent (p = ) whose expertise will increase while the expertise of agent 2 will decrease. Asymptotically, agent will have an expertise of x=.82, while the knowledge of agent 2 is zero. In Figure 3 (i), the temporal evolution is shown. In contrast, when choosing p =.5, the company ends up with two equally trained agents. In Figure 3 (ii) we compare how the expertise level of two agents will evolve in an M/M/2 queueing system with an average utilization of 28%. Each policy is work conserving, in that a call will never wait while an agent is free. They differ in that under best quality (BQ), if both agents are available the call is taken by the more proficient agent. This leads to p =.68. In fair sharing (FS), the two agents will either alternate or be randomly assigned such a call with equal probability. As would be expected, BQ leads to one relative expert and one relative novice, while FS leads to two equally proficient agents. The two-agent case thus provides an interesting insight: for maximizing the sum of the knowledge in the firm, it is better to route to the less experienced agents in order to give him or her the possibility to learn. This result is largely independent on the form of the learning-forgetting curves as long as the increase of expertise x by learning-on-the-job is a concave function as a function of x (less increase at higher expertise level), the gain of cumulative knowledge is always larger when routing to the less experienced agent. Thus, under a knowledge management perspective, a balanced routing is always preferable. This may be in contrast to other goals such as waiting time reduction. A second observation can be made: the difference between the customer s utility function for the two policies is marginal (less than % under fast learning), but the difference in the firm s utility function is considerable (up to 5% under fast learning). Classical routing policies, treating knowledge as an exogenous variable and consequently optimizing the customer s utility function, lead to extreme rather than to balanced routing, resulting in a marginal increase to the average service level per customer, but a significant decrease in the company s overall knowledge. 5
(i) (ii).8.8 expertise, x(t).6.4 expertise, x(t).6.4.2 policy p= policy FS.2 policy BQ policy FS 5 5 2 25 3 35 4 time t 5 5 2 25 3 35 4 time t Figure 3: (i) Temporal evolution of the expertise levels x i (t) for the two agent case when always choosing the agent with maximum expertise (solid line) or splitting the jobs evenly between the agents, policy FS (dashed line). (ii) Temporal evolution of policies BQ (solid line) and FS (dashed line). 4. Conclusions and Model Extensions Several managerial insights stand out from this analysis. First, since the difference between extreme and balanced policies in the two-agent case we studied tends to have more effect on the total service provider portfolio than on the average customer experience, it may not be ideal to optimize the customer s utility function. Choosing routing policies that optimize the knowledge acquisition of the agents may have a large impact on the knowledge pool of the firm, while not affecting the customer s average experience significantly. Classical routing rules thus might be suboptimal with respect to service quality under these conditions. Second, we observe that extreme (greedy) policies always lead to more heterogenous expertise levels. Even if the expected value of the experienced expertise increases for the BQ policy, there are more cases where customers experience an agent with low expertise. This might have negative consequences. For example, the reputation of a firm is often more affected by bad service experiences than by good ones. Thus, in the case of an extreme routing policy, the reputation risk is higher. Finally, the risk mitigation of a diverse, trained workforce appears to be of great value, particularly in environments of high turnover. Interestingly, all these considerations favor a balanced routing over an extreme routing. This is in contrast to classical routing policies that treat expertise as an exogenous parameter which, for the studied cases, lead to extreme routing policies. We can conclude that taking into account the learning and forgetting of agents can lead to quite different optimum routing policies. Neglecting the learning-on-the-job effect for routing may have significant negative overall consequences for the firm. Several natural extensions of this phenomena are yet to be explored. For example, in the case of multiple agents and multiple job classes, knowledge management becomes even more challenging. We expect to see similar features of knowledge sharing, but the level of cross training and utilization of individual specialists should be carefully balanced. 6
For another model that is closely related to our subject, we recommend Pinker and Shumsky (2). They also model service quality in call centers as a function of agent expertise. In their case, the mechanism that limits experience acquisition is turnover, not forgetting. Both turnover and forgetting are of interest when analyzing patterns of expertise development; it is interesting that Pinker and Shumsky suggested forgetting as a way to augment their model. In our terms, we can consider an agent leaving as an extreme form of forgetting an agent very occasionally forgets all the way to zero expertise, and these rare events are governed by a separate random process. We are presently investigating the integrated effects of learning, forgetting, and turnover. In this study, we have focused on relatively low utilization levels (28%) combined with learning, which Pinker and Shumsky note leads to a situation where the balance of specialists and generalists becomes particularly important. We anticipate that by extending our analysis to larger numbers of agents and job types, together with more specific knowledge objectives, we will see similar properties arise and be able to ensure the appropriate balance through routing strategies. Acknowledgments As a final note, the authors would like to extend their sincere thanks to the editors and anonymous reviewers for their valuable insights. References Badiru, A. 992. Computational survey of univariate and multivariate learning curve models. IEEE Transactions on Engineering Management 39(2) 76 87. Gans, N., Koole, G., and Mandelbaum, A. 23. Telephone call centers: tutorial, review, and research prospects. Manufacturing and Service Operations Management 5(2) 79 4. Globerson, S., Levin, N. 987. Incorporating forgetting into learning curves. International Journal of Operations and Production Management 7(4) 8 94. Howick, S. and Eden, C. 27. Learning in disrupted projects: on the nature of corporate and personal learning. International Journal of Production Research 45(2) 2775 2797. Nembhard, D.A., and Osothsilp, N. 2. An empirical comparison of forgetting models. IEEE Transactions on Engineering Management 48(3) 283 29. Pinker, E., and Shumsky, R. 2. The efficiency-quality trade-off of cross-trained workers. Manufacturing and Service Operations Management 2() 32 48. Shafer, S., Nembhard, D., and Uzumeri, M. 2. The effects of worker learning, forgetting, and heterogeneity on assembly line productivity. Management Science 47(2) 639 653. Sikström, S., and Jaber, M. 22. The power integration diffusion model for production breaks. Journal of Experimental Psychology: Applied 8(2) 8 26. Whitt, W. 26. The impact of increased employee retention on performance in a customer contact center, Manufacturing and Service Operations Management 8(3) 235 252. 7