A Mathematical Model of Service Failure and Recovery Strategies*

INTRODUCTION

Service failures and recovery strategies have been topics of keen interest to service researchers (e.g., Gronroos, 1988; McCollough, Berry, & Yadav, 2000; Tax, Brown, & Chandrashekaran, 1998) and practitioners (e.g., Brady, 2000; Metz, 2000; Quick, 2000). While researchers have made important contributions to the literature on service recovery, there is a continuing need for additional scholarly research and new insights in this domain, especially in light of the growth in services' share of economies around the world and the increasing role of technology in marketing to and serving customers (Bitner, Brown, & Meuter, 2000; Parasuraman & Grewal, 2000). As we highlight below, several critical knowledge gaps still exist.

First, although previous research has extensively examined customer satisfaction in service-failure and recovery encounters, few studies have explored recovery solutions from a service firm's perspective. In particular, the cost of providing effective recovery has rarely been considered in past research. Likewise, there is a lack of research on the link between the value of a customer to the firm and its recovery decisions. Managerial insights pertaining to issues such as how much service-recovery effort is optimal in a given failure scenario, and whether different recovery strategies should be employed for different customers within the same failure scenario, are scant in the extant service-recovery literature.

Second, researchers have classified failure encounters into outcome failures (e.g., an online bookstore shipping the wrong book to the customer) and process failures (e.g., a night-shift hospital nurse treating emergency patients rudely). Similarly, service-restitution strategies have been clustered into outcome recovery (e.g., a 50% discount on the purchase price) and process recovery (e.g., an extensive apology) strategies. Some studies have compared the relative efficiency of outcome versus process recovery strategies (e.g., Webster & Sundaram, 1998), as well as possible complementary effects of the two types (e.g., Miller, Craighead, & Karwan, 2000; Tax et al., 1998). However, several key questions remain unanswered: Are the two types of recovery strategies substitutable and, if so, under what conditions is the substitution likely to be effective? For example, can a process-recovery effort, by itself, be sufficient to overcome a minor outcome failure? Can an outcome-recovery solution, by itself, compensate for a major process failure? What are the implications of such substitutions for the firm's profitability?

Third, previous research on service recovery has not incorporated important contextual factors, such as individual differences across customers, into analytical models or empirical testing. Although customer differences in loyalty and risk perception toward the firm have been found to influence service evaluation in failure/recovery encounters (Bolton, 1998), there is a dearth of modeling efforts incorporating such differences. For instance, how should a firm structure its recovery strategy if an elaborate expression of apology would mean nothing for an angry customer but would pave the way for a smooth recovery for another? What are the implications for service recovery of differences across customers in terms of the relative importance of service outcome versus process components? Insights pertaining to questions such as these are still pending.

To address the aforementioned knowledge gaps, we develop and evaluate an analytical model for determining the optimal allocation of a firm's service recovery resources. We first propose a general conceptual framework for examining service failure and recovery encounters from both customer's and firm's perspectives. Based on this conceptual framework we develop and analyze a mathematical model that incorporates (a) both outcome and process service dimensions-including the customer's sensitivity to failure and recovery along each dimension, the relative importance of each dimension to the customer, and the firm's cost functions for providing recovery along each dimension; and (b) the firm's dual objective of restoring the customer's value perception following recovery to a target level and doing so at minimum cost. We then derive comparative statics to assess the optimal solution's sensitivity to changes in model parameters, and summarize the comparative statics in the form of research propositions. Finally, we employ numerical examples to examine the features of the model and illustrate ways of extending it to more complex scenarios.

To our knowledge the mathematical model we propose is the first such effort to offer a rigorous approach for minimizing service recovery expenses and optimally allocating the recovery budget. As such, it contributes significantly to the services literature by complementing current approaches, which are predominantly descriptive and survey-based. Our approach also specifically answers a call for incorporating financial concerns into services marketing decisions (Anderson & Mittal, 2000; Rust, Zahorik, & Keiningham, 1995; Zeithaml, 2000). Yet another contribution of the proposed model is its value-oriented focus, in terms of both the value perceived by a customer following recovery and the target value the firm desires the customer to hold after the recovery encounter. Finally, because the model integrates customer characteristics (such as risk-aversion level and prior perceived value provided by the service firm) and the relative importance of process versus outcome dimensions into the firm's decisions pertaining to recovery strategies, it offers managerial insights for crafting customized recovery strategies for specific service failure scenarios.

BACKGROUND LITERATURE

Classification of Service Failures and Recovery Strategies

Researchers have proposed a variety of detailed typologies of service failures or dissatisfying incidences (e.g., Bitner, Booms, & Mohr, 1994; Bitner, Booms, & Tetreault, 1990; Kelley, Hoffman, & Davis, 1993). Underlying all those typologies is a more general and parsimonious outcome-process classification (Gronroos, 1988; Parasuraman, Berry, & Zeithaml, 1991), wherein an outcome failure is a core-service failure (e.g., an airline losing a passenger's luggage) and a process failure refers to inconvenience or unpleasantness experienced during service delivery (e.g., a flight attendant being rude to a passenger). From a resource-exchange-theory perspective, outcome and process failures pertain to problems that customers experience with utilitarian and symbolic exchanges, respectively (Smith, Bolton, & Wagner, 1999). A firm's recovery strategies can correspondingly be classified broadly into the same utilitarian (outcome) and symbolic (process) categories.

Link between Failure Types and Recovery Strategies

Findings from service-recovery studies suggest that the most appropriate approach for addressing service deficiencies is to provide monetary compensation in the case of outcome failures and to offer apologies and speedy responses in the case of process failures (Smith et al., 1999; Tax et al., 1998). This stream of research, invoking social equity, justice, and resource-exchange theories, posits that customers (a) place greater value on the exchange of similar resources than on the exchange of dissimilar resources, and (b) categorize economic loss (outcome failure) and social/ psychological loss (process failure) into different mental accounts (Brinberg & Castel, 1982; Smith et al., 1999). These postulations imply that "matching" recovery strategies are most cost-efficient for a firm. The literature, however, does not address questions such as, "Is there a higher level of mental accounting that combines both economic and social loss/gain for an overall service failure and recovery event?" and "If so, are outcome and process recovery complementary or substitutable in the customer's overall value accounting?"

Customer Evaluation of Service Failure/Recovery Encounters

Studies grounded in belief-updating theory (Hogarth & Einhorn, 1992) and examining longitudinal relationships suggest that customers update their beliefs about the future value of a relationship by digesting new, sequential information from service deviations (i.e., service failures) and recoveries (Bolton, 1998; Slovic & Lichenstein, 1971). Similarly, taking the perspective of value accounting resulting from a service experience, studies invoking prospect theory and equity theory conceptualize service failure/recovery encounters as value exchanges in which the failure causes the customer to experience a value loss and the firm's recovery effort contributes to a value gain (Kahneman & Tversky, 1979). Value losses or gains could occur on both utilitarian (e.g., outcome-related money, goods, or time) and symbolic (e.g., process-related status, esteem, or empathy) dimensions (Smith et al., 1999).

Financial Aspects of Service Recovery

There is anecdotal evidence of the importance and financial benefits of effective service recovery management. For instance, data from leading service companies such as Xerox and Federal Express indicate that the costs of recovering from poor service can account for as much as 25-30 percent of sales revenue, and that typically around 70 percent of the recovery efforts are misdirected due to missing or exceeding customer service requirements during the recovery (Carr, 1992). However, despite the detrimental financial consequences of suboptimal recovery decisions, and despite calls for greater rigor and financial accountability in making service-investment decisions (Filiatrault, Harvey, & Chebat, 1996; Rust et al., 1995), scholarly research focusing on the financial aspects of service recovery is virtually nonexistent. Much of the service recovery research to date has focused solely on customer-perceived value or satisfaction; few studies have taken the perspectives of both customers and service firms and no study has formally incorporated the costs of providing service recovery. In particular, there is a dearth of research on the impact of a firm's cost structures-that is, the response functions linking recovery expenditure levels with corresponding customer-perceived benefit levels-on the firm's recovery decisions.

In summary, the extant literature offers some useful insights along four broad themes: classification of service failures/recovery strategies, link between failure types and recovery strategies, customer evaluation of service failure/recovery encounters, and financial aspects of service recovery. However, significant research and modeling gaps remain. Building on the contributions of pervious research and focusing on the unanswered questions, we next develop a conceptual framework and a mathematical model of service failure and recovery strategies.

MODEL DEVELOPMENT

Conceptual Framework

Figure 1 contains our proposed conceptual framework incorporating both the customer's and service firm's perspectives. On the customer side, we divide service failure into two components: failure type (i.e., outcome vs. process) and failure magnitude. When a failure occurs, the customer experiences a value loss, with the perceived value loss being moderated by the customer's sensitivity to each type of failure and perceived relative importance of the outcome and process dimensions. Depending on the type and magnitude of the firm's service recovery effort, the customer experiences a value gain, with the perceived value gain also being moderated by the customer's sensitivity to each type of recovery and perceived importance of the two dimensions. The customer's overall mental accounting of perceived value of the service is collectively determined by his/her previous perceived value, perceived value loss from failure, and perceived value gain from the recovery.

On the service firm's side, the recovery efforts have a two-fold objective. First, the firm aims to restore the customer's cumulative perceived value to a desired target level that is necessary for retaining the customer (this level is the "Value Recovery Target" in Figure 1). Second, the firm attempts to minimize the overall recovery expenditure needed to realize the value recovery target. The overall recovery expenditure is the sum of the expenditures on outcome and process recoveries, which, in turn, depend on the firm's cost functions associated with the two types of recovery.

We propose that the type and magnitude of service recovery do not depend solely on the severity of the failure and the principle of matching mental accounting. Instead, our model conceptualizes a value-driven approach. That is, to determine the optimal recovery strategy, the service firm first needs to decide on a target for its value recovery efforts (i.e., how much value the firm expects the customer to perceive after recovery). This target value should be based on the customer's current and potential profitability as well as other criteria that might influence the customer's importance to the firm. Zeithaml, Rust, and Lemon (2001), for instance, suggest that the most valuable customer group is the segment that spends more with the firm over time, costs less to maintain, and spreads positive word of mouth, while the least valuable customer group is the segment that costs the firm more time, effort, and money to retain and yet does not provide the return the firm expects. A detailed exploration of specific drivers of the targeted recovery level is beyond the scope of this article. As such, while we discuss them briefly later, they are not explicitly incorporated into the framework. The value recovery target and the firm's cost functions for outcome and process recovery jointly determine the type and magnitude of the optimal recovery strategy. We next develop mathematical representations of the various components of our framework and state the assumptions underlying each.

Model Elements, Assumptions, and Notation

Nature of service failure

The magnitudes of the outcome and process failures are denoted as F^sub o^ and F^sub p^, respectively. We assume that the magnitudes of both types of failure are continuous and that F^sub o^ > or = 0 and F^sub p^ > or = 0. In a service failure scenario, either one or both types of failures can occur.

Relative importance of outcome and process dimensions

We denote the relative importance of the outcome dimension as [theta], with 0 < [theta] < 1 (although theoretically and conceptually 0 < or = [theta] < or = 1, assuming a value of 0 or 1 for [theta] is not feasible when deriving comparative statics due to division by zero; however, this assumption does not affect the model conclusions substantively). Correspondingly, the weight for the process dimension is 1 - [theta]. Consistent with findings from past studies (e.g., Goodwin & Ross, 1992; McCollough et al., 2000), we propose that [theta] is mainly influenced by across-industry differences. Situational factors, such as the urgency of customer demand for a service, could also influence the relative importance of the outcome and process dimensions (Webster & Sundaram, 1998).

Customer risk profile and sensitivity to recovery

A customer's evaluation of a service failure and the recovery encounter is influenced by the level of his/her risk aversion. We represent this risk aversion in our model by a set of parameters: [delta]^sub o^ for sensitivity to loss from outcome failure, [delta]^sub p^ for sensitivity to loss from process failure, [gamma]^sub o^ for sensitivity to gain from outcome recovery strategy, and [gamma]^sub p^ for sensitivity to gain from process recovery strategy. All four parameters are assumed to be positive; that is, [delta]^sub o^ > 0, [gamma]^sub o^ > 0, [delta]^sub p^ > 0, and [gamma]^sub p^ > 0. According to prospect theory, most customers are more sensitive to loss caused by service failure than to gains from recovery (Kahneman & Tversky, 1979; Thaler, 1985). That is, in most cases, [delta]^sub o^ > [gamma]^sub o^ and [delta]^sub p^ > [gamma]^sub p^. But our model accommodates situations wherein customers take higher-than-normal risks ([delta]^sub o^ < [gamma]^sub o^ and/or [delta]^sub p^ <[gamma]^sub p^). This phenomenon is true, for instance, for technology-savvy consumers who voluntarily take risks on new innovations and derive satisfaction from the process of trial or retrial, irrespective of their own failure (and recovery) or that of the service firm (Greco & Fields, 1991; Novak, Hoffman, & Yung, 2000).

Customer's prior perceived value

We denote a customer's prior perceived value of the service as B. When B > 0, the customer has a net positive image of the firm based on experiences before the current failure. When B < 0, the customer has already experienced some value loss from the firm and has a net negative image. While some studies suggest that loyal customers are more dissatisfied by failures, others suggest that customers who have higher previous perceived value are more tolerant of failures and more receptive to recovery. Given these equivocal results, interactions between customers' prior perceived value and their sensitivity to service failure and recovery are not explicitly incorporated in our model.

Value recovery target

The service firm's goal is to restore the customer's overall perceived value of the service to a certain target level, k, which represents the cumulative value the firm desires the customer to have after experiencing the service recovery. We assume that the customer's mental accounting occurs across the entire failure/recovery process; therefore the customer's prior perceived service value, B, influences the perceived service value after the failure/recovery experience. A value of k < B implies that the service firm does not consider the customer to be very important and decides to recover for the customer only part of the value loss resulting from a service failure. Likewise k > B implies that the firm considers the customer to be important enough to recover more value for the customer than what was lost during the service failure. The level of k, in addition to reflecting customer characteristics such as customer lifetime value (Blattberg & Deighton, 1996; Brady, 2000), can also be influenced by level of competition, industrial practices such as compensation given for a flight delay or cancellation (Iacobucci, 1998) and guarantees offered regarding certain failure scenarios (Hart, Schlesinger, & Maher, 1992).

Cost structure and recovery strategy

The cost structure for outcome and process recovery determines the efficiency of converting resource inputs into recovery performance (i.e., the magnitude of outcome and process recovery achieved). Our model includes separate cost functions for the two types of recovery and depicts the outcome recovery magnitude [lambda]^sub o^(X^sub o^) as a function of outcome recovery expenditure X^sub o^, and the process recovery magnitude [lambda]^sub p^(X^sub p^) as a function of the process recovery expenditure X^sub p^. Although other resources, such as a more customer-oriented organizational culture, could indirectly enhance recovery efficiency, our model's focus is on monetary recovery investment, which has a direct impact on the managerial objective of recovery-cost control. We assume [lambda]^sub o^(X^sub o^) and [lambda]^sub p^(X^sub p^) are nonlinear functions with the following properties. First, both outcome and process recovery expenditures are positively related to the recovery magnitudes (i.e., [lambda]'^sub o^(C^sub o^) > 0 and [lambda]'^sub p^ (X^sub p^) > 0). Second, when recovery expense increases, its marginal unit contribution to recovery magnitude diminishes (i.e., [lambda]''^sub o^(X^sub o^) < 0 and [lambda]''^sub p^(X^sub p^) < 0). The posited nonlinear relationship between recovery expenditure and recovery performance is consistent with previous research findings regarding the diminishing effect of financial investment on perceived service-quality improvement (Rust et al., 1995) and the nonlinear value function of recovery performance evaluation (Smith et al., 1999). Third, the relative degree of the diminishing slope for outcome recovery and process recovery could vary across different service industries, firms, and seasons.

The Model

In equation (2), -[delta]^sub o^[theta]F^sub o^ refers to the perceived value loss from outcome failure and -[delta]^sub p^(1 - [theta])F^sub p^ refers to the value loss from process failure. [gamma]^sub o^[theta][lambda]^sub o^)(X^sub o^) refers to the perceived value gain from outcome recovery and [gamma]^sub p^ (1 - [theta]) [lambda]^sub p^ (X^sub p^) refers to the value gain from process recovery. The cumulative perceived value after recovery, k, is composed of the elements on the left side of equation (2)-prior perceived value, perceived value loss, and perceived value gain. From the decision-making perspective, this cumulative perceived value after recovery is set to be the service firm's recovery target for the failure situation. We note that our modeling effort is the first attempt to formally integrate recovery cost control and customer value perception of the recovery in a single conceptual framework. The general model we propose can serve as a basis for future model extensions. All parameters are assumed to be independent of one another.

The optimal solutions for expenditures on outcome recovery X*^sub o^ and process recovery X*^sub p^ (and hence total recovery expenditure X*) can be obtained by simultaneously solving equations (6) and (9). However, satisfying the first-order conditions of the Lagrangian function does not guarantee an optimal solution to the original problem as it depends on the structure of the functions involved; moreover, by using the method of Lagrange multipliers, we assume that the values of X^sub o^ and X^sub p^ returned by the method will not be negative. These theoretical assumptions are consistent with existing modeling approaches and can also be verified in actual implementation through graphical representation as we illustrate later. We also note that closed-form solutions may exist only for simple functional forms; to determine the optimal solution for more complicated forms, an evaluation of all possible solutions will be necessary.

The optimal solutions can be converted to managerial decision choices concerning outcome versus process recovery according to the following general guidelines. When X*^sub o^ [much greater than] X*^sub p^, a strategy emphasizing outcome recovery is preferred; when X*^sub o^ much less than X*^sub p^, a strategy emphasizing process recovery is preferred; and when X*^sub o^ [asymptotically =] X*^sub p^, a mixed strategy involving both types of recovery is preferred. Thus, the optimal solutions provide insights for designing the most appropriate recovery strategy-that is, deciding on the types of recovery strategy to emphasize and the amounts of resources to allocate to each type. Converting the allocated resources into specific executional elements (different types of compensation and redress, different employee training programs, etc.) would require superior operational management, the details of which are beyond the focus and scope of this article.

Comparative Statics

Comparative statics facilitate examining how changes in model parameters (such as [theta]) affect the optimal solutions (i.e., X*^sub o^ and X*^sub p^). For our general model, comparative statics are calculated by differentiating the FOC equations (6) and (9) with respect to each parameter and simultaneously solving the two resulting equations (e.g., we get [partial differential]X*^sub o^/[partial differential][theta] and [partial differential]X*^sub p^/[partial differential][theta] as the comparative statics with respect to [theta]). Because optimal total recovery expense X* is the sum of the optimal outcome recovery expense X* and the optimal process recovery expense X*^sub p^, the comparative static result of X* with respect to each parameter is the sum of the comparative statics of X*^sub o^ and X*^sub p^ with respect to the same parameter (e.g., [partial differential]X*/[partial differential][theta] = [partial differential]*^sub o^/[partial differential][theta] + [partial differential]X*^sub p^/[partial differential][theta]), In the interest of space, detailed derivations are not provided here. Detailed expressions for 27 comparative statics results are listed in the Appendix and directional results are summarized in Table 1.

Evaluating comparative statics is important for two reasons. First, they offer a calibration mechanism for deciding on appropriate recovery strategies. Most current service managers rely on their gut feelings to gauge appropriate solutions to the dynamics of a service failure. By following insights from the general model, managers can make more accurate resource allocations based on their understanding of the relevant parameters. Second, and perhaps more important, because the comparative statics are derived from the theoretical model, the results can be considered propositions that can be empirically tested. We now turn to such propositions.

DEVELOPMENT OF PROPOSITIONS

Insights from the comparative statics in the Appendix can be synthesized into a propositional inventory for highlighting and clarifying the relationships between the optimal recovery costs and their critical influencing factors.

Proposition 1a: As the magnitude of a certain type of failure increases, not only does the optimal expenditure on the same (or matching) type of recovery strategy increase, but the optimal expenditure on the other (or mismatching) type of recovery strategy also increases.

Proposition 1b: Overall, as the magnitude of a failure increases, the optimal total expenditure on recovery increases, with the magnitude change determined by customer's sensitivity to value loss, recovery sensitivity, and the slope of the firm's cost function for the matching type of recovery.

This proposition is based on the first six comparative statics results shown in the Appendix. It indicates that the more severe a failure, the greater the resources needed to be invested in both outcome and process strategies. That is, as the magnitude of an outcome failure increases, the optimal expenditure on process recovery also increases (see comparative statics results 2 and 4). Moreover, the efficient correction of the failure may require the firm to expend more resources on the mismatching dimension than the matching one, depending on the cost structures of the recovery strategies. The logic is that when a customer's total mental accounting can be changed by either type of value update (i.e., outcome or process type), the firm should use more the type of recovery that will contribute more efficiently to the value correction, given a fixed recovery budget. For instance, the results indicate that when the magnitude of an outcome failure increases, the ratio of the expenditure increases on the outcome dimension to that on the process dimension (i.e., ([partial differential]X^sup *^^sub o^/[partial differential]F^sub o^/([partial differential]X*^sub p^/[partial differential]F^sub o^)) numerically equals [[lambda]'^sub o^(X^sub o^)X''^sub p^(X^sub p^)]/[[lambda]''^sub o^(X^sub o^)[lambda]'^sub p^(X^sub p^)]. Compared to a slight delay in responding to a customer's request (i.e., a small process failure), rude behavior from a service employee (i.e., a bigger process failure) normally raises more attention and anger from the customer, which often demands discounts or coupons to fully resolve the customer's complaint in addition to a more elaborate expression of regret. In such a situation, compensation can be more efficient to demonstrate recovery intention.

The increase of the optimal overall recovery expense, however, is only influenced by factors related to the matching service dimension ([partial differential]X/[partial differential]F^sub o^ = [delta]^sub o^/([gamma]^sub o^[lambda]'^sub o^(X^sub o^); [partial differential]X/[partial differential]F^sub p^ = [delta]^sub p^/([gamma]^sub p^[lambda]'^sub p^ (X^sub p^)). In particular, for a more severe outcome failure, the increase of the overall recovery cost is positively related to a customer's sensitivity to the outcome failure, and negatively related the customer's sensitivity to outcome recovery and the slope of the cost function for outcome recovery.

Proposition 2a: As the customer's sensitivity to value loss from one type of failure increases, the optimal expenditure on both types of recovery increases.

Proposition 2b: As the customer's sensitivity to value loss from one type of failure increases, the optimal total expenditure on recovery increases, with the magnitude change determined by the magnitude of the failure, the customer's recovery sensitivity, and the slope of the firm's cost function for the matching type of recovery.

The reasoning behind this proposition is that a customer's sensitivity to value loss can amplify the effect of failure on recovery expenditure. Expenditures on recovery strategies that match the type of service failure are positively related to the customer's sensitivity to the value loss for that type of failure. For instance, airlines find that business travelers are usually more sensitive to flight delays or cancellations than are nonbusiness flyers. For a given flight cancellation, significantly more resources might be needed to compensate business travelers compared to others. Similar logic can be used for process failure situations (comparative statics results 7 and 11).

Meanwhile, the optimal expenditure on the mismatching type of recovery also increases, if the customer's sensitivity to the value loss for a failure increases. This seemingly counterintuitive result (compared to the matching recovery strategies described in the literature) actually captures the complementary relationship between outcome recovery and process recovery. For instance, if a wedding catering service prepared less food than reserved (an outcome failure) and the wedding party is very sensitive to the failure, the failure can be corrected not only by a large discount, but also by a quick make-up plan and a sincere apology. Here, outcome and process recovery complement each other to produce an optimal solution.

Taken together, when the customer's sensitivity to the value loss from a failure increases, the optimal total recovery expenditure increases. The amount of cost increase is only influenced by factors related to the matching service dimension ([partial differential]X/[partial differential][delta]^sub o^ = F^sub o^/([gamma]^sub o^[lambda]'^sub o^(X^sub o^); [partial differential]X/[partial differential][delta]^sub p^ = F^sub p^/([gamma]p[lambda]'^sub p^(X^sub p^)). In the example above, when the wedding party's sensitivity with the value loss in the outcome failure increases, the overall cost increase is positively related to the outcome failure magnitude, and negatively related to the customer's sensitivity to outcome recovery and the slope of the cost function for outcome recovery.

Proposition 3a: As customers' sensitivity to value gain from one type of recovery strategy increases, the optimal expenditure on the same (matching) type of recovery strategy does not necessarily decrease; however, the optimal expenditure on the other (mismatching) type of recovery strategy decreases.

Proposition 3b: As customers' sensitivity to value gain from one type of recovery strategy increases, the optimal total expenditure on recovery decreases, with the magnitude change determined by the customer's sensitivity to the matching type of recovery and the firm's cost function for the matching type of recovery.

Based on the comparative statics analysis (see results 13 and 17 in the Appendix), the change in the optimal expenditure on a recovery strategy has a nonmonotonical relationship with the change in the customer's sensitivity to the value gain from the same type of recovery. This finding differs from the prediction based on "in kind" mental accounting assumptions in that the latter infers that less "in kind" resource is needed to restore a failure when it gets more efficient. When taking into consideration the possible substitution effects across different recovery strategies, the service firm makes recovery decisions based on the trade-off between two opposite drivers. Taking the outcome dimension as an example, the change of the optimal expenditure on outcome recovery is determined by two elements. On the one hand, the need for outcome recovery expenditure decreases because the customer perceives more value from each unit of the recovery performance. On the other hand, the motivation to substitute process recovery with outcome recovery increases because, from an economic standpoint, outcome recovery expenditures can generate more value gain for that specific customer. As a result, the optimal expenditure on a recovery strategy is nonmonotonically related to the customer's sensitivity to value gain from that recovery strategy.

Interestingly, another result of the substitutability of recovery effect is that less resource is needed for service recovery strategies that are not "in kind." As the customer's sensitivity to the value gain from the process recovery increases, the optimal expense on complementary outcome recovery decreases (see comparative statics results 14 and 16 in the Appendix). This pair of relationships has implications for refund management. For example, when a customer is a low-income shopper or cares more about the economic utility of a service, (s)he can tolerate a long waiting time for a refund/compensation as long as (s)he is sure the refund will be delivered.

Overall, the optimal total expenditure on recovery always decreases when a customer's sensitivity to the value gain from recovery increases (results 15 and 18 in the Appendix). Specifically, the amount of cost increase is only influenced by factors related to the matching service dimension ([partial differential]X/[partial differential][gamma]^sub o^ = [lambda]^sub o^(X^sub o^)/([gamma]^sub o^[lambda]'^sub o^(X^sub o^); [partial differential]X/[partial differential][gamma]^sub p^ = - [lambda]^sub p^(X^sub p^)/([gamma]^sub p^[lambda]'^sub p^(X^sub p^)). When the customer becomes more sensitive to outcome recovery, the overall recovery cost change is negatively related to the current level of sensitivity to outcome recovery, the return of recovery investment on outcome recovery performance, and the slope of the cost function for the outcome recovery.

Proposition 4a: For a customer with higher prior perceived value of the service firm, the optimal expenditures on both outcome and process recoveries are lower, given a certain recovery target.

Proposition 4b: Optimal total expenditure on recovery is lower for a customer with higher prior perceived value of the service firm. The magnitude of change is determined by customer's sensitivity to a recovery strategy, relative importance of that service dimension, and the slope of the firm's cost function for that recovery strategy.

This proposition demonstrates that given a certain value recovery target (k), recovery strategies could be customized based on each customer's prior perceived value of the same firm (see comparative statics results 22 and 23 in the Appendix). Because our model considers a customer's prior perceived value as the starting point of the customer's mental accounting, the higher the accumulated value, the easier is the firm's task of recovering the customer's value perception to at least a neutral level. Customers who usually perceive higher value of the service are more likely to consider the failure as an accident, be less dissatisfied with a failure, and more cooperative in the service-recovery stage. This proposition offers face validity to the theoretical model. In addition, the proposition is consistent with the results from empirical studies showing that customers with positive experience with the service firm are likely to have a higher tolerance for value loss in failures and are also able to perceive higher value from the firm's recovery efforts (e.g., Bolton, 1998).

When a customer has a higher previous value perception about the service, less resource is needed to restore the value perception to a certain desired level. For a customer who has experienced value loss before, a second failure demands much greater efforts to return the cumulative value perception to the targeted level (Johnston & Fern, 1999). Numerically, given an outcome (or process) failure scenario, the marginal cost change, [partial differential]X/[partial differential]B = -1/([gamma]^sub o^[theta][lambda]'^sub o^(X^sub o^)) [or [partial differential]X/[partial differential]B = - 1/([gamma]^sub p^(1 - [theta],)[lambda]'^sub p^(X^sub p^))], is positively related to the customer's outcome (or process) recovery sensitivity, relative importance of outcome versus process service dimensions, and the slope of the firm's cost function for outcome (or process) recovery.

Proposition 5a: As the service firm's recovery target (for a customer's cumulative perceived value after recovery) increases, the optimal expenditures on both outcome and process recovery increase.

Proposition 5b: As the service firm's recovery target increases, the optimal total expenditure on recovery increases, with the magnitude change determined by the customer's sensitivity to a given recovery strategy, the relative importance of the corresponding recovery dimension, and the slope of the firm's cost function for that recovery strategy.

When the recovery target k increases, the firm needs to expend more on recovery. The effects implied by Proposition 5(a and b) are similar to those corresponding to changes in the customer's prior perceived value B (Proposition 4). Numerically, given an outcome (or process) failure scenario, the marginal cost change, [partial differential]X/[partial differential]k = 1/([gamma]^sub o^[theta][lambda]'^sub o^(X^sub o^)) [or [partial differential]X/[partial differential] k = 1/([gamma]^sub p^(1- [theta])[lambda]'^sub p^(X^sub p^))], is negatively related to the customer's outcome (or process) recovery sensitivity, relative importance of outcome versus process service dimensions, and the slope of the firm's cost function for outcome (or process) recovery. The marginal cost change is not related to the failure magnitude and customer sensitivity to failure.

NUMERICAL ILLUSTRATIONS

In this section we examine through numerical examples the effects of (1) across-firm differences in cost structures, (2) across-industry differences in weights for the outcome and process dimensions, and (3) across-customer differences in substitutability between the two dimensions. In the examples we use hypothetical model-parameter values that are consistent with the model's assumptions. We also illustrate how the proposed model can be extended to incorporate more complex scenarios.

Impact of Across-Firm Differences in Cost Structures

The influence of a firm's cost structure can be examined by comparing two service firms-say, A and B-with contrasting recovery-cost functions. To satisfy the assumptions of the positive but diminishing effect of recovery expenditure on perceived recovery performance in the general model, we set the cost structures to be logarithmic functions. Specifically, for Firm A we operationalize perceived outcome recovery magnitude as [lambda]^sub o^(X^sub o^) = 50 * L^sub n^(0.03X^sub o^ + 1) and perceived process recovery magnitude as [lambda]^sub p^(X^sub p^) = 12 * L^sub n^(0.25X^sub p^ + 1), with [lambda]'^sub o^(X^sub o^) > 0, [lambda]''^sub o^(X^sub o^) < 0, [lambda]'^sub p^(X^sub p^) > 0, [lambda]''^sub p^(X^sub p^) < 0. Firm A is typical of service businesses that have lower basic process-recovery costs than basic outcome-recovery costs (as implied by the higher coefficient for X^sub p^ than for X^sub o^), but whose process-recovery expenditures will increase dramatically if they want to further impress their customers with higher-than-normal recovery speed or superior complaint handling processes. Conversely, Firm A's basic outcome recovery is comparatively less cost efficient but the marginal effects of higher levels of outcome recovery will diminish slower than for process recovery. Various e-service businesses share this cost structure. Firm B's cost functions are the reverse of Firm A's cost functions. Specifically, [lambda]^sub o^(X^sub o^) = 12 * L^sub n^(0.25X^sub o^ + 1) and [lambda]^sub p^(X^sub p^) = 50 * L^sub n^(0.03X^sub p^ + 1). Airlines and banks, for instance, have cost functions similar to those of Firm B.

Scenario 1

Consider a service for which customers place equal importance on the outcome and process (i.e., [theta] = 0.5). A customer with no prior experience with a firm (B = 0) experiences a process failure of magnitude 30 (F^sub o^ = 0; F^sub p^ = 30). This customer's sensitivities to value loss or gain for outcome and process dimensions are all set at 1.0 ([delta]^sub o^ = 1; [delta]^sub p^ = 1; [gamma]^sub o^ = 1; [gamma]^sub p^ = 1). The firm decides to recover the exact amount of value loss due to the failure; that is, the "value recovery target" is zero (k = 0).

In the above scenario, Firm A should use a "mismatching" recovery strategy emphasizing outcome recovery because its cost structure makes outcome recovery methods more efficient. Under such a strategy the optimal overall cost to achieve the recovery target turns out to be $21.31, with $13.96 for outcome recovery expense and $7.35 for process recovery expense. For Firm B under the same scenario, a matching strategy is optimal. Although the total cost is the same, it should invest more on process recovery ($13.96) and less on outcome recovery ($7.35). Thus, within the same context, cost structure differences across firms have a significant impact on the optimal recovery strategy. The top two panels of Figure 2 show differences in the optimal recovery strategy for Firms A and B for various magnitudes of process failure.

Impact of Across-Industry Differences in Service-Dimension Weights

The impact of across-industry differences can be examined by considering two industries-say I and II-that differ in terms of the relative importance of the outcome and process dimensions. In Industry I the outcome dimension is more important (e.g., medical services) while in Industry II the process dimension is more important (e.g., tour-guide services). For illustration, we set [theta] = 0.7 in Industry I and [theta] = 0.2 in Industry II.

Scenario 2

Consider a new customer with the same risk profile as in Scenario 1 who experiences a service problem involving outcome and process failures, both with severity level of 30. The service firm in question has the same cost structure as that of Firm A and decides to recover the exact amount of value loss due to the failure. Numerically, we set B = 0; k = 0; F^sub o^ = 30; F^sub p^ = 30; [delta]^sub o^ = 1; [delta]^sub p^ = 1; [gamma]^sub o^ = 1; [gamma]^sub p^ = 1; [lambda]^sub o^(X^sub o^) = 50 * L^sub n^(0.03X^sub o^ + 1); and [lambda]^sub p^(X^sub p^) = 12 * L^sub n^(0.25X^sub p^ + 1).

If the firm in Scenario 2 operates in Industry I, the total expense associated with the optimal recovery strategy turns out to be $43.80, with $40.23 for outcome recovery and $3.57 for process recovery. In contrast, if the firm operates in Industry II the optimal solution for the same failure situation will largely be a process recovery strategy, with a total recovery expense of $71.69, partitioned into $22.29 for outcome recovery and $49.40 for process recovery. As the lower left panel of Figure 2 shows, the optimal outcome and process expenses have nonmonotonical relationships with [theta]. As this example implies, even for firms with superior executional advantage on one type of recovery (outcome recovery in Scenario 2), the optimal recovery strategy will be a "mixed" strategy whose composition is influenced by industry characteristics.

Impact of Across-Customer Differences in Substitutability between Service Dimensions

Customers can vary in terms of the extent to which they consider outcome and process dimensions to be substitutable during recovery. This variation will be reflected by differences in the relative values of the sensitivity parameters [gamma]^sub o^ and [gamma]^sub p^. For instance, [gamma]^sub o^ [much less than] [gamma]^sub p^ implies that correcting a process failure with an outcome recovery is not very acceptable; [gamma]^sub o^ = [gamma]^sub p^ implies moderate substitutability between the two dimensions; and [gamma]^sub o^ [much greater than] [gamma]^sub p^ implies a high degree of substitutability of outcome recovery for process failure. The impact of the degree of substitutability on the optimal recovery strategy can be illustrated by considering three customers-say I, II, and III-with all assumed to have the same sensitivity to process recovery (say, [gamma]^sub p^ = 1 for simplicity), but varying degrees of sensitivity to outcome recovery: [gamma]^sub o^ = .2 for Customer I, [gamma]^sub o^ = 1 for Customer II, and [gamma]^sub o^ = 2.5 for Customer III.

Scenario 3

This scenario is the same as Scenario 1 except for the aforementioned variations in the customers' sensitivity to outcome versus process outcomes. Specifically, all input values other than for [gamma]^sub o^ are the same as in Scenario 1 and the cost functions are the same as for Firm A: B = 0; k = 0; F^sub o^ = 0; F^sub p^ = 30; [theta] = 0.5; [delta]^sub o^ = 1; [delta]^sub p^ = 1; [gamma]^sub p^ = 1; [lambda]^sub o^(X^sub o^) = 50 * L^sub n^(0.03X^sub o^ + 1); and [lambda]^sub p^(X^sub p^) = 12 * L^sub n^(0.25X^sub p^ + 1).

For recovering from the process failure in the above scenario, the optimal solution for Customer I (with [gamma]^sub o^ = .2) is primarily a process recovery strategy, with $40.55 invested in process recovery and only $3.79 in outcome recovery, for a total of $44.34. But for Customer II (with [gamma]^sub o^ = 1) and Customer III (with [gamma]^sub o^ = 2.5), it is best for the firm to take advantage of its greater cost efficiency on the outcome dimension and invest more on outcome recovery. The optimal solutions in these two cases are X^sub o^ = $13.96, X^sub p^ = $7.35, and X = $21.31 for Customer II, and X^sub o^ = $8.98, X^sub p^ = $.06, and X = $9.04 for Customer III. As these illustrative results imply, the matching recovery strategies typically suggested by past studies are optimal only when customers' value substitution between the two dimensions is very low; as the degree of substitutability increases firms will have more flexibility in designing their recovery strategies. By leveraging that flexibility and designing recovery strategies, which capitalize on their cost advantages, firms can significantly reduce their total recovery expenditures. The lower right panel in Figure 2 pictorially demonstrates the decline in total recovery expenditures as substitutability increases and optimal strategies that take advantage of the increasing substitutability are employed.

Model Extensions to Examine More Complex Scenarios

This article represents an inaugural attempt to introduce a mathematically rigorous approach into the service-recovery domain, which thus far has been dominated by descriptive and survey-based studies that focus solely on customers' assessments of service failure and recovery encounters. As such, our proposed model is necessarily basic, with simplifying assumptions. However, one can easily relax the assumptions and extend the model to analyze more complex scenarios. For example, consider a customer's sensitivities to value loss ([delta]^sub o^ and [delta]^sub p^) and value gain ([gamma]^sub o^ and [gamma]^sub p^). Our basic model assumes the various sensitivity parameters to be static. This assumption can be relaxed to accommodate scenarios wherein a customer's sensitivities to value losses and gains may vary with the magnitude of value loss or value gain, as predicted by prospect theory (Kahneman & Tversky, 1979). Under a relaxed assumption to accommodate the predictions of prospect theory, a customer's sensitivity to value gain will decline gradually as the amount of gain increases, while his or her sensitivity to value loss will increase as the amount of the loss increases. To reflect the dynamic nature of the four sensitivity parameters we can augment the basic model by replacing [delta]^sub o^ with [mu]^sub o^(F^sub o^), [delta]^sub p^ with [mu]^sub p^(F^sub p^), [gamma]^sub o^ with v^sub o^([lambda]^sub o^(X^sub o^)), and [gamma]^sub p^ with v^sub p^([lambda]^sub p^(X^sub p^)) in equations (2) and (3) where [mu]'^sub o^(F^sub o^) > 0, [mu]'^sub p^(F^sub p^) > 0, v'^sub o^([lambda]^sub o^(X^sub o^)) < 0, and v'^sub p^([lambda]^sub p^(X^sub p^)) < 0.

Simultaneously solving equations (12) and (14), we can obtain the optimal solutions for expenditures on outcome recovery (X*^sub o^ = 0), process recovery (X*^sub p^ = 0), and total recovery (X*). As mentioned previously, satisfying the first order conditions of the Lagrangian function does not guarantee an optimal solution to the original problem as it depends on the structure of the functions involved. Closed-form solutions may exist only for simple functional forms; to determine the optimal solution for more complicated forms, an evaluation of all possible solutions by means of numerical simulations will be necessary. Next, we offer a numerical illustration to demonstrate changes to the optimal solution when the sensitivity parameters are dynamic rather than static.

Scenario 4

Except for the dynamic nature of the sensitivity parameters, this scenario is the same as Scenario 3, with the focal firm having the same cost functions as Firm A. That is, B = 0; k = O; F^sub o^ = 0; F^sub p^ = 30; [theta] = 0.5; [lambda]^sub o^(X^sub o^) = 50 * L^sub n^(0.03X^sub o^ + 1); and [lambda]^sub p^(X^sub p^) =12* L^sub n^(0.25X^sub p^+ 1). We assume the following functional forms for the four sensitivity parameters to make them dynamic: [delta]^sub o^ = .2 + F^sub o^/20; [delta]^sub p^ = .5 + F^sub p^/30; [gamma]^sub o^ = 1 -[lambda]^sub o^(X^sub o^)/150; and [gamma]^sub p^ = 1.2 - [lambda]^sub p^(X^sub p^)/120.

In the above scenario, the rationale for the specific expressions used to represent the four sensitivity parameters is as follows. When the most commonly seen mistakes occur (e.g., F^sub o^ = 16 and F^sub p^ = 15 in this scenario), the customer's sensitivities to outcome and process losses are "about average" and turn out to be [delta]^sub o^ = 1 and [delta]^sub p^ = 1 (same as those for Customer II in Scenario 3). When less serious failures occur (i.e., F^sub o^ < 16 and F^sub p^ < 15), the customer is more forgiving and perceives less value loss than would Customer II in Scenario 3 (i.e., [delta]^sub o^ and [delta]^sub p^ < 1). However, when failures that are more serious than usual occur (i.e., F^sub o^ > 16 and F^sub p^ > 15), the customer perceives greater value loss (i.e., [delta]^sub o^ and [delta]^sub p^ > 1). The customer's sensitivities to outcome and process recovery are set to start at 1 and 1.2, respectively, and then decrease gradually as the magnitude of the restitution increases.

Compared with Customer II in Scenario 3, for whom the optimal recovery strategy (as illustrated before) was to spend $13.96 on the outcome dimension and $7.35 on the process dimension, the customer in Scenario 4 experiences a value loss of 45 units (because F^sub p^ = 30 and the corresponding [delta]^sub p^ = 1.5), correcting which will require at least $27.89 for outcome recovery (with [gamma]^sub o^ = .80) and $ 17.37 for process recovery (with [gamma]^sub p^ = 1.03). If the firm considers this customer to be desirable and sets a higher value recovery target, say k = 3 (instead of 0), the expenditures for outcome and process recovery will be even higher ($36.50 and $23.04, respectively). This illustration demonstrates that our basic model can be extended to incorporate various complexities and study their impact on the optimum recovery strategy.

DISCUSSION

Key Contributions

Previous research on service recovery has focused primarily on customer reactions to service failures and recovery efforts, and has generally employed experimental or survey-based approaches. This article contributes significantly to the extant literature by (a) proposing a framework that incorporates both the customer's and firm's perspectives and (b) employing a mathematical approach for arriving at optimal recovery strategies. Based on insights from the proposed mathematical model, comparative statics, and illustrative results, this article makes several additional contributions. First, it highlights the importance of recognizing and understanding the degree of substitutability between outcome and process recovery strategies. While conventional wisdom suggests that a recovery strategy that matches the type of failure is likely to be the best, insights from our proposed model suggest that matching recovery strategies may be suboptimal and increase the recovery expenses when the degree of substitutability is high. Second, our framework also demonstrates that, apart from the magnitude of a failure, the customer's prior perceived value and the firm's value recovery target could significantly influence the total recovery expenditure and its optimal allocation between outcome and process recovery. Thus, a recovery strategy that explicitly incorporates customers' current perceptions of value and future contributions to revenues is likely to be more efficient than one that simply mimics the type and magnitude of the service failure. Finally, as illustrated earlier, our basic framework can be extended to accommodate more complex scenarios. As such it serves as a foundation for future modeling efforts and scholarly research on service recovery.

Guidelines for Model Validation and Empirical Testing

As mentioned previously, the framework discussed and illustrated in this article is an initial attempt at developing a mathematical and value-based approach for making sound service-recovery decisions. The proposed model has face validity by virtue of its being consistent with the extant findings and established theories that were invoked in developing it. The reasonableness of the patterns of results obtained in illustrating the model under various hypothetical scenarios also contributes to its face validity. Nevertheless, more extensive empirical assessment of the model is needed to further validate it and enhance its value to managers. In particular, the robustness of the model's predictions must be examined in various settings involving different industries, customer segments, and failure contexts (Miller et al., 2000, offer an exemplar for such empirical validation).

The firm's actual implementation of the optimal recovery strategy is composed of five steps: (1) operationalizing model parameters based on past data, existing cumulative knowledge of the consumer, and managerial intuition; (2) inputting into the model information pertaining to the specific failure-recovery incident; (3) deriving optimal outcome and/or process recovery strategies from the model; (4) allocating resources to outcome and process recovery tactics based on the optimal solution; and (5) executing the strategies. Implementing the model and linking it to managerial action on an ongoing basis will require the use of methodologies such as conjoint analysis (in which managerial actions or service attributes are linked and scaled to a common measure of customer utility, similar to the units of analysis used to represent customer value perceptions in our illustrative scenarios), regression analysis (to establish model parameters based on generalization across customers or customer segments), and what-if sensitivity analysis (to estimate the impact of different assumptions and input values on the optimal strategies). It is worth noting that ongoing use of the model after implementation might involve a learning period during which a company continues to calibrate the model, and the measurement units/scales evolve to be consistent with the company's data collection and analysis systems. Therefore, the initial scale of measurement-for example, whether the customer's perceived value is measured on a scale of 0-1 (as is typically done in conjoint analyses or discrete choice probability models) or 1-100 (to enable more convenient classification and discussion)-does not matter because in steady-state implementation of the model, the other model parameters (e.g., sensitivity parameters) and computations will be adjusted automatically to be consistent with the scale of measurement. Moreover, the model's main benefit is to provide insight into the relative emphasis that a firm should place on outcome versus process recovery strategies in a given context so as to minimize recovery costs.

Operationalizing the model's parameters, a prerequisite for empirical testing, requires careful data collection and close cooperation between service managers and academic researchers. While such cooperation is important in any field study, it is especially critical for research involving service failures and recovery since this is a potentially sensitive area and could trigger ethical questions (e.g., when service failures are to be "manipulated" in an experimental setting). Moreover, inputs and educated judgments from managers would be necessary to operationalize certain parameters such as k, the value recovery target. Drawing upon relevant insights from previously employed methodologies and scales, we next suggest general approaches for gathering the data needed to operationalize the various parameters.

Field surveys of failure events, failure severity, and recovery satisfaction can be conducted and the collected data analyzed (e.g., through regression analysis in which overall customer satisfaction is the dependent variable and failure severity and recovery satisfaction on the outcome and process dimensions are independent variables) to estimate customers' sensitivities to value loss from failure and value gain from recovery (i.e., [delta]^sub o^, [delta]^sub p^, [gamma]^sub o^ and [gamma]^sub p^). Field surveys to collect data about customer reactions to service failures and recovery can be supplemented or substituted with other data-collection methods such as the critical incidence technique and scenario-design methods (e.g., Bitner et al., 1994; Smith et al., 1999). Data to operationalize customers' prior perceived value (i.e., B) can also be collected through similar surveys.

The calibration of [theta], the importance of the outcome dimension relative to the process dimension, can be done through several alternative methods. For instance, importance weights can be derived by asking customers to respond to constant-sum scales such as the ones used to measure the relative importance of the SERVQUAL dimensions (Parasuraman, Zeithaml, & Berry, 1988). Secondary data sources that are externally available (e.g., industry indices) can also be used to derive baseline estimates of [theta]. In addition, managers in the focal firms or industry experts could be asked to estimate [theta].

The firm can specify the value of k for each individual customer (or customer segment if the customer base is too large) to be consistent with an appropriate overall measure or index of the customer's (or segment's) importance to the firm. For instance, Zeithaml et al. (2001) offer guidelines for categorizing a firm's customers into platinum, gold, iron, and lead tiers based on overall profitability. A firm employing such a segmentation approach could use, say, a scale of 1-100 for k and assign target recovery values of 100, 75, 50, and 25 for its platinum, gold, iron, and lead customers, respectively. While the metric used for scaling k and the specific values assigned to different groups are subjective, management can set these values to be consistent with its assessment of each customer's or segment's importance to the firm; moreover, management can vary the input values of k to reflect alternative assumptions about customers and examine the output values to see whether and to what extent the optimal strategies would change. In this regard, our model has the characteristics of "decision calculus" models-first discussed by Little (1970) and subsequently adopted by others-which are mathematically based decision-making aids for managers.

Data for estimating other parameters can be obtained from company sources. To estimate the cost functions for the outcome and process recovery strategies (i.e., [lambda]^sub o^(X^sub o^) and [lambda]^sub p^(X^sub p^)), researchers need cost data from the company (Rust et al., 1995). Because companies within a given service industry may have different strategic advantages (e.g., some are known for their personal assistance, while others excel at technology-based delivery), their cost structures can vary. In addition, researchers should consider potential cost variations across different service seasons. For instance, the cost associated with a hotel's outcome-recovery strategy of providing a free-accommodation coupon will be substantially lower during a lean season than a peak season. Similarly, a restaurant's process-recovery cost of an apology or explanation could be higher during the busy dinner time than during mid-afternoon.

Table 2 lists the model's parameters and potential approaches for operationalizing them.

Managerial Implications

By incorporating the traditional goal of profitability and the customer-relationship-management goal of enhancing customer-perceived value, the proposed model offers managers a framework for designing optimal service recovery strategies after taking into account the trade-offs between the two potentially conflicting, but equally important, goals. In particular, the model emphasizes the need for service managers to focus on critical customer factors (e.g., customers' sensitivity to value losses due to failures and value gains from recovery) and variations in those factors across customers and contexts. Such a focus will help managers craft customized service recovery strategies that address the dual goals of profitability and customer-perceived value in the most effective and efficient fashion.

Once the proposed model has been operationalized in a given context, service managers can conduct cluster analyses to segment customers based on their sensitivity to monetary (or outcome) gain/loss ([delta]^sub o^ and [gamma]^sub o^) and quality (or process) gain/loss ([delta]^sub p^ and [gamma]^sub p^). Informed by the optimal solutions to varying sets of parameter values, managers can then design tailored recovery strategies for each cluster. While general value-based clusters of customers have been identified in the literature-for example, "price butterflies" and "quality/convenience pursuers" (Reichheld & Schefter, 2000)-the approach proposed herein will produce more context-specific clusters that are especially appropriate for developing service recovery strategies.

Companies that are constrained to using a standardized recovery strategy (e.g., due to limited recovery resources or rigid government regulations) can still use this model to identify defection-prone customers. Given a fixed recovery budget (i.e., X^sub o^ and X^sub p^), a predetermined cost structure [lambda]^sub o^(X^sub o^) and [lambda]^sub p^(X^sub p^), and other parameter values ([delta]^sub o^, [delta]^sub p^, [gamma]^sub o^, [gamma]^sub p^, and [theta]), the model can predict a dissatisfied customer's perceived value after recovery efforts (i.e., k). An examination of customers' postrecovery perceived values predicted by the model vis--vis the prerecovery values can help identify the most vulnerable customers and take appropriate corrective actions.

Our model also prompts service managers to focus attention on potential cost-structure differences across companies and industries for providing outcome and process recovery, and the implications of those differences for recovery strategies. For instance, if a company and its key competitor have dramatically different cost structures, simply copying the competitor's recovery strategies for similar service failures, which is a common practice, will be suboptimal. There is thus a need for detailed recovery efficiency analyses along both the outcome and process dimensions when formulating service recovery strategies.

Understanding the cost structures associated with service recovery is especially important in emerging service sectors wherein the cost structures are still at a formative stage. For instance, in the domain of technology-based self-service, the process dimension involves no interpersonal contact. Instead, the convenience of the technology application or service-process design is crucial for customer satisfaction (e.g., Meuter, Ostrom, Roundtree, & Bitner, 2000). However, the cost of improving the process design (e.g., increasing the degree of customization of the machine-to-customer interfaces) could be much higher than in the case of traditional employee-delivered services. In situations like these, a general recommendation from the proposed model is to choose recovery strategies that involve low additional cost but are likely to deliver high value from the customer's perspective. For instance, most self-service encounters (such as an e-book download or an ATM transaction) consist of repetitive customer operations with negligible incremental cost to the company. As such, for failures in these self-service encounters, an effective solution could involve prompting customers to "self-recover" (e.g., compensating for problems triggered by an inconvenient design of an e-book delivery process with free downloads), which could more than satisfy customers at minimal cost.

Future Research Directions

As already mentioned, empirical testing of the proposed model is aresearch priority. In addition, a promising avenue for further research is to extend the model by incorporating potential interactions between parameters and examining their impact on the optimum recovery solutions. We earlier discussed one such extension-namely, making the sensitivity parameters ([delta]^sub o^, [delta]^sub p^, [gamma]^sub o^ and [gamma]^sub p^) functions of the value losses due to failures and value gains due to recovery. Several other extensions are possible and could provide new insights. For instance, a customer's previous perceived value (B) and risk profile ([delta]^sub o^ and [delta]^sub p^) probably have a negative relationship as suggested by Bolton's (1998) study, which found that as customers become more loyal to a service firm, they are more likely to be trusting (i.e., less risk averse) based on a belief that the firm will not take advantage of them. Another possible extension is to incorporate a positive relationship between a customer's prior perceived value (B) and the service firm's value recovery target (k). The plausibility of such a relationship is supported by findings suggesting that retaining loyal customers-relative to acquiring new customers-leads to higher and more stable customer lifetime value predictions (Blattberg & Deighton, 1996), which, in turn, should prompt the firm to set higher value recovery targets for such customers.

The proposed model suggests that a customer's prior perceived value (B) and the firm's recovery target (k) could have a stronger effect on optimal recovery expenditures (outcome, process, or total) than that of the magnitude of outcome or process failure (F^sub p^ or F^sub p^). This suggestion is consistent with findings from Bolton's (1998) wireless-service study showing that a customer's prior satisfaction or perceived value significantly influenced the formation of updated service-value perceptions. Collectively, these insights underscore the need for increased research attention on developing broad-based, integrated approaches for designing recovery strategies-approaches that go beyond merely considering failure type and magnitude, and explicitly incorporate the customer's assessment of the value delivered by the firm (e.g., as represented by B in the model) and the firm's assessment of the customer's value (e.g., as represented by k in the model). [Received: February 2003. Accepted: January 2004.]