internal customers and internal suppliers

8/6/2019 Internal Customers and Internal Suppliers

1/14

JOHN R. HAUSER, DU NCAN I. SIMESTER, and BIRGER WERNERFELT*To push a customer and m arket orientation deep into the organization,many firms have adopted systems by which internal customers evaluateinternal supp liers. The internal su pplier receives a larger bonu s for a high-er evaluation. The authors examine two internal customer-internal sup-plier incentive systems. In one system, the internal customer provides the

evaluation implicitly by selec ting the percentage of its bonus that is basedon market outcomes (e.g.. a combination of net sales and customer sat-isfaction if these measures can be tied to incremental profits). The inter-nal supplier's reward is based on the percentage that the internal cus-tomer chooses. In the second systenn. the internal customer selects tar-get market outcomes, and the internal supplier is rewarded on the basisof the target. In each incentive system, some risk is transferred from thefirm to the em ployees, and the firm must pay for this; but in return, the firmneed not observe either the internal supplier's or the internal customer'sactions. The incentive systems are robust even if the firm guesses wrong-ly about what employees perceive as costly and about how employeeactions affect profit. The authors discuss how these systems relate tointernal customer satisfaction systems and profit centers.

Internal Customers and Internal SuppliersIn order to drive customer satisfaction with our cus-tomers, IBM employees need to be satisfied with theorganization and strive to exceed their own intemal cus-tomer expectations.

Brooks Carder and James D. Clark. "The Theoryand Practice of Employee Recognition"

[Metropolitan Life Insurance Company of New York]developed a comprehensive program of measuring theexpectation of all its customers, including both extemaland internal [employee] customers.... only 25% [of theemployees] are servicing the outside customer.Valarie A. Zeithaml, A. Parasuraman. and LeonardL. Berry, Delivering Qua lity Service

[At Weyerhaeuser] staff support departments such ashuman resources, accounting, and quality control haveused "customer requirements analysis deployment"with line departments, such as sales, marketing, and

*John R. Hauser is the Kirin Professor of Marketing, and BirgerWemerfell is Professor of Marketing, Sloan School of Management.Massachusetts Institute of Technology. Duncan I. Simester is AssistantProfessor of Marketing, Graduate School of Business, University ofChicago,. This research was funded by the International Center forResearch on the Management of Technology (ICRMOT). It has benefitedfrom presentations before the member companies and, in particular, from atwo-day ICRMOT special interest conference on the "Marketing/R&DInterface" that was held at 3M. This paper has benefited from seminars atthe Marketing Science Conference at the University of Arizona, DukeUniversity, University of Florida, University of Minnesota, MassachusettsInstitute of Technology, U niversity of Pennsylvania, and the U.S. ArmySoldier Systems Laboratory.

branch production.... [internal] customers are thenasked to rate the suppliers ... in meeting each of theirrequirements.Donald L. McLaurin and Shareen Bell, "MakingCustomer Service More Than Just a Slogan

DEVELOPING A CUSTOMER ORIENTATIONTHROUGHOUT THE FIRMIn the 1990s, many firms believe that they will be moprofitable if they can push a marketing orientation deep inthe organization, particularly in new product developmeand research and development (R&D). In fact, these goactre the top-listed and top-ranked research priorities of tMarketing Science Institute (1992-1994). Implementingmarketing orientation (including employees and supplieremains one of Marketing Science Institute's three "capittopics for 1994-1996. One aspect of this market orientatiis to focus intemal suppliers on serving their intemal cu

tomer who. in tum. serves the extemal customers. To mafirms, such intemal suppliers are the next challenge implementing a marketing orientation. The epigraphs reto IBM. Met Life, and Weyerhaeuser, respectively; othexamples include 3M. ABB. Battelle. Berlex. Cable Wireless. Chevron. Coming. Hoechst Celanese. KodaHonda, and Xerox.' Marketing departments are often tintemal customers of product development or R&D. thou'Examples (in order) are based on studies by Mitsch (1990), Ha(1993), Freundlich and Schroeder (1991), Azzolini and Shillaber (199personal communication with Cable & Wireless, Chevron, ComiHoechst Celanese, and Kodak; and studies by Henke, Krachenberg, Ly(1993) and Menezes (1991).

Journal of Marketing ResearchVol. XXXIII (Augu st 1996 ), 26 8- 28 0 268


2/14

Internal Customers and Suppliers 269in some cases, the marketing department is the intemal sup-plier that provides information on customer needs andrequirements (Kem 1993). In most cases, marketing profes-sionals are called on to help the firm develop a customer ori-entation for its intemal suppliers.In many of these firms, the intemal customers evaluatethe intemal suppliers. For example, at an imaging firm, theintemal customers evaluate their intemal suppliers on bothshort-term and long-term profit indicators. At an automobileparts firm, the evaluations include measures that can bejinked to the intemal customer's ability to maximize thefirm's profits. In some cases, the intemal supplier's com-pensation is tied directly to the evaluations; in other cases itis tied indirectly with the more qualitative job performanceevaluations. W hether the compensation is explicit or implic-it, most intemal suppliers recognize that, all else beingequal, they are mo re likely to be rewarded if they are evalu-ated well by the intemal customers.

There are at least two motivations for the intemal cus-tomer-intemal supplier evaluation systems. First, the goalsof the intemal supp liers may conflict with those of the firm.In addition to the usual problem that effort is costly toemployees, intemal suppliers may have different objectivesthan those of the firm. For example, one extensive studysuggests that many R&D scientists and engineers focus onpublication and discovery of kno wledge rather than on facil-itating the ability of the firm (through the intemal cus-tomers) to maximize profit (Allen and Katz 1992). In anoth-er example, Richardson (1985) suggests that the R&Ddepartment works on the technologies it prefers rather thanon the technologies needed by the business areas.Furthermore, these conflicting objectives are not limited toR&D groups (Finkelman and Goland 1990). Without incen-tives to the contrary, the research suggests that intemal sup-pliers underprioritize their customer's (and the firm's)concems.Second, the intemal customer often can evaluate theeffects of the intemal supplier's decisions, whereas manage-ment may not have the skill, information, or time to do so aseffectively. For example, Henke, Krachenberg, and Lyons(1993) give an example of how an intemal customer, theinterior trim team, had better knowledge of how to solve aproblem than did the overseeing product management team.This is especially true in R&D, where the decisions oftenrequire specialized scientific or engineering knowledge notpossessed by top managers. (In some cases, top managerscome from R&D, but this is the exception rather than therule.) Thu s, top managem ent direction or involvement is dif-ficult at best. On the other hand, intemal customers, such asmarketing groups that are affected by R&D's decisionsabout where to direct its actions and efforts, can often eval-uate R&D better than top management.Another factor, true in many but not all cases, is the sig-nificant time lags between the decisions made by the inter-nal supplier and the market outcomes. For example,McDonough and Leifer (1986) suggest that planning andmonitoring techniques rarely work for R&D teams, becausecomm ercial success is often not known for five to ten years.In these cases, it may be better to reward the intemal sup-plier on the basis of an intemal customer's evaluation thanon the basis of market outcomes. Although the time lag forthe intemal customer may be less than that for the intemal

supplier, it could still be significant. In this case, the intemalcustomer might, in tum, be rewarded on the basis of an eval-uation by its downstream customer. Altematively the firmmight choose to use other indicators that measure whetherthe intemal customer is making the decisions that are bestfor the firm (for one example, see Hauser, Simester, andWemerfelt 1994).We formulate the problem in terms of a marketing groupas the intemal customer and an R&D group as the intemalsupplier. For example, R&D might supply the technologythat the marketing (or product development) group uses todevelop a new product, or R&D might supply a more devel-oped product that the marketing group must then sell to theextemal customer.Although intemal customer evaluation systems are popu-lar, they are not always easy to implement. One issue is thatintemal customers may have a tendency to report favorablyon their colleagues. In fact, intemal suppliers might rewardsuch behavior with various perks to the intemal customer.Starcher (1992) gives an example in which the intemal sup-plier faced an aggressive goal to reduce the number of

defects found by the intemal customer. The intemal cus-tomer found fewer faults, but only because it allowed moredefects to be passed on to the final assembly group. Thiswas costly to the firm because it required more rework (formany other examples, see Zettelmeyer and Hauser 1995).The temptation for increasing an evaluation is greater ifthere is no cost to the intemal cu stomer for providing a high-er evaluation. For example, Zettelmeyer and Hauser (1995)report many examples in which intemal customers give uni-formly high evaluations if the intemal customer provides anevaluation on a one-to-five scale, if the intemal supplier istold the evaluation it receives (by whom), and if manage-ment never questions any of the intemal customer's evalua-tions. This temptation to provide high evaluations might be

counteracted if there is some cost to the intemal customerfor providing a higher evaluation. This might be as simpleas management questioning a history of "all fives"; it mighttake the form of management holding the intemal customerto higher standards if the intemal customer reports that itgets uniformly good input from its suppliers (i.e., gives allfives); or it might be formalized.We examine two reward systems that use intemal cus-tomer evaluations. The essential idea underlying both of theincentive schemes is that the intemal customer need notevaluate the intemal supplier by providing a written evalua-tion. It can reveal its evaluation of the irttemal customers byselecting the parameters of its own reward function.^ Bothsystems provide incentives to both marketing and R&D

groups such that, acting in their own best interests, eachchooses the actions that the firm would choose to maximizefirm profits if it had the information and ability to do sodirectly and had to reimburse emp loyees only for their cost-ly actions (as if the employees bear no risk). These systemsshare the properties of using simple-to-specify reward func-tions and being relatively robust to errors that the firm mightmake in selecting the parameters of the reward functions.We are interested in simple systems because they are more

^Systems in which an agent selects parameters of its incentive systemhave been used in sales force compensation (Basu et al. 1984; Lai andSrinivasan 1993; Mantrala and Raman 1990).


3/14

27 0 JOURNAL OF MARKETING R ESEAR CH, AUGU ST 1996Figure 1

CONCEPTUAL REPRESENTATION

R&D(Internal Supplier)Marketing

(Internal Customer)

Customerlikely to be implemented than more complicated systemsand/or systems that are more sensitive to the parameters ofthe reward functions.^ Although the two systems share manyproperties, they have distinct interpretations and thus pro-vide two altematives that firms can choose.

A FORMAL MODELWe consider two employee groups and one group ofextemal customers. This suffices to illustrate the basicpoints. For simplicity we call the intemal supp lier "Researchand Development," label it as R, and refer to it as theupstream employee group. We call the intemal customer"M arketing," label it as M, and refer to it as the downstreamgroup (see Figure 1).Research and Development (R) expends effort, r. This r

refers to the time and energy R expends to identify, discov-er, or improve technology that M, in tum, uses to developHa the language of agency theory, we seek first-best actions. However,because we allow noise in outcomes (and implied noise in the agents'rewards) and agents to be risk averse, agency theory recognizes that thefirm may need to reimburse agents for any additional risk that the incentivesystem im poses on them. The actions that minimize the net of profits minusthis extra compensation are called the second-best actions. First-bestactions may not be optimal in a second best world. However, the defmitionof second best does not consider the administrative costs of extremely com-plex system s. The defmition also assumes that the parameters of the rewardsystems, no matter how complex, can be set exactly. Thus, we sacrificesome additional compensation costs to obtain systems that are simple, easyto implement, and robust with respect to errors the firm might make inselecting parameters.

products for customers. Effort (r) also refers to decisionsthat R might make, which R views as costly because thedecisions confiict with R's personal objectives. This effor(r) is incremental above and beyond the effort R must allocate in the absence of an intemal customer-intemal supplier incentive system. It is important to think of r as costlyeffort. Research and Development (R) may work longhours, but if part of the time is on-the-job consumption thaconfiicts with the needs of the firm, then r may be less thanthe long hours would suggest. For example, Allen andKatz's (1992) and Richardson's (1985) studies (as well asour own experience) suggest that R prefers those technolo-gies that are new, interesting, and lead to peer recognitionand patents. These techno logies may confiict with the needof the intemal customer. Research and Development's (Refforts, r, might include the time and energy necessary tounderstand M's needs beyond that which R would otherwiseallocate. We represent the perceived costs to R as CR(r)where CR is thrice differentiable, increasing, and convexBecause the costs are incremental, we normalize CR(O) = 0Formally, we assume that after R chooses and expends r, Mcan observe r, but top management (the firm) cannot. Foexample, consider a situation drawn from our experiencewith the R and M divisions at a major oil company: R waworking on the problem of getting more information to Mfrom remote oil fields. In this situation, M (but not top m anagement) might be able to evaluate whether R's new datacompression algorithm allows enough information to betransferred so that M can meet its customer's needs.

Marketing (M) uses the technology that R develops andexpends its own incremental effort, m, to serve the customers. We define m to represent incremental and cosdyefforts, actions, and decisions. (Henceforth, we simply calm, efforts.) We represent the perceived costs to M as Ci (m)where C|^ is thrice differentiable, increasing, and convexWe norm alize c^^(O ) = 0. If R expen ds m ore effort, r, then Mfinds that its own efforts, are more effective. For example, abetter data compression algorithm might enable M to provide better service to its customers. However, M must alsoexpend costly effort to provide that better service. The firmdoes not observe m directly.We assume that the firm observes an indicator of the profits that it obtains from the actions of R and M. It uses thisprofit measure as a (noisy) indicator of r and m. In our example, the firm might observe the increase in profits (more oirecovered, reduced costs) due to the new data transfer system. That is, the firm might compare the profits it nowobtains with those it would have obtained using the old dattransfer system. (Here we assume the firm can account foother effects on the profitability of the remote oil field.)In practice, this profit indicator can take many formsZettelmeyer and Hauser (1995) report that one firm usemeasures of quality, cost-effectiveness, timeliness, communications, and satisfaction from the (extemal) customer aan indicator of profits from r and m. They also report thaanother firm uses downstream production cost, labor costquality cost, and production investments as indicators of theffect of r and m on short- and long-term profit. If we are touse these measures as proxies for incremental profit, wmust assume that the r and m that maximize these ind icator(net of cost) are the same values of the r and m that maximize incremental profit.


4/14

Internal Customers and Suppliers 271Hauser. Simester. and Wemerfelt (1994) provide anotherexample. They demonstrate that if the internal customermaximizes a weighted sum of (extemal) customer satisfac-tion and sales (net of costs) then the intemal customerchooses the efforts that maximize the firm's long-term prof-its. In their case, we would use a weighted sum of satisfac-tion and sales (net of costs) as a proxy for the incrementalprofits due to r and m (see also Anderson. Fomell. and

Lehmann 1994). For our purposes here, we only need thefirm to be able to observe some measure that indicates theincremental impact of r and m on the firm's profits. For sim-plicity, we call this outcome measure profits, or it(r.m). Weassume that the firm can scale the measure (or combinationof measures) in the units of currency so that it represents theincremental contribution to profits from R and M.Because no measure is perfect, we model the error itintroduces. We write the measure as equal to its mean,ir(r.m). plus zero-m ean and normally distributed noise."* e.That is.

Figure 2ORDER OF ACTIONS IN FORMAL REPRESENTATION

(COOPERATION ALLOWED)

(1 ) = Tr(r.m) + e.where e ~ N(0. CT^); T is thrice differentiable. increasing, andconcave in both arguments; and a^ > 0. We model the risk-neutral firm as using the expected value of TT in the profit-maximization equation that relates to R and M. (The expect-ed value is TT.)After observing r. M chooses an evaluation, s. that indi-cates to the firm how it values r. (We subsequently use S|and S2 to distinguish between the two reward sys tems w eanalyze.) We use s as a mnemonic device because we thinkof this evaluation as an indicator of how well the intemalsupplier satisfies the intemal customer. However, s may notbe measured on a typical satisfaction scale. In both of ourreward systems, we allow the interpretation that the firminfers s from M's choice of reward functions.

Marketing (M) chooses s before selecting m. but M antici-pates how it will set m. That is. M evaluates R and does soanticipating how it will use R's output to serve M's cus-tomers. For example. M might choose its bonus plan, andhence evaluate R. after observing a demonstration of the datacompression algorithm. Marketing (M) would do so. antici-pating how it would use that algorithm to serve its customersand knowing that s affects its own rewards. (Technically, wealso could have stated the sequence as M choosing s simulta-neously with m. because no one except M observes m direct-ly. Subsequently, we m odify this sequence of events to enab leR and M to cooperate on the selection of s; see Figure 2.)After observing s. the firm gives a reward, v. to R thatdepends on s. We write this function as v(s). At a later time,

the firm observes the profit measure, TT. and provides areward, w, to M that depends on this measure and M'schoice of s. We write this function as W(S.TT). We restrict ourattention to incentive systems with integrable and thrice dif-ferentiable^ v and w. which are concave in s. In keeping w ith^We here assume normally distributed error beeause that enables us toderive analytical expressions for linear and quadratic reward systems. Ourpropositions also apply to the special case of no error. We expect that thequalitative concepts apply, at least approximately, for more general errordistributions. See example approximations in Wemerfelt, Simester, andHauser's (1996) study.'These functions are integrable and thrice differentiable, except atboundaries imposed by any extemal constraints imposed on s.

1. Rewaidsystems.v(s) andw(s,*).announced.

2. R chooses ror does notparticipate.

3. M observesr. Firm doesnot.4. R and Magree on g

and s, or Mdoes notparticipate.

5. M evaluatesR with s. Rpays g to M.

6. R receives itsreward, v( s). 7. M chooses m.Firm d oes notobserve.8. Firm observes

the profitindicator , -fr.

9. M receives itsreward, w (s,*) .

the managerial statement of the problem, we considerrewards to R that are larger for higher implicit evaluations(increasing in s). We also want s to be an indicator of r'seffect on ir; thus, we restrict our attention to w such thatdwVdTds > 0.It is convenient to think of v and w as m onetary rew ards;however, they need not be. Any set of rewards that R and Mvalue and for which the firm must pay would be appropri-ate, including new equipment, training, recognition, andawards (Feldman 1992; Mitsch 1990). For simplicity, weassume that the amount that the firm pays is equal to thevalue that the employee group receives.We assume that the firm is risk-neutral and profit maxi-mizing and that both R and M act in their own best intereststo maximize their expected utilities. We assume that both Rand M are risk averse and that perceived costs to R and Mare measured on the same scale as are rewards.^ The utili-ties. U R and U |^. are

and

where U R and U ^ are integrable. thrice differentiable.increasing, and concave.We assume that the net utilities. D R and D |^. required byR and M to participate are set by the market that is. by theother options that R and M have available. (If there are anyswitching costs favoring the firm, then these are included inthe definition of D R and D ^ ) Thus, the total expected utili-ty of R's and M's rewards minus their costs for allocating rand m and reporting s must exceed D R and D ^ . We normal-ize the utility functions such that they imply that (riskless)market options for R and M are equal to zero. (If the marketoptions have risk, then they must be such that R and M con-sider them equivalent to a riskless option that is scaled tozero.) In its maximization of expected profits, a risk-neutralfirm attempts to set the expected utility to each employee

*The choice of a scaling constant is a nontrivial practical problem. Wesubsequently address the sensitivity of outcomes to the choice of parame-ters of V and w.


5/14

27 2 JOURNAL OF MARKETING R ESEARCH , AUGU ST 199group just above D R or D ^ if, by doing so, it can earn non-negative profits. For example, the oil firm would select vand w such that R and M are willing to develop and use anew data compression algorithm rather than continue toserve the customer with the old system. R ecall that the prac-tical problem requires that the oil firm do this without know-ing the technical details of compression algorithms.We summarize the sequence of moves described so far:

1. The firm announces an internal customer-internal supplierincentive system, v(s) and W(S,TT).2. R chooses its actions, r, or does not participate.3. M observes R 's actions, but the firm does not.4. M chooses s or does not participate.5. R is rewarded based on v(s).6. M chooses its actions, m, but the firm can not observe theseactions.7. The profit indicator, TT, is observed and the firm pays W(S,TT).This sequence of moves is a well-defined contracting prob-lem, and we could, in principle, evaluate the performance ofalternative v's and w's. In this contracting problem, withnoise and risk aversion, simple contracts do not do well. Thefirm can do better by tying pay to performance than by justpaying a fixed salary.In the formal contracting problem, we focus on one set ofactions, r and m. In practice, the firm would not reset v andw for every decision that R and M m ust make. The firmmight set v and w such that they apply to repeated interac-tions between R and M. We do not solve this problem for-mally. However, we show that our incentive schemes arerobust with respect to the specifics of TT, Cp , and c^; hence,it is likely that the key parameters of v and w do not need tobe set for every interaction. We formalize this robustnessissue subsequently.

RESEARCH AND DEVELOPMENT MIGHT REWARDMARKETING TO CHANGE ITS EVALUATIONInternal customers might have more opportunities tointeract with internal suppliers than do outside customers.' 'Hence, they might cooperate in setting s. We illustrate theconcept of cooperation^ with an example between a sales-person and the external customer. A colleague recently pur-chased an automobile. As part of the delivery transaction,our colleague w as asked to complete a customer-satisfactionsurvey. He did so to the best of his ability. After looking atthe customer's ratings, the salesperson said that the ratingswere unacceptable and that he would be fired if the ratingswere not increased. Our co lleague agreed to increase the rat-

'Th e ability of R and M to cooperate on s depends on there being a smallnumber of evaluators for any given evaluatee. If the number of evaluatorsis large, for example, hun dreds, then it is likely to be too costly for the eval-uatee to seek out every evaluator, and the cost imposed on the evaluator forproviding a higher evaluation would be small. Such systems look more liketraditional (external) customer satisfaction systems. For a discussion of themechanisms by which R and M groups communicate and cooperate, seeGriffin and Hauser (1992, 1996).*For the remainder of the article, we use the more positive term cooper-ation rather than the negative term collusion. Although the latter term ismore common in the economic literature, it has implicit connotations thatgo beyond those we wish to discuss here. Indeed, if the firm can anticipatehow R and M might cooperate in setting s, they might factor this into thereward system. In any case, we define precisely what we mean by cooper-ation and derive what cooperation im plies for how R and M interact. Wereturn to this issue in the final section.

ings, but in return for a year of routine maintenance paid foby the salesperson. The salesperson agreed, and the ratingwere increased. We spoke to an executive vice-president othe consulting firm that designed the ratings-based bonusystem that gave the salesperson a monetary bonus for high rating. The executive vice-president said that the automobile company was aware that instances such as the onwith our colleague might happen. The automobile companwanted satisfied customers (in the delivery transaction) anwas willing to pay for them. In the long run, the companhoped that the salesperson would find other methods of satisfying the customermethods that the salesperson woulfind less onerous than sharing his or her bonus with the customer. If the salesperson became more efficient in satisfyinthe customer, this would create surplus that, depending on future reward system, might be shared among the customerthe salesperson, the dealership, and the automobilcompany.

More recently, one of us purchased a new car. Not onldid the sales manager instruct the customer on how to complete the customer-satisfaction questionnaire and imply thahis access to supply depends on the ratings, but he sent thcustomer an expensive gift the day prior to the completionof the satisfaction questionnaire. (We were told that thmanufacturer allocated a supply of this popular car to dealerships on the basis of the ratings. Presumably the dealership found it more efficient to increase customer satisfactiowith this gift than with other forms of service. Certainly, thcustomer was satisfied.) We presume that, similar to thsalesperson ex ample, the automob ile company ho pes that, ithe long run, the dealership will find other, more efficienways to satisfy the customer.Managers and reward systems consultants have indicateto us that they believe that modest sharing of rewards icommon in internal customer-internal supplier systems. Fotwo documented cases see Gouldner's (1965) account of small gypsum mine and Sidrys and JakStaite's (1994account of the Lithuanian university system. See also Boston Globe (1994) editorial applauding frequent fiyeprograms.We now analyze cooperation with the formal model. Tsimplify the analysis, we follow Tirole (1986) and assumthat R and M find a way to make a binding agreemenexchanging goods or services that are valued at g in returfor a higher evaluation. The enforceability of the agreemencould come from expected future interactions between Rand M or from social norms (e.g., in Sidrys and JakStaite'[1994] data, agreements occur more often with local instructors than with foreign instructors). In the agreement, w e notthat the assumption is that R and M can cooperate on s bunot on r. (The effort, r, has already been set.) The paymeng, cannot be contingent on TT. In situations in which R anM can cooperate directly on r as well as on s, this assumption restricts the domain of our analysis. However, wbelieve that this assumption is an important starting poinand applies to most of the situations we have observed. Wfind that it is much harder to monitor agreements abouaverage effort over a month (including detailed technologdecisions) than it is to monitor agreements about a singlperformance evaluation. We leave cooperation on r to further research. Thus, formally, we augment the sequence oevents such that R and M can make a binding ag reemen


6/14

Internal Customers and Suppliers 27 3(g,s), after M has observed r but before M has selected s (seeFigure 2).The gains (if any) from the agreement can be split inmany ways between R and M. To simplify the exposition wemodel the split as a take-it-or-leave-it offer of (g,s) from Rto M. This means that M receives only as much as is neces-sary to induce M to report the agreed-upon s. This assump-tion does not affect the qualitative interpretations. We couldderive similar results for other sharing mechanisms.

We define ifi and s as the efforts and evaluation that Mselects to maximize U^j for a given r with no cooperation.For concave UM( ), this maximization of expected utility byM defines three continuously differentiable functions, m(r),s(r), and TT(r). Tha t is , after R selects r, these functions tellus the efforts, iti, and evaluation, , that M would select ifcooperation were precluded. Now suppose that for a given r,R wants to infiuence M to choose another s that is morefavorable to R. This s implies an m(r,s) that maximizes M'sexpected utility, given r and s. It also implies a 'ir(r, m).(No te that rh may differ from iti, and IT may differ from IT ifs differs from s.) To infiuence M to select s, R must give Man amount, g, that at least compensates for M's loss. Thismeans that M's expected utility with an agreement, (g,s),must at least equal the expected utility that M could obtainwithout accepting g. Thus, the minimum g that M willaccept is defined by Equation 3.(3) g J =Research and Development (R) has no incentive to givemore than this g in retum for s, thus R will attempt to get gdown to that defined in Equation 3, and M will try to get gup to that defined in Equation 3. Thus, Equation 3 defines acritical value of g for every r. We write this critical value asg(r). Research and Development (R) wants to maximize itsown well-being. That is, R will select f and s to maxim ize itsown expected utility:(4 ) EU R [ S , r] = EUR [V(S) - CR (?) - g(7)].where g is implied by Equation 3 and rh, ifi, and s are implic-it in M's maximization problems.In summary, R maximizes the expression in Equation 4subject to the constraints imposed by the two maximizationproblems that define Equation 3. The right-hand side ofEquation 3 describes what M would do if there were nocooperation, and the left-hand side of Equation 3 describeswhat M would do if cooperation were allowed. Naturally,both sides of Equation 3 must be at least as large as thatwhich M could obtain by not participating. Equation 4 mustbe at least as large as that which R could obtain by not par-ticipating (R m ust consider M 's participation b ecause it can-not get s if M does not participate). The firm is interested inmaximizing its profits, so it will attempt to select v and wsuch that it gets the efforts it wants and does not pay R andM m ore than is necessary. Once the firm specifies v and w,these constraints and maximization problems are sufficientto solve for r, ifi, s, tfi, s, and the implied g and it.

We now use this stmcture to examine two differentreward systems. We study these reward functions becausethey are simple and relatively robust with respect to errorsthe firm might make in selecting the reward functions. We

anticipate that the firm would choose the system that bestmatches its culture.TWO PRACTICAL INTERNAL CUSTOMER-INTERNALSUPPLIER INCENTIVE SYSTEMS

Our analysis of these reward systems is driven by themanagerial problem faced by R and Mselecting the"right" technology. For example, Zettelmeyer and Hauser(1995) report that chief executive officers and chief techni-cal officers are more concemed that R and M select the righttechnology than they are about minimizing the extra incen-tives for which they must pay R and M for any risk that theincentive system imposes on R and M. (Chief executiveofficers and chief technical officers are concemed aboutincentive system costs and would like to keep them small,but this appears to be a less critical problem than providingthe incentives for the right technology.) Thus, in our analy-ses, we focus on reward systems that provide R and M withthe incentives to select those actions, r* and m*, that maxi-mize the (risk-neutral) firm's expected profits if it had thepower and knowledge to dictate actions, observe actions,and reimburse emp loyees only for their costly actions (as ifthe employees bear no risk). That is.(5 ) r' and m* maximize TT(r, m) - CM (m) - CR (r).For each of the two reward systems that we study, we seekthose particular v's and w's that cause R and M to select r'and m*. In the language of agency theory, r' and m* arecalled the first-best actions.Although we concentrate on r* and m', we cannot neglectthe costs that risk in the incentive system imposes on R andM. Because the intemal customer-intemal supplier systemsforce risk-averse employees to accept risk, the firm mustreimburse those employees for accepting that risk. Theamount that die firm must pay is called a risk penalty. Wecompute the implied risk penalty and show how the para-meters of the reward functions affect that risk penalty. Thefirm can then select the reward system and parameters (fromthe two systems we analyze) to minimize risk costs.Altematively, it can weigh these costs against the ease ofimplementing the reward system.The analysis of the problem of choosing incentive sys-tems for risk-averse agents whose actions are unobservableis a topic in agency theory (Holmstrom 1979). One bench -mark in agency theory, called the second best, is to seekoptimal incentive systems that maximize the net of profitsminus the risk penalty. According to this benchm ark, it maynot be optimal to have agents choose r' and m*. Thus, oursystems might not lead to optimal profits as defined byagency theory.' On the other hand, optimal solutions areoften extremely complicated and sensitive to model specifi-cation (Hart and Holmstrom 1987). However, the definitionof optimal does not take into account that complex systemsmight impose administrative costs or that complex systemsmight be confusing for real employees and hence lead tononrational actions that neither maximize employee utilitynor firm profits. Our systems are less likely to impose suchcosts because they are simple and robust.

'For the case of no noise in profits and/or risk-neutral employees, oursystems are optimal. For the case of low risk (as implied by noise and riskaversion), our systems are close to optimal.


7/14

274 JOURNAL OF MARKETING RESEARCH. AUGUST 199To simplify exposition, we conduct our analyses in the

context of employee groups with constant (absolute) risk-averse utility functions (Keeney and Raiffa 1976). Tha t is.(6a)(6b)

U RU M

Variable Outcome-Based Compensation SystemsWe begin with one of the simplest specifications of v andwlinear functions of s. Linear functions provide a valuablestarting point (and a useful benchmark) iind. in single-agentproblems, have proven to be robust (Holmstrom and Milgrom1987; Lilien. Kotler. and M oorthy 1992). We begin by statingthe general form (Equation 7) and then derive a set of para-meter values that provide incentives to R and M such that theyselect r* and m*. Formally, the variable outcome-based com-pensation system is given by the following functions, whereyo' yi ' ^o' 2 |' ^nd ^3 ^ constants chosen by the firm.( 7 a ) R : v ( S | ) = y ^ + y i S , ;( 7 b ) M : w ( s , . I T ) = Z Q + z , ( l - S | ) + 2 3 8 , 7 : . S | e [ 0 . 1 ]

That is. after observing r and anticipating m. M is askedto evaluate R on a scale from 0 to 1. This evaluation deter-mines the portion of its compensation that is determined byincremental profits. The remaining portion of M's bonus isfixed. (In an altemative interpretation. M is simply asked toselect the percentage of its compensation plan that is basedon incremental profits, and the firm interprets M's selectionas an implicit evaluation of R.) We call this system the vari-able outcome-based compensation system because theimplicit evaluation, s. determines how much of M's bonusdepends o n the (variable) incremental profit.In other words, if S| = 1. then M receives its fixed bonus.ZQ . plus a bonus proportional to the profit indicator. Z3'fr. Onthe other hand, if S| = 0 . then M receives only a fixed bonus.ZQ + Z|. For intermediate s^ , the portion is determined by Sj.(We could also specify Sj as a percentage.) Intuitively welink this implicit evaluation, s,. to R because if R does itsjob well, then M will prefer to be rewarded on the profitindicator; and if R does its job poorly, then M will prefer theguaranteed bonus. The firm attempts to select the parame-ters of the functions so that R and M choose r* and m*. (Toparticipate, R and M are compensated for their efforts andany risk they must bear.)

The variable outcome-based com pensation system is a for-malization of the linear reward systemspopular in market-ing and Total Quality Managementthat we have seen inpractice. If that evaluation is an intemal customer satisfac-tion rating, if there is some cost to M in providing that rating,and if R's and M's bonuses are linear in M's rating, then thefollowing proposition gives us formal tools with which tointerpret and improve intemal customer satisfaction systems:P | (variable outcome-based compensation): For z, and y, abovecritical values and for Z3 = 1. the variable outcome-basedcompensation system gives incentives to R and M such that,acting in their own best interests, they select r* and m*.

The proof and the critical values are in the Appendix.'^The basic idea is that if Z| is above a critical value, then Min the absence of cooperation, will set S| =0 . If y| is abovits critical valu e. R has sufficient incentive to prov ide g to Min order to obtain S] = 1. Research and Development (Rwants to keep g as small as possible, and keeping g smacoincides with selecting r* and m*.For P|. we can compute g. In addition, because M and bear risk, we can compu te the risk penalty that the firm m upay. To compute this penalty we recognize that, in the soltion to Equation 5. the firm would only need to pay R and Mfor their effort costs. CR(r*) and CM(m'). The risk penalty the amount by which v + w exceeds the sum of these costThus, with algebra we obtain

g = Zlil = 1.

Risk Penalty =

(8a)(8b)and(8c)

The firm can make the g small by selecting a Zj close tits critical value, but the risk penalty is not affected by and y|. The risk penalty implied by this system is equal tthat which the firm would incur by transferring all risk to M(We subsequently investigate a system with a smaller rispenalty.)With the parameters of P|. the optimization implies thextreme value solution. Z3S1 = 1. That is. M's co mpensatibecomes a constant plus TT. Thus, in equilibrium, the firoffers M the opportunity to accept responsibility for thincremental outcome, -fr. and M accepts this responsibilitby choosing Sj = I. Research and Development (R) rewarded w henever M gives an evaluation that indicates thM accepts this responsibility. This system gives R the incetives to provide r and g so that M will accept thresponsibility.Transferring responsibility to M is similar in some wato a mechanism that the agency-theory literature (e.gMilgrom and Roberts 1992. pp. 236-39) calls "selling thfirm to the agent." However, in our case. M becomes thresidual claimant only for the incremental outcom es of r anm and o nly for this interaction. The firm retains responsibity for those outcomes (other than the measurement errothat do not depend on r and m. Although the actions and oucomes are the same as making M the residual claimant fthe incremental outcomes of r and m. we have found thmany managers find a linear evaluation system more resonable than "selling the firm." The latter, perhaps unintetionally and inadvertently, implies transferring assets, futuresponsibility, and future rights for global rather than incrmental actions. Interpreted with this perspective, the syste

"The parameters in P, cause M to report s, = I. Under some special coditions, the nrm can choose an altemative paran^eterization such that treported s, is at an intermediate value. The conditions (for r = r') requthat the risk, HXTV2. be so large that IT - c^ - (ji,a2/2 is sm alle r at m* thIT - C| is at m = 0. that is, when M 's function is primarily that of a supvisor for R. This altemative parameterization retains the proflt-center aresidual claimant interpretations but is not as robust as that in P\ . Also,the special case of a constrained linear w with no noise, it is possiblechoose a set of parame ters such that the Tirm maximizes profits w ithout clusion. Proofs of both results are available from the authors.


8/14

Internal Customers a nd Suppliers 275in Equation 7 is a practical means to im plement a profit cen-ter-like approach.The profit center relationship may be a new perspective.For example, Harari (1993) argues that internal customersatisfaction systems should be abandoned and replaced withprofit center systems. We have spoken to many managerswho are strong advocates of intemal customer-intemal sup-plier systems. None have described such systems as a meansto implement a profit center. Finally, and we discuss thissubsequently, the variable-compensation system is surpris-ingly robust.Target-Value Compensation Systems

We now introduce nonlinearity into the system by makingM's rewards a nonlinear function of s. Specifically, weselect a quadratic function of s - iT. The linear and quadrat-ic functions are not the only functional forms for w that willyield r* and m*, however they suffice to illustrate many ofthe principles of intemal customer-intemal supplier incen-tive systems. Each has a different, but practical, interpreta-tion. Our experience suggests that firms are more willing touse simple than complex functional forms in compensationsystems (see also Lilien, Kotler, and Moorthy 1992).Formally, the target-value system is given by the follow-ing functions:(9a)and(9b) M :

V(S2) = Vo + V1S2,

w ( s 2 , IT ) = Wo - W2 (S2 -where v^, v,, w^, and W2 are constants chosen by the firm.That is, after observ ing r and anticipating m, M selects a tar-get valu e, S2, for th e profit in dicator, TT. Marketing (M)receives its maximum bonus if the realized profit indicator,IT, equals the target and is penalized for deviations from thetarget. Note that the target-value function penalizes targetsthat are set too high an d too low. We have discussed thisconcept with managers at commercial banks, computermanufacturers, imaging firms, chemical companies, oilcompanies, and automotive companies. In each instance,they found the idea of a target-value system appealing andbelieved that the benefit of an accurate target could out-weigh concems about penalizing an employee group forexceeding its target. The target-value concept is similar toGo nik's reward functions (Gonik 1978; Mantrala andRaman 1990) used in sales force compensation. (Gonikreward functions encourage salespeople to make accurateforecasts by penalizing them for selling more or less thanthe targets they se t." )

P2 (target value): For v, = I and 0 < Wj < I/(2JI.CT2), the target-value compensation system gives incentives to R and M suchthat, acting in their own best interests, they select r* and m*."Gonik rew ard functions use abs olute deviations rather than quadraticdeviations and apply to a single agent rather than to an intemal cus-tomer-internal supplier dyad. By comparing the linear and quadratic sys-tems , w e see that quadratic functions can provide low er risk penalties.Gonik ab solute-value functions share the "make-or-break" properties of thelinear system. For the riskless case, it is pos sible to prove P2 for any con-cave function of (Sj - *) w ith a finite maximum.

The proof in the Appendix is constmctive. We first com-pute rewards for R and M that are implied by Equation 9.We use Equation 3 to compute the implied g. We use this gin Equation 4 to compute R's net rewards. After these sub-stitutions, we maximize E quation 4 subject to the constraintsimposed by Equation 3. This yields the equations for thegoals of R and M. We show that these goals yield the samesolution as Equation 5the firm's goals. Finally, we set v^,and Wj, so that both R and M get sufficient rew ards so thatthey prefer participating to not participating.

To get an intuitive feel for how the target-value functionworks, notice that, in the absence of cooperation, M wouldwant to minimize the expected deviation of S2 from TT and,hence, set S2 equal to IT. Because v, = 1, R's rewcU'ds arethen proportional to IT. With a positive g, R can get M toincrease S2 slightly. This makes R's rewards sensitive to M'scosts. When W2 is set in the given range, R's incentives aremaximized at r* and m'.It should not be surprising that we can find a family ofnonlinear reward functions, v and w, that yield r* and m*.There are a limited number of first-order equations impliedby the firm's optimization. Many functional families haveenough parameters so that these equations can be solved;however, some simple functional families, like constantreward s, do not. P2 shows that a quadratic system, w hich hasan intuitive interpretation in terms of targets, has sufficientparameters. General families may not be as simple or robust.(We analyze the robustness of P2 in the subsequent section.)Using the parameters of P2 as a basis, we compute g, theimplied evaluation, and the risk penalty.(10a) g = CM(m*) + 1/(4 w 2)(10b) S2 - I T ' = 1/(2 W2) - H a^(10c) Risk Penalty = -(2 j i)" ' log( I - 2^ o^ W2)

First, note that when there is no noise (a^ = 0), there is norisk penalty; but g is still positive, and the reported targetexceeds the amount that M will achieve. (The condition onW2 is required for other reason s, but it also assures that gexceeds M's costs and the evaluation exceeds the targetprofit.)Second, note that both g and the risk penalty depend onthe firm's choice of W2. If the firm chooses W2 close to itsupper bo und, then it can make g sm aller, but its risk penaltyincreases. Thus, for the target-value system, there is aninherent tension between g and the risk penalty. Supposethat we make M's penalty for misforecasting small(W2 - 0). Then, g becom es large, the distortion in the eval-uation (S2 versus IT ) becomes large, and the risk p enaltyapproaches what the firm would have incurred had it trans-ferred all risk to M . (If M bo re all the risk, its risk premiumwould be JJLCT2/2.) In other words, in systems in which thereis only mild social pressure for M to get the forecast right(W2 is small), selected targets are much larger than achiev-able targets, g is large, and the firm incurs a larger riskpenalty.Fo r n,(T > 1, the risk pena lty can be m inimized for a W2between the extremes, and this minimum is less than


9/14

27 6 JOURNA L OF MARKETING RE SEARC H, AUGUST 199In repeated situations, the firm might get to thatminimum by trial and error, but in the formal game, it needsto know M's risk aversion coefficient and the noise in theprofit m easure. For (JLO < 1, the firm can still get r* and m*,but the risk penalty exceeds

Figure 3RISK PENALTY

SENSITIVITYBoth of the incentive systems that we have examinedshare the property that if the firm sets the constants, v^ andWQ or y^ and z^, too low, either M o r R (or b oth) will choosenot to participate (if they are well informed). If the firm setsthese constants too high, either M or R (or both) will beoverpaid. This property is not unique to the systems westudy here.In the formal game, the firm must know such values asTT*, c ^ ' , CR' , |i,, a-2, and TT(r*,O) to set the fixed componentsof com pensation, v^ and w , or y^ and z^. This conceptualproblem is shared with all incentive systems.However, to implement the variable outcome-based com-pensation system or the target-value system, the firm mustdo more than select the fixed components of compensation.

It also must select the coefficients that determine how com-pensation varies as a function of the actions and evaluations.In the linear system, the firm must select the relative coeffi-cient, Z1/Z3, that sets the trade-offs that M must makebetween compensation that depends on outcomes and com-pensation that is guaranteed. The firm also must set the coef-ficient, yi , that determines how R is rewarded on the basisof s. In the quadratic system the firm must select the coeffi-cient, W 2, that pena lizes M for deviation s from its target.(Recall V, = 1.) W e now examine the sensitivity of the com-pensation systems to these variable parameters.W e have already shown that the variable compensationsystem is not sensitive to Zi and y, as long as they are abovetheir critical values and that the target-value system is notsensitive to W2 as long as it is within a reasonable range. W estate these facts as corollaries for emph asis. (The proofs areobvious by recognizing the conditions of Pi and P2, but forcomp leteness are given in the Append ix.) In C orollary 1, y]and Z| refer to parameters just above the critical values inPp In C orollary 2, W 2* refers to the W2 that minimizes therisk penalty.'3 The firm may have a hard time setting W 2'because it needs to know \L and cr^ to select this value.C orollary 1 (variable outcome-based compensation sensitivity):If the firm makes an error and selects a value, Zj',that is different than Z|, or a value, yi', that is dif-ferent than y| , then the system still yields r* and m*as long as Zj' > Zj" and y i' > yi. The risk penalty is

unaffected.\i,a > 1, the W2 that produces the minimum risk penalty isl - ((jLcr)-2]. For JICT < 1, the minimum occurs as W2 - 0. An inte-

rior minimum occurs because the quadratic target value function introducesboth a linear error term, (1 - 2|jLcr2w2)e, and a quadratic error term, -W2^^ 1and W2 e (0,W2"] for a W 2" E (W2*, (2(J ,O-2)-I]. See the Appendix for details.

RiskPenalty

[ " T a r g e t V a l u e1 Var iab le Compensat ion

/

C orollary 2 (target-value sensitivity): If the firm m akes an errand selects a value, W 2', that is different than toptimal value, W2*, then the target-value com penstion system still yields r* and m* as long as 0 < W< 1/(2|J,CT2). The risk penalty increases, but it can less than the risk penalty for the variable outcombased compensation system when JJLCT > 1.COMMENTARY

Both intemal customer-intemal supplier systems haintuitive interpretations and provide incentives to R and Mto select r* and m*. However, the target-value system can bsuperior to the variable compensation system on the criteron of the risk penalty. (Figure 3 is an example plot of the ripenalty as a function of W2. W e also plot die [constant] ripenalty for the linear system. In this plot, |x = 5 and JJICT= .25.)W e state the variable compensation system as if M choa percentage and state the target-value system as if M choan incremental profit target. P| and P2 are not limited these situations. The firm could choose to implement eithsystem by defining a rating scale such as the five-, sevenor nine-point Likert or semantic-anchor scale. To implemethe variable compensation system, the firm would need transform the rating scale into a percentage and announthis to both R and M. To implement the target-value systemthe firm would need to transform the rating scale into a taget. Naturally, the firm would need to scale the parameteof the reward functions to assure consistent units in the reevant equations, either Equation 7 or Equation 9. To implment the variable compensation system, the upper bound the scale, for example, a 5 on a five-point scale, would corespond to 100%. To implement the target-value system, tfirm would need to select the upper bound (if there weany) such that the constraint is not binding on tbe implieoptimization problems in equations 3 and 4.

P[ implies an evaluation at its upper bound, and implies evaluations, S2, that are greater than TT. W e have spken to many managers who feel that intemal customer satfaction evaluations are inflated. In fact, it is not uncommto see evaluations clustered toward the top of the sca


10/14

Internal Customers and Suppliers 277Menezes (1991) provides a published example in whichXerox has chosen a target (extemal) satisfaction that impliesthat every customer believes the company is at the top of thescale. We have seen many other examples in practice.Indeed, if there is no cost to M of providing a higher evalu-ation for R, then cooperation should be easier and the eval-uations should be more inflated.Many of the intemal customer satisfaction systems thatwe observe have properties similar to the formal models.The linear system is a logical first cut, and the quadratic sys-tem provides an evolution to which managers can move. Inmany observed systems, management begins by giving theevaluating group (M) little or no penalty, W2, for misreport-ing S2. We predict that these firms could improve their sys-tems by making the evaluator's (M's) compensation dependmore steeply on the evaluation.

RELATIONSHIP TO OTHER INCENTIVE SYSTEMSA profit center system also attempts to use the superiorlocal knowledge. In a profit center system M signals its pur-chase of R's technology with the acceptance of a transfer

price, and its perfonnance becomes more dependent on thequality of the inputs. For example, at Chevron, the operatingdivisions can "purchase" projects from the R division (orfrom outside the firm). In both the intemal customer-inter-nal supplier systems and profit center systems, M signals itsuse of an R project with its choice of s; and once M choos-es s, its variable compensation depends on the quality of R'sperfonnance.Another common practice, often used in combinationwith other systems, is management by objectives (MB O). Ina typical MBO system, top management consults with M todevelop a set of objectives for use in subsequent evalua-tions. For example, M might select a sales target of $5 mil-lion and a custom er satisfaction target of 90 % extremely sat-isfied. In the target-value system, M selects a goal, S2.However, S2 itself can be comprised of indicators such as netsales and satisfaction. Compared to an MBO, the target-value system specifies a specific reward function, and thetarget, S2, is used to reward R.Finally, many firms have adopted integrating mechanismsthat enable M and R to communicate on both customerneeds and technological solutions. These systems are com-plementary to intemal customer-intemal supplier systemsnot substitutes.SUMMA RY AND SUGGESTED R ESEARCH

There is considerable pressure to push a marketing orien-tation deep into the organization. In many cases, this meansthat intemal suppliers, such as R&D, see downstreamgroups, such as Marketing, as their intemal customers. In avariety of firms, the intemal customer is asked to evaluatethe intemal supplier, and the intemal supplier is rewardedbased on that evaluation. We have proposed two systemsthat can yield r' and m*. These are certainly not the only sys-tems possible, but they are among the simplest. The target-value system, in particular, should be relatively easy toimplement and, in many cases, will yield a reasonable riskpenalty. Judging by our field experience, the linear systemseems to be the first system that management adopts. Itssimplicity is appealing.

There is certainly room for further research. We use ourtheory to illustrate the properties that s and TT should have.There are important empirical challenges in developing suchmeasures.Our systems may not minimize the risk costs that theemployees must bear (and for which the firm must pay), butthey are simple and perform fairly well for a fairly general

set of functions, IT, C|^, and CR. In other words, our systemsmay not be optimal from the standpoint of minimizing therisk costs. However, if second-best systems are much morecomplex or less robust, then they might impose yet-to-be-identified implementation costs that overwhelm any savingsin risk costs. Implementation costs pose an empirical ques-tion that can only be answered with further research.An altemative research strategy is to establish that somesimple V and w minimize the risk cost for some reasonableprofit and cost functions. Researchers can also study moregeneral v and w, which yield r* and m*. For example, w e canreplace the quadratic function with more-general, asymmet-ric, concave functions of s - -IT, or we can attempt to reducethe risk penalty with higher-order polynomials in s and TT.Another direction of research is to extend the analyses toother error distributions besides normal and other utilityfunctions besides constantly risk averse. Each of our sys-tems allow for cooperation. Researchers also might investi-gate the conditions under which cooperation is an inherentproperty of intemal customer-intemal supplier systems andwhether this is costly to firms.''Our systems assume that the intemal customer and theintemal supplier can cooperate on the (implicit) rating, butfind it more difficult to cooperate on the effort of the inter-nal supplier. Examining and/or relaxing this assumption is

an interesting area for further research.Other areas of research include extensions to longerchains of employee groups, multiple actions by R&D andMarketing, multiple evaluating groups, multiple groupsbeing evaluated, multiple evaluations, and cases in which raffects C[^ or a^ directly. Additional research also mightextend the analyses to other dyadic relationships besidesthose of the internal customer and internal supplier.Empirical research might investigate the practical implica-tions of the systems we analyze and/or the implications ofsuch systems as a means to coordinate Marketing and otherfunctions. Finally, it might elaborate on the formal model toinclude those implementation issues that explain why manyfirms choose an intemal customer-intemal supplier systemas a means to implement a profit center-like approach.'"Hauser, Simester, and Wemerfelt (1996) show that under fairly gener-al conditions g is positive whenever s is not constrained. They also showthat if the firm can choose v and w to implement a set of actions, r" and m"while precluding collusion, then it can choose a different v and w to imple-ment the same actions while allowing collusion. It does this by paying Mmore and R less such that, under collusion, they get just enough to partici-pate. Because the new v and w introduce no new risk costs, the firm earnsthe same profits under the new system as under the old. Wemerfelt,Simester, and Hauser (1996) demonstrate similar results for nonproductivesupervisors and analyze a related case in which collusion is avoided byexpanding Equation 7 to a discrete menu of linear functions.


11/14

278 JOURNAL OF MARKETING RESEARCH, AUGUST 19APPENDIX

Proofs T o Propositions and CorollariesP | (variable outcome-based compensation): For z^ and y,above critical values and for Z3 = 1. the variable outcome-based com pensation system will give incentives to R and Msuch that, acting in their own best interests, they select r'and m*. The critical conditions are Z| > max{TT(r*.O).TT*-2 * *CM } and yi > z, - {-n* - c^* -f froof Define -it = Tr(r.O). Let fl^, be M's certainty equiv-alent (left-hand side of Equation 3) and let OR be R's cer-tainty equivalent (Equation 4). That is.

= Zo + Z| (1 - S |) + Z 3 S| IT-zjsjil O /2 - CM + g-

(Ala)(Alb)and(Ale)

RM

g = ZI S I - Z3 s i TT + Z 3 i l H o ^ / 2 + C M- Zl S | + Z3 SI TT - Z3 S? ^CT 2 - C M

The condition for S| = 0 is that aft^/ as i < 0 when g s 0.(1=0 implies m = 0 by Equation Al.) The condition forS| = 1 is that 3 n|^ /3 s 0.First, note that both conditions hold in the neighborhood ofr*. m*. C ondition 1 holds because Z] > TT by the statement ofthe proposition. C ondition 2 holds because y, > Zj - IT* +Si(jLa-2 for all S| e [0.1] by the statement of the proposition.If conditions 1 and 2 hold, then S] = 1 and S| = 0. and we get(A2a) Q R = y^ +and(A2b) g = z i -

At r = r*. ifi = m*. we have z, > Tr(r.iTi) - Cy^iJ) - |XCT2/2according to the statement of the proposition. Thus, g > 0and the maximization of OR yields r'. m*. This means thatif conditions 1 and 2 hold, which they do for r* and m*. thenR and M choose r' and m*. Furthermore, we can choose y^such that R participates, and we can choose ZQ such that Mparticipates. We need only establish that R never chooses r.and the implied m. for conditions 1 and 2 to be violated.Suppose condition 1 is false, but condition 2 is true. Then.1 = 1. S| = 1. and g = 0. This implies that HR = y^ -H yi -CR(r). But condition 1 being false implies that -z, 4- 'n-(r.O) -lCJ2 > 0. Adding this positive number to HR yields QR < y ,+ yi - C R(r) - z, -I- ^(r.O) - 1X0-2 = y^ + y, - z, - JJLCT2/2 +[Tr(r.O) - C R(r) - CM(O)] - JJ,(J2/2 < y^ + y, - z, - M,a2/2 +[TT(r*.m') - CR(r*) - CM(m*)]. which is what R can get withr* and m*. The last inequ ality is by the definition of the opti-mal. Recall c^^(0) = 0. Thus. R can do better if it chooses anr such that condition 1 is true than it can such that condition1 is false.

Suppose con dition 2 is false. (C ondition 1 can be truefalse.) Then. S| = 0. g = 0. and HR = yg. This is less thancan get at r*. m* (see Equ ation A6 . in which QR > y^ accoing to the definitions of y| and Z] in the statement of tproposition). Thu s. R can do better if condition 2 is true thif condition 2 is false.To summarize. R can choose and prefers to choose r suthat conditions 1 and 2 are true. When these conditions atrue, r = r'. m = m*. and S| = 1.

P2 (target value): For v, = 1 and 0 < W2 < 1/(2|JI,CT2), the targvalue compensation system will give incentives to R and such that, acting in their own best interests, they select r* am*.Proof. We demonstrate'5 that for constantly risk-aveutility and for normally distributed noise, the certainequivalent, or c.e.. of EUj^ (w - c^ + g) is given by the flowing (W2 < [2|jLcr2]-i assures that the logarithm defined):

c.e. = [wo + (l/[2n] )log(l- 2H O^ wi (1 - 2H Cp - W2 )" ' (S2 - TT )2 - CM + g= wo' - W2' (S2 - I T ) ^ - CM + g

Because EU R and EU^, are increasing transformations the certainly equivalents, we evaluate equations 3 and 4 terms of certainly equivalents. Rewriting Equation 3 yield(A3 ) w o ' - W2 ' [s 2 - f f( '" . m) ]^ -C M( m) + g

= w o ' - W 2' [s2 - TKr. m) ]^ -

Because 2 maxim izes the right-hand side of Equation Ait is easy to see that 2 = 'ir(r.ifi) when W2 > 0. Similarly, wshow iti = 0. Thus.

g = W2' [s2 - iT(r. m) f + CM (m)-Because R receives its reward on the basis of 2. which reported prior to market outcomes, TT. there is no risk adjusment for R. Hence, we incorporate M's maximization prolems. Equation A3 . into R's maximization problemEquation 4. by substituting for g. Thus, the maximizatioproblem in Equation 4 becomes

maximize {vo + vis2 -C R ( r ) - W2'[s2 - ir(r. m )f - CM(mThe first-order conditions for m. r. and 2 are giverespectively, by equations A4. A5. and A6. (ir. c^,. and C

are shorthand for ir[r.iTi]. c^^Cm]. and CR[r]. respectively.)(A4)

(A5)(A6)

= 0, ^ OTT a C x- IT ) ^r ^dm om

dr- 2 W 2 ' (S2 - TT ) = 0

"This proof is available from the authors.


12/14

Internal Customers and Suppliers 279

By setting V| = 1, Equation A6 becomes 2w2'[s - ir(r,tTi)] =1. By substituting this relationship in equations A4 and A5,we get the first-order conditions for r* and m*. Finally, weselect the constants, VQ and WQ, such that M's c.e. and R'smaximum are preferred to nonparticipation. The firm partic-ipates if the profit in Equation 5 exceeds the risk penaltyimplied by M's c.e. Note that different V|'s implement dif-ferent actions.

Corollary I (variable outcome-based competisation sensitivity):If the firm makes an error and selects a value, Z]',that is different than Z|, or a value, y,', that is dif-ferent than y|, then thesystem still yields r* and m*as long as z, ' S Z| and yf :Sy,o. The risk penaltyis unaffected.Proof. Notice that P] only requires that Z| and y| be a b o v etheir cr i t ica l va lues . Adjus t ing the coeffic ients , z and y ,assures part ic ipat ion. The r isk penal ty in Equat ion 8 d o e snot depend on Z| and y^Corollary 2 (target value sensitivity): If the firm makes an errorand selects a value, W2', that is different than the

optimal value, W2*, then the target-value compensa-tion system still yields r* and m* as long as 0 < W 2'< \l(2\L(j^). The risk penalty increases, but it can beless than the risk penalty for the variable outcome-based compensation system when JJLCT > I.Proof. Notice that P2 only requires 0 < W2'

internal customers and internal suppliers

Documents