Customer Acquisition and Retention: A Fluid Approach For Staffing

Vol. 30, No. 11, November 2021, pp. 4236–4257 DOI 10.1111/poms.
13520
ISSN 1059-1478|EISSN 1937-5956|21|3011|4236 © 2021 Production and Operations Management Society
Customer Acquisition and Retention:

A Fluid Approach for Staffing
Eugene Furman*
Rotman School of Management, University of Toronto, Toronto, ON M5S 3E6, Canada, eugene.furman@rotman.utoronto.ca
Adam Diamant, Murat Kristal

Schulich School of Business, York university, Toronto, ON M3J 1P3, Canada
e investigate the trade-off between acquisition and retention efforts when customers are sensitive to the quality of
W service they receive, that is, whether they get timely access to a company’s resources when requested. We model
the problem as a multi-class queueing network with new and returning customers, time-dependent arrivals, and abandon-
ment. We derive its fluid approximation; a system of ordinary linear differential equations with continuous, piecewise
smooth, right-hand sides. Based on the fluid model, we propose a novel approach to determine optimal stationary staffing
levels for new and returning customer queues in anticipation of future time-varying dynamics. Using system accessibility
as a proxy for service quality and staffing levels as a proxy for investment, we demonstrate how to apply our approach to
two families of time-varying arrival functions motivated by real-world applications: an advertising campaign and a clini-
cal setting. In a numerical study, we demonstrate that our approach creates staffing policies that maximize throughput
while balancing acquisition and retention efforts more effectively (i.e., equitable abandonment from each customer class)
than commonly used near-stationary methods such as variants of square-root staffing policies. Our model confirms that
acquisition and retention efforts are intimately linked; this has been found in empirical studies but not captured in the
operations literature. We suggest that in time-varying environments, focusing on either alone is not sufficient to maintain
high levels of throughput and service quality.
Key words: acquisition and retention; feedback; queueing networks; time-varying demand
History: Received: February 2018; Accepted: June 2021 by Michael Pinedo, after 2 revisions.
*Corresponding author.
Motivated by the task of allocating resources to

1. Introduction support CRM initiatives, we investigate the opera-
Marketing researchers describe customers as dis- tional trade-off between acquisition and retention
counted streams of revenue and suggest that extend- efforts when customers are sensitive to the quality of
ing a client’s relationship with a firm by focusing on service they receive, that is, whether they get timely
their retention can increase a firm’s profit margin access to a company’s resources when requested. The
(Berger and Nasr 1998, Jain and Singh 2002, Malhotra value of capturing the effect of service quality in
2007). Not surprisingly, as part of various customer retention efforts is alluded to both in the literature
relationship management (CRM) programs, execu- and industry reports. For instance, Keaveney (1995)
tives have looked for better ways to retain clients identifies quality-of-service as one of the major rea-
while simultaneously acquiring new ones (Buttle sons customers seek a different service provider. Sim-
2004, Chen and Popovich 2003). Acquisition and ilarly, Oracle claims that service quality substantially
retention practices are costly, however, and allocating contributes to a customers’ decision to leave (Oracle
enough resources to support them is critical. For 2011). We focus on the relationship between a cus-
example, BCE invests $110 million into acquiring new tomer’s ability to access services and quality. Specifi-
clients, and its investments into customer retention cally, increased access to company resources implies
programs exceed this amount 10-fold (BCE 2016). Fur- higher levels of service quality (Dean 2002, Feinberg
thermore, the CRM industry is large and diverse—it et al. 2002) which, in turn, results in fewer unhappy
includes hotels, insurance companies, health care customers who change providers (Reichheld and Sas-
organizations, and universities (Milovic 2012)—and ser 1990, Reichheld and Schefter 2000). There is also
its value is expected to reach 40 billion dollars in 2020 an inverse effect, customers who find it more difficult
(Columbus 2016, Taylor 2018). to access services will be more likely to leave.
4236
Furman, Diamant, and Kristal: Customer Acquisition and Retention
Production and Operations Management 30(11), pp. 4236–4257, © 2021 Production and Operations Management Society 4237
We model a service system with new and returning allocation decisions in a dynamic environment.
customers as a three-station queueing network with Indeed, we find that a more realistic model yields
feedback and abandonment. After new customers additional managerial insight to what can be obtained
receive service at the first station, they decide whether when demand is assumed to be stationary (e.g.,
to join the customer base (the second station). Return- Afèche et al. 2017, King et al. 2016, Yom-Tov and
ing customers periodically request service (station Mandelbaum 2014).
three) and after each service episode, may choose to To evaluate the effectiveness of our approach, we
remain with the company (returning to the second conduct a simulation study and compare the perfor-
station for a period of time) or defect (leave the sys- mance of our policy to several benchmarks based on
tem). Customers seeking service are impatient and the square root staffing (SRS) rule (see, e.g., Feldman
may decide to leave if they wait sufficiently long. To et al. 2008, Janssen et al. 2011, Liu 2018) as well as a
capture the non-stationary nature of demand, we First-Come-First-Served (FCFS) heuristic that
assume that new customers arrive according to a non- assumes servers are fully flexible. As compared to the
homogeneous Poisson process. The exact analysis of SRS-based polices, we demonstrate that our approach
the stochastic system is intractable, and thus, we recommends staffing levels that better balance acqui-
derive its fluid approximation which results in a sys- sition and retention efforts over a wide-range of sys-
tem of ordinary linear differential equations (ODE) tem parameters and arrival functions. We also find
with continuous, piecewise smooth right-hand sides. that although our objective is to ensure new and
Conforming to our goal of balancing acquisition vs. returning customers are adequately served, overall
retention, we consider the limiting behavior of the throughput does not suffer. As compared to the FCFS
system of ODEs. We formulate a mixed-integer pro- policy, which represents an upper bound on perfor-
gram that determines how much capacity to allocate mance, our policy does not lead to as significant a
to each customer type in order to maximize through- degradation in throughput as compared to the poli-
put while ensuring no customer type is dispropor- cies based on the SRS rule. Finally, we demonstrate
tionally neglected (i.e., abandonment). We focus on that employing time-varying staffing policies when
the development of appropriate stationary staffing demand is time-varying may not always be necessary.
policies in a time-varying environment as frequent This is important as there are many settings where
staffing changes are costly (Batta et al. 2007, Kitaev dynamically reassigning staff is not feasible (Chan
and Serfozo 1999, Koole 1997, Park and Bobrowski and Sarhangian 2018).
1989), can result in added employee stress, and turn- We make several contributions to the literature.
over costs exceeding 20% of an employee’s annual First, our research extends previous work on station-
compensation (Boushey and Glynn 2012). ary staffing policies for queueing systems by intro-
We apply our theoretical results to two practical ducing a new method that better accounts for time-
scenarios: an advertising campaign (Araghi 2014) and dependent demand. Specifically, we show that our
a clinical setting (Yom-Tov and Mandelbaum 2014). policies dominate the near-stationary benchmarks by
The first scenario assumes that the number of new ensuring both new and returning customers receive
customers seeking service increases following a pro- timely service while maintaining similar levels of
motion. Thus, the firm must decide how much atten- throughput. Furthermore, we identify regimes where
tion to reserve for new customers while also ensuring our policy results in higher levels of throughput (e.g.,
that existing ones are not neglected. The second sce- high traffic intensity, long service times, high service
nario assumes that the arrival rate of new customers frequency, low abandonment rate). Thus, our
is periodic, like in the case of patients arriving at a approach performs well without neglecting certain
medical facility. The institution must determine the customer classes. An added benefit is that our staffing
amount of resources to allocate to newly arriving polices can easily accommodate customers with vary-
patients vs. those that need more frequent, sustained ing profit contributions, a common feature in the
service. In both cases, our model incorporates the CRM literature and in many service systems.
time-variability of arrivals and customer abandon- We add to the literature on customer acquisition
ment which represent features that affect the relation- and retention by introducing a model that explicitly
ship between staffing assignments and service quality links these quantities to service quality. Furthermore,
over extended time periods. Although these represent in this more realistic setting, we confirm and refine
different operating regimes, our objective is to present existing results on resource allocation strategies. Simi-
the full generality of our methodology by demonstrat- lar to King et al. (2016), for instance, we find that
ing that it accounts for a wide range of arrival pat- acquisition efforts are vulnerable to diminishing
terns while proposing stationary staffing policies that returns. However, in contrast, dedicating too few
are domain specific. Furthermore, it allows a manager resources to acquisition efforts at any time results in
to better analyze the long-term effects of resource poor performance. In fact, our results suggest that
4238 Production and Operations Management 30(11), pp. 4236–4257, © 2021 Production and Operations Management Society
customer acquisition and retention efforts are inti- that is, Harrison and Zeevi (2004); Bassamboo et al.
mately intertwined, which agrees with the marketing (2006).
literature (see, for instance, Thomas 2001). Whereas The square-root staffing (SRS) rule is a well-known
assigning few staff to returning clients results in a dynamic staffing model in the operations literature
performance decrease, too few resources dedicated to (see, for instance, Borst et al. 2004, Feldman et al.
acquisition efforts undermines retention activities. 2008, Hampshire et al. 2009, Janssen et al. 2011). A
Hence, balancing this trade-off requires careful con- challenge for SRS, as described in Green et al. (1991,
sideration of both acquisition and retention practices 2007), is its application to non-stationary systems.
in a time-varying environment. Furthermore, most applications of the SRS rule
assume that staffing levels can be changed in real-
time, a strong assumption that is typically not true in
2. Literature Review and Contribution practice (e.g., health care, unionized shift work).
Our study contributes to the literature on determinis- Instead, staffing levels are set at asynchronous
tic fluid models for multiserver queueing networks. intervals by management over relatively long time
Specifically, we apply a dynamical system analysis to horizons. In this case, a manager must anticipate the
determine an optimal stationary staffing policy in a non-stationarity of the arrival process. Our model
multi-server queueing network with returning cus- accounts for this and we, to quantify the benefit of our
tomers (feedback), time-dependent arrivals, and approach, compare our policy to several variants of
abandonment. Many authors investigate queueing the SRS policy (Liu 2018).
systems with arrivals that follow a non-homogeneous As discussed, we analyze the dynamical system
Poisson process; see the surveys by Defraeye and Van corresponding to the fluid limit of the original
Nieuwenhuyse (2016), Whitt (2016), and Whitt (2018). stochastic queueing network (Halfin and Whitt 1981).
The problem of staffing in dynamic service systems Foundational results on fluid approximations can be
has also been an active area of study (e.g., Akcali et al. traced back to Mandelbaum et al. (1998), Whitt (2006),
2006, Bhandari et al. 2008, Henderson et al. 1999, Rob- and Kang et al. (2010) for cases with reneging. The
bins and Harrison 2010). Furthermore, several papers precision of fluid approximations are evaluated in
investigate staffing in queueing systems with time- Bassamboo and Randhawa (2010); Daw and Pender
dependent arrival processes and abandonment; (see (2019) while Pender et al. (2017) use the approach to
Bassamboo and Zeevi 2009, Bekker and de Bruin analyze the impact of delay announcements. Unlike
2010, Defraeye and Van Nieuwenhuyse 2013, Har- previous work which grounds the evaluation of sys-
rison and Zeevi 2005, Niyirora and Pender 2016 for tem performance on a set of distributional assump-
example). However, few papers examine stationary tions applied to the fluid model (e.g., Bassamboo and
staffing models in queueing networks with time- Randhawa 2015, Jouini et al. 2013), we focus solely on
varying arrivals and both returning and impatient analyzing the performance of the fluid model.
customers. Our work obtains insights on how to balance cus-
Our research proposes a methodology to assign a tomer acquisition vs. retention in dynamic service
limited number of servers to customers of different environments. These systems are characterized by a
classes given that customers may seek service multi- set of new customers (a priori homogeneous) and
ple times. The dynamic and static assignment of flexi- returning customers (possibly heterogeneous in their
ble server has been investigated by several service requirements). Although several case studies
researchers; see the paper by Andradóttir et al. (2001). and empirical research describe this phenomenon in
The objective in these papers is to maximize the long- the marketing literature (e.g., Reinartz et al. 2005,
run average throughput or minimize the long-run Thomas 2001), there are few papers that model and
average costs. For problems where staff can be allo- analyze the trade-off. Fruchter and Zhang (2004) for-
cated dynamically, several models have been pro- mulate differential game, with two competing firms
posed; single-server queueing systems (Andradóttir and a fixed market, to investigate how effective acqui-
et al. 2003), tandem queues (Andradóttir et al. 2007), sition and retention efforts are at generating sales.
assembly-type queues (Tsai and Argon 2008), and dis- They find that under various conditions, a focus on
crete assignment intervals (Chan and Sarhangian either customer retention or acquisition can be an
2018). Our approach involves solving a static opti- optimal long-run strategy for a firm but do not dis-
mization problem to determine the number of servers cuss how to dynamically balance these efforts over
assigned to each customer class. Although several time. Dong et al. (2011) develop incentive mecha-
papers investigate the assignment of a fixed-set of ser- nisms to determine the sales channel (i.e., direct sell-
vers (e.g., Futamura 2000, Hillier and So 1989, Lee et ing vs. delegation) for customer retention and
al. 2014, Smith and Barnes 2015, Smith et al. 2010), acquisition. They find that when acquisition and
few pursue this objective for time-varying systems, retention efforts are in conflict, a firm should choose
to focus on retention. King et al. (2016) introduce a process with time-varying intensity λ ≡ λ(t). Class-b
discrete-time, deterministic dynamic programming clients are customers that have received service some-
model to optimally allocate the resources of a profit time in the past and may require service again.
maximizing firm. They derive an optimal investment Class-i ∈ I clients have service requirements that are
policy which indicates that a firm should shift its exponentially distributed with rate μi . There is also a
focus from acquisition to retention as the size of a fixed-sized pool of s > 0 staff that are able to serve
firm’s customer base grows over time. Finally, Afèche any class of customer and a decision maker that must
et al. (2017) present a static optimization model to decide how many staff to dedicate to each customer
find a set of optimal staffing levels when customers class. We assume that the staffing policy is stationary,
differ in their service request rates, have rewards that that is, once decided it cannot be easily changed. Let
may depend on service, and have different return sa and sb be the staffing assignment to new (Qa ) and
probabilities. They find that an optimal policy has a returning (Qb ) service stations, respectively, such that
“bang-bang” structure—certain customers get service sa þ sb ≤ s. Servers attend to customers of an assigned
or not at all. We extend this literature by incorporat- class under the standard first-in, first-out (FIFO) ser-
ing time-varying demand and abandonment. Further- vice policy.
more, the trade-off between whether to invest in We account for the revenue differential, or the
customer acquisition and retention efforts is modeled penalty of customer abandonment, by assigning
directly into the dynamics of the problem as the abil- weights ρa and ρb to new and base customers
ity of a firm to allocate enough resources to satisfy attempting to receive service in accordance with
time-varying demand. the instantaneous value of serving that customer.
The customer retention and acquisition literature is That is, we define ρi to be relative value of a class-
related to the analysis of service systems with feed- i client seeking service where, without loss of gen-
back (de Véricourt and Jennings 2011); Jacobson et al. erality, ρb ¼ 1 and ρa ∈  ≥ 0 . The inclusion of these
(2012) model an emergency response environment weights links our queueing model to the CRM lit-
using a closed-queueing network model with feed- erature where balancing acquisition and retention
back and a fixed number of customers who may leave efforts is achieved by maximizing the long-term
the system. Ding et al. (2015) analyze a call center profit of the firm.
with unspecialized servers, retrials, and reconnects. After a class-a customer arrives to the system,
Yom-Tov and Mandelbaum (2014) and Huang et al. she is served immediately if there is an available
(2015) propose a SRS policy to stabilize the perfor- server at station Qa , and waits for service other-
mance of a time-varying network with returning cus- wise. All customers are impatient and may decide
tomers without abandonment. Liu and Whitt (2017) to abandon the queue before receiving service. We
extend this work by developing an offered-load assume abandonment times of class-a customers
approximation in a time-varying, many-server queue- are exponentially distributed with rate τa . Then,
ing system with customer abandonment and return- class-a clients who complete their service either
ing customers. Chan et al. (2014) present an analysis leave the system or join the customer base (which
of a piecewise smooth dynamical system with discon- we model as a station Qc with infinite capacity);
tinuous right-hand sides where the service rate is they become a class-b client. If a client joins the
dependent on the number of customers in the system. customer base, she may again seek service, this
Our study takes an alternative approach: we deter- time at station Qb , after an exponentially dis-
mine an optimal stationary staffing level in anticipa- tributed amount of time with mean 1/r. Customers
tion of time-varying dynamics. Furthermore, we may also decide to cease their relationship with the
demonstrate that our asymptotic analysis and static service provider; attrition times from station Qc are
optimization model can be adapted to a variety of ser- exponentially distributed with rate ζ. Class-b clients
vice settings. that seek service at station Qb are, again, impatient;
abandonment times are exponentially distributed
with rate τb . Class-b customers who abandon the
3. Model Formulation queue rejoin the customer base (i.e., Qc ) with prob-
In this section, we introduce a stochastic queueing ability θc or leave the system.
network to model a firm’s ability to acquire and retain Let QN ≡ ðM, I, PÞ define the topology of the
customers, derive its fluid limit, and formulate the queueing network where M ¼ fQa , Qb , Qc g is the set
corresponding dynamical system. of stations in the network, I is the set of customer
Let I ≡ fa, bg be a set of customer classes that corre- classes, and P : ðM MÞ ! ½0, 1 is a function defin-
spond to new and returning clients, respectively. ing the routing probabilities among the stations for
Class-a clients represent new customers that arrive to clients who complete their service requirements.
the system according to a non-homogeneous Poisson More specifically,
0 1 0 1
aa ab ac 0 0 θac The stochastic processes in (2) describes the evolu-
B C B C tion of the system. New customers that finish ser-
P : ðM MÞ ! @ ba bb bc A ¼ @ 0 0 θbc A,
vice at Qa may join the customer base, Qc , or they
ca cb cc 0 1 0 may leave the system entirely. Returning customers
leave Qc to seek service at Qb or may exit the sys-
where  j j0 is the probability that a customer who fin- tem. After a returning customer receives service at
ishes service at station Q j is routed to station Q j0 for Qb , she may either leave the system or rejoin the
j, j0 ∈ fa, b, cg; θac is the probability that a new cus- customer base Qc . Thus, the non-stationary arrival
tomer joins the customer base and θbc is the probabil- rate of new customers has a knock-on effect where
ity that a returning customer rejoins the customer the arrival, departure, and abandonment processes
base. We refer to a queueing network with the speci- from all stations in the network QN are non-
fied arrival, service, and abandonment rates by its stationary.
network topology QN. Let W ¼ fW a ðtÞ, W b ðtÞ,
W c ðtÞg be a set of headcount stochastic processes 3.2. Fluid Limit and Dynamical System
corresponding to the number of customers waiting A cornerstone of constructing the fluid (limiting)
for service in Q j where j ∈ {a, b, c}. Let X ¼ fXa ðtÞ, approximation of the stochastic equations in (2) is to
Xb ðtÞ, Xc ðtÞg be a set of headcount stochastic pro- describe the asymptotic dynamics of a corresponding
cesses corresponding to the number of busy servers large (scaled-up) queueing system. We employ the
at each station. The system states evolve according to heavy traffic limit theorems described in Halfin and
the following equations. Whitt (1981) and Mandelbaum et al. (1998). The pro-
cedure is used pervasively in the literature (see, for
Qa ðtÞ ¼ W a ðtÞ þ Xa ðtÞ, W a ðtÞ ¼ ðQa ðtÞ sa Þþ , Xa ðtÞ ¼ ðQa ðtÞ^sa Þ; example, Borovkov 1967, Iglehart 1965, Iglehart
Qb ðtÞ ¼ W b ðtÞ þ Xb ðtÞ, W b ðtÞ ¼ ðQb ðtÞ sb Þþ , Xb ðtÞ ¼ ðQb ðtÞ^sb Þ; 1973a,b). Specifically, we consider a family of queue-
Qc ðtÞ ¼ Xc ðtÞ: ing networks parametrized by η such that the arrival
(1) rate and the number of servers of system QNη are
scaled up by a factor η while the traffic intensity is
With a slight abuse of notation, we denote stations held constant. We also scale up the processes in Q
by Q j and the corresponding headcount stochastic and the initial conditions in Q0 . This transforms (2)
processes by Q j ðtÞ. The states of QN are described into
by the stochastic vector Q ¼ ðQa ðtÞ, Qb ðtÞ, R R
Qc ðtÞÞT ∈ 3≥ 0 . The dynamics of the system is pre- Qηa ðtÞ ¼ Qηa ðt0 Þ þ N aλ t0 ηλðuÞdu N aμ t0 μa Qηa ðuÞ^ηsa du
t t
sented in Figure 1. R þ
N aτ t0 τa Qηa ðuÞ ηsa du ,
t
R R
Qηb ðtÞ ¼ Qηb ðt0 Þ þ N bλ t0 rQηc ðuÞdu N bμ t0 μb Qηb ðuÞ^ηsb du
t t
3.1. Stochastic Queueing Model
j R þ
Let N λ , N μj and N τj be the standard Poisson arrival, N bτ t0 τb Qηb ðuÞ ηsb du ,
t
departure, and abandonment processes with time- R R

Qηc ðtÞ ¼ Qηc ðt0 Þ þ N aμ t0 θac μa Qηa ðuÞ ^ ηsa du þ N bμ t0 θbc μb Qηb ðuÞ ^ηsb du
t t
varying intensities for station j, respectively, and let R þ R R
Q0 ¼ ðQa ðt0 Þ, Qb ðt0 Þ, Qc ðt0 ÞÞT be a vector of initial þN bτ t0 θc τb Qηb ðuÞ ηsb du N cμ t0 rQηc ðuÞdu N cτ t0 ζQηc ðuÞdu :
t t t
conditions, that is, the number of customers at station (3)

j at some starting time t0 . Due to the distributional
assumptions introduced in Section 3, the system can The relations in (3) describe a sequence of queueing
be modeled as three non-stationary, Erlang-A queues. systems where the number of servers and the num-
This, as shown in Mandelbaum et al. (1998), allows us ber of arrivals grow as η → ∞. The compact conver-
to express the stochastic vector Q in the following gence of (3) to its fluid limit is established in
functional form. Mandelbaum et al. (1998) and we provide the result,
without proof, in the following lemma.
R R
t t
Qa ðtÞ ¼ Qa ðt0 Þ þ N aλ t0 λðuÞdu N aμ t0 μa ðQa ðuÞ^sa Þdu LEMMA 1. Let Qη ðtÞ be a scaled up queueing process
R
N aτ t0 τa ðQa ðuÞ sa Þþ du ,
t
parametrized by η that corresponds to the original queue-
R
t
R
t
ing process Q(t) for all t ≥ 0. Then, we have
Qb ðtÞ ¼ Qb ðt0 Þ þ N bλ t0 rQc ðuÞdu N bμ t0 μb ðQb ðuÞ^sb Þdu
R
N bτ t0 τb ðQb ðuÞ sb Þþ du ,
t 1
qðtÞ≡ lim sup Qη ðtÞ a:s:, (4)
t∈ T η
R R η!∞
t t
Qc ðtÞ ¼ Qc ðt0 Þ þ N aμ t0 θac μa ðQa ðuÞ^sa Þdu þ N bμ t0 θbc μb ðQb ðuÞ^sb Þdu
R R R
þ N bτ t0 θc τb ðQb ðuÞ sb Þþ du N cμ t0 rQc ðuÞdu N cτ t0 ζQc ðuÞdu :
t t t
where T⊂  ≥ 0 is a compact set such that q(t) solves
(2) the integral equation
Figure 1 Dynamics of Service System
Z t where f : 3 ! 3 is a continuous, vector-valued,

þ
qðtÞ ¼ qðt0 Þ þ λðuÞ μðuÞðqðuÞ^sÞ τðuÞðqðuÞ sÞ du: piecewise smooth function. Thus, q(t) is differen-
t0
tiable for any t ≥ 0. Relation (7) with initial condi-
(5) tion q0 ¼ ðqa ðt0 Þ, qb ðt0 Þ, qc ðt0 ÞÞT defines the initial
value problem (IVP)
Process Qη ðtÞ is said to converge compactly to q(t), a
0 1 0 1
fluid approximation of Q(t). q_ a ðtÞ λðtÞ μa qa ðtÞ^sa τa ðqa ðtÞ sa Þþ
B C B C
B_ C B C
B qb ðtÞ C ¼ B rqc ðtÞ μb qb ðtÞ^sb τb ðqb ðtÞ sb Þþ C
Consider the queueing network QN whose dynam- @ A @

A
q_ c ðtÞ θac μa qa ðtÞ ^sa þ θbc μb ðqb ðtÞ^ sb Þ þ θc τb ðqb ðtÞ sb Þþ ðr þ ζÞqc ðtÞ,
ics are governed by the stochastic functional equations
in (2). If q ¼ ðqa ðtÞ, qb ðtÞ, qc ðtÞÞT , then by evaluating a q0 ¼ ðqa ðt0 Þ, qb ðt0 Þ, qc ðt0 ÞÞT :
limit of the sequence of queueing networks QNη as (8)
η → ∞ and applying Lemma 1, we have that

Rt Rt Rt þ This IVP is a piecewise smooth system of ODEs
qa ðtÞ ¼ qa ðt0 Þ þ t0 λðuÞdu t0 μa qa ðuÞ^ sa du t0 τ a qa ðuÞ sa du,
Rt Rt Rt þ with continuous right-hand sides.
qb ðtÞ ¼ qb ðt0 Þ þ t0 rqc ðuÞdu t0 μb qb ðuÞ^ sb du t0 τ b qb ðuÞ sb du,
Rt h þ i LEMMA 2. A solution to (8) exists, is unique, and is pie-
qc ðtÞ ¼ qc ðt0 Þ þ t0 θac μa qa ðuÞ^sa þ θbc μb qb ðuÞ^ sb þ θc τb qb ðuÞ sb du
Rt Rt cewise smooth.
t0 rqc ðuÞdu t0 ζqc ðuÞdu:
(6) Lemma 2 demonstrates that for any smooth time-
which implies that q is the fluid approximation of Q varying arrival function, the IVP in (8) has a unique
component-wise. solution and the solution is piecewise smooth with a
The right-hand sides of (6) are piecewise smooth continuous right-hand side. Nevertheless, our analy-
functions, and if we differentiate (6) piecewise with sis of staffing rests on a stronger assumption, namely,
respect to t, the following relation is obtained that the dynamical system is stable, that is, converges
to a limit as t → ∞. The stability is especially impor-
_ ¼ fðqðtÞÞ,
qðtÞ (7) tant for systems with monotonous arrival functions.
The limit (equilibrium point) and the corresponding
speed of convergence provide insight into the beha- the first coordinates of the eigenvectors correspond-
vior of the solution for small values of t. The next ing to ψ and χ. Finally, qi ðtÞ for i ∈ {b, c} is the non-
result states a necessary and sufficient condition for homogeneous term, which converges to a stationary
the stability of (8). solution as t → ∞. Given a description of the system
state q(t), at every point in time, we can calculate
LEMMA 3. If λðtÞ ∈1 is a smooth function and the value of the customers in the system using the
lim λðtÞ < ∞, (8) is stable. relation ρq(t) where ρ ¼ ðρa , ρb , ρb ÞT . That is, the
t!∞
value generated at each station, at each moment in
Although stable systems with continuous monoto- time, is the expected profit associated with serving
nous arrival functions (e.g., exponential) are mathe- each customer class multiplied by number of cus-
matically convenient, there are many practical tomers in each station.
instances where the conditions of Lemma 3 are too
strict. For example, we consider a particular instance 3.3. Model Discussion
from the family of periodic arrival functions. To Our model captures the trade-off between customer
address this, we define the concept of boundedness acquisition and retention. Specifically, the number of
for any initial condition. Specifically, a solution to (8) customers that abandon and the waiting time at each
is bounded if there exists a K > 0 such that q(t) < K station are functions of the arrival rate and the staffing
for any t > 0. Notice that all vector inequalities and policy. Determining an optimal staffing assignment,
products are defined component-wise. then, involves balancing the length of idle and busy
periods for each customer class while also consider-
LEMMA 4. Suppose lim λðtÞ is undefined for λðtÞ ∈ 1 ing their relative importance or profitability. More
t!∞ servers assigned to a customer class implies shorter
but there exists a K0 > 0 such that λðtÞ < K0 for any busy periods in the stochastic regime. Conversely,
t > 0. Then, q(t) is bounded. having fewer severs results in longer busy periods
with higher numbers of customers who abandon.
Because the solution to (8) defines a piecewise Thus, any staffing decision must balance short, new
smooth, continuous, dynamical system in 3≥ 0 , customer waiting times (i.e., better quality-of-service)
qa ðtÞ ¼ sa and qb ðtÞ ¼ sb describe planes (i.e., the with a focus on retention and an increased access to
boundary Σ) that split the state space into four company resources. Our model explicitly incorpo-
subspaces, Si , i = {1, 2, 3, 4}. The right-hand sides rates the effects of service quality by connecting it to
of (8) change each time the solution crosses the abandonment, a relationship that has not been
boundary. We rigorously define the boundary, as included in the customer retention literature, while
well as the subsets Si , in Appendix A. Neverthe- also including time-varying demand and feedback.
less, because our system has continuous right-hand Analyzing the behavior of new and returning cus-
sides, solutions cross the boundaries and immedi- tomers when demand is time-varying is cumbersome
ately enter into adjacent regions, that is, the in the stochastic regime. Instead, we study the sys-
amount of time spent on the boundary has mea- tem’s fluid limiting behavior. This simplifies the anal-
sure zero. This is in contrast to systems with dis- ysis of the dynamics, allows us to capture both
continuous right-hand sides (e.g., Chan et al. 2014), overloaded and underloaded regimes, and ensures
where solutions can remain on the boundary for that optimal control decisions (e.g., staffing) are tract-
extended periods of time. able and implementable. Since a fluid approach
For either stable or bounded systems, as character- assumes a heavy traffic environment, classical appli-
ized by Lemmas 3 and 4, the queue for new customers cation of these results is confined to settings with a
is described by a non-homogeneous ODE and the relatively large number of arrivals and servers. How-
complexity of its solution depends on the form of the ever, our simulations suggest that our approach does
arrival function. The dynamics governing the queue consistently well in environments with a small num-
of returning clients and the customer base requires ber of servers and steadily improves as the number of
solving a system of differential equations which take servers increases.
the form Our work is grounded in classical call center litera-
qb ðtÞ ¼ ueψt vðψÞ þ weχt vðχÞ þ qb ðtÞ, ture and we inherit many of its structural and proba-
(9)
qc ðtÞ ¼ ueψt þ weχt þ qc ðtÞ, bilistic assumptions. For example, we assume that
new customers arrive according to a non-stationary
where w and u are constants derived from initial Poisson process. Although this may be not an accu-
conditions, ψ < 0 and χ < 0 are the smallest and lar- rate reflection of reality in all settings, the approxima-
gest negative eigenvalues associated with the homo- tion does improve as the system size increases. This
geneous solution, respectively, and v(ψ) and v(χ) are also conforms to the heavy traffic assumption used to
derive the fluid approximation of the stochastic sys- function of the queue they enter and not an attribute
tem. Second, we assume that the service duration and of the customer class they identify with. The routing
time-to-abandonment are exponentially distributed. probabilities after a service completion follow a simi-
Although other more sophisticated distributions can lar relationship.
be used, our focus is on how a firm balances acquisi-
tion and retention efforts when demand is non-
stationary. Using more realistic distributions would
4. Staffing Analysis
increase model complexity and may obfuscate man- In this section, we consider two families of arrival
agerial insight. Third, we assume routing probabili- functions to better understand the trade-off between
ties are independent of a customer’s service acquisition and retention. Theoretically, we capture
experience. Fourth, following Hu et al. (2016), we con- two broad types of asymptotic behavior: convergence
sider congestion and fitness abandonment. Customers to a limiting point and convergence to a periodic func-
who wait sufficiently long to obtain service may aban- tion with an undefined limit. Practically, the first fam-
don the system due to congestion (τi ). This type of ily is motivated by an advertising campaign, that is,
abandonment is not related to the quality of service an event which results in an immediate boost of new
they receive and is dependent only on system accessi- customers followed by a gradual decay (e.g., Araghi
bility. Those who experience poor service quality may 2014). The second addresses environments with a
leave the system after a service completion (fitness periodic pattern of arrivals, such as a clinical setting
abandonment). We assume θbc > θc as customers are (e.g., Yom-Tov and Mandelbaum 2014). In both cases,
more likely join the customer base after successfully we study how resources should be allocated given the
completing service. firm’s objective is to appropriately serve their cus-
We assume that a manager does not dynamically tomers, that is, provide timely access to their
adjust staffing. In these settings, it may not be feasible resources when requested. To this end, we introduce
to quickly change s. As noted, regularly hiring/firing a general four-stage modeling approach that attempts
workers is expensive (e.g., Del Boca and Rota 1998, to maximize throughput while ensuring no customer
Moon et al. 2020). Furthermore, although temporary type is disproportionally neglected (i.e., large number
workers maintain staffing flexibility, there are many of abandonments).
documented downsides such as a drop in overall pro-
1. Partition the domain into managerially relevant
ductivity (Rodrı́guez-Gutiérrez 2007) and the emer-
regions.
gence of safety and/or quality issues (Lu and Lu
2. In each region, analyze the limiting behavior of
2017). A manager may also decide against regularly
the dynamical system as t → ∞.
re-assigning staff (i.e., from sa to sb and vice versa) as
Systems that converge to limiting points: Establish
frequent switching has been shown to increase service
an ordering relationship among the points to
times, reduce performance quality, and exacerbate
determine which piecewise smooth region is
workplace stress (Duan et al. 2020, Rogers and Mon-
most preferable.
sell 1995, Rubinstein et al. 2001). Indeed, there are
Systems that converge to periodic functions: Find
many documented examples in banking, computing,
the extrema of the periodic orbits and deter-
and health care where experience and specialization
mine the times when they cross the boundaries
result in productivity gains (e.g., Madiedo et al. 2020).
of the piecewise smooth regions.
In addition, even flexible employees may need to re-
3. Investigate the behavior of the system at finite
train in order to refresh their skills before being
times (state of disorder).
deployed to another service. In call centers, for
4. Formulate a mixed-integer program where the
instance, employees need regular training sessions to
objective represents the sum of limiting points
remain effective and to update them on product
(stage 2) weighted by a convex combination of
changes and new company processes. In a typical call
the results from stage 3. The optimal solution of
center, experienced agents receive between 6 and 15
the optimization model is a stationary staffing
days of training per year on average (Holman et al.
policy that accounts for the resource capacity of
2007).
the system and its time-varying dynamics.
For simplicity, we assume two classes of cus-
tomers, that is, new and returning clients. Our
model, however, can be naturally extended to 4.1. Advertising Campaign: Exponential Decay
include several classes of returning clients with a We consider a new advertising campaign launched
dedicated queue for each additional base customer by a service provider. At the start of the campaign the
class. Servers can also be assigned to subsets of cli- number of customers that request the firm’s services
ents. Nevertheless, our model assumes that once cus- rises to its maximum. The increased intensity of arri-
tomers seek service, their rate of abandonment is a vals causes an expansion of the customer base,
provided the firm’s ability to retain clients is not intensity that jumps upward after each arrival and
somehow compromised. As the effect of the campaign decays exponentially to a baseline level (Hampshire
diminishes, the initial spike in the acquisition of new et al. 2020). Mathematically, an exponentially decay-
customers fades and the arrival rate returns to pre- ing arrival function assures tractability of the time-
advertisement (or more stable) levels. To capture this varying dynamical system in (8).
behavior, we model the intensity of the arrival process Substituting (10) into (8), we obtain the following
of new customers as a decaying exponential function IVP. By Proposition 2 and Lemma 3, the solution of
of the form (11) is piecewise smooth, and its form changes as the
system transitions from one smooth region to another.
λðtÞ ≡ λ1 eδt þ λ0 , (10) Furthermore, in each region of the domain (there are
0 1 0 1
q_ a ðtÞ λ1 eδt þ λ0 μa ðqa ðtÞ ^ sa Þ τa ðqa ðtÞ sa Þþ
B C B C
B q_ ðtÞ C ¼ B rqc ðtÞ μb ðqb ðtÞ^ sb Þ τb ðqb ðtÞ sb Þþ C
@ b A @ A (11)
þ
_qc ðtÞ θac ðqa ðtÞ^sa Þμa þ θbc ðqb ðtÞ^ sb Þμb þ θτc τb ðqb ðtÞ sb Þ ðr þ ζÞqc ðtÞ,
q0 ¼ ðqa ðt0 Þ, qb ðt0 Þ, qc ðt0 ÞÞT :
where λ1 > 0 is the intensity of the campaign, 4 such regions), the queue for base customers in
δ ∈ (0, 1) controls the steepness of the exponential (11) admits a closed-form solution as in (9), and
curve (i.e., the rate of decay), and λ0 is the intensity the equation for new customers has a closed-form
of the pre-advertisement arrival process. The arrival solution obtained by standard methods. We argue
function in (10) has three intuitive properties: it that characterizing the behavior of the system in
achieves the global maximum at λ1 þ λ0 , it is mono- only a few regions is sufficient for developing a
tonously decreasing for all t > 0, and lim λðtÞ ¼ λ0 . systematic approach to stationary staffing in a
t!∞
Representing the arrival rate of new customers in time-varying environment. To this end, we parti-
an advertising campaign as an exponentially decay- tion the domain into three regions characterized
ing function is a convenient choice both practically by distinct modes of operation. The first mode has
and mathematically. In Blattberg and Deighton high rates of customer acquisition, which in turn,
(1996), for instance, a similar function is used to pre- lead to high rates of customer retention. The abil-
dict the number of customers acquired over a time- ity to acquire new customers decreases in the sec-
horizon. Furthermore, this choice remains in line ond mode (as compared to the start of the
with the marketing literature, which models adver- campaign), which results in a slower growth/de-
tising elasticity (a measure of an advertising cam- cline of the customer base. In the third mode, the
paign’s effectiveness in generating new sales) as an ability of the firm to acquire new customers
exponential function of time (e.g., Dant and Berger returns to pre-advertisement levels and the cus-
1996, Parsons 1975). Similar considerations are tomer base declines until it reaches a steady-state.
employed in economics where demand is generally Mathematically, we have
assumed to be an exponential function of price (see
Mode 1 - Launch Region: qa ðtÞ > sa and qb ðtÞ≠ sb ,
Thompson and Teng 1984, for instance) and in the
Mode 2 - Loyalty Region: qa ðtÞ < sa and qb ðtÞ > sb ,
health care operations literature where, at emergency
Mode 3 - Lessening Region: qa ðtÞ < sa and qb ðtÞ < sb .
call centers, there may be a burst of incoming traffic
due to emergency events such as a fire or power out- Given the partition of the domain into modes of
age (L’Ecuyer et al. 2018). Furthermore, in situations operation, our goal is to investigate how a firm
where there are clusters of arrivals, a Hawkes pro- can remain in the launch and loyalty regions as
cess—where an arrival of a new customer increases long as possible. This is appropriate because, in
the likelihood that another will arrive soon after— the lessening region, some servers are idle indefi-
can capture decaying dynamics by assuming an nitely, which implies that a new staffing assign-
exponential kernel function (Daw and Pender 2018, ment may be more appropriate. Thus, in the
Daw et al. 2020). Generally speaking, under this second stage of our approach, we examine the
assumption, the Hawkes process has an arrival limiting behavior of (11) and focus specifically on
the regions of interest, that is, the launch and loy- towards is larger than the corresponding limiting
alty regions. point when q(t) is in the loyalty region. Furthermore,
when q(t) transitions from the launch region to the
PROPOSITION 1 (Limiting Points). The system state loyalty region, qc ðtÞ is moving to a larger limiting
qðtÞ ¼ ðqa ðtÞ, qb ðtÞ, qc ðtÞÞT is attracted by point than a system in which q(t) never reached the
8
>
> λ0 þ sa ðτa μa Þ rθac sa μa þ sb ðτb μb Þðr þ ζÞ sb rðθτc τb θbc μb Þ θac sa μa þ sb μb ðθbc θc Þ
>
< , , , qb ðtÞ ≥ sb ,
τa τb ðrð1 θτc Þ þ ζÞ rð1 θc Þ þ ζ
ðqa1 , qb1 , qc1 Þ :¼
>
> λ0 þ sa ðτa μa Þ rθac sa μa θac sa μa
>
: , , , qb ðtÞ< sb ,
τa μb ðrð1 θbc Þ þ ζÞ rð1 θbc Þ þ ζ
the equilibrium solution in the launch region (i.e., the launch region. As a result, for long time horizons,
first mode of operation), and by QN1 is more managerially preferable than QN2 .

λ0 rθac qa2 ðt0 Þμa þ sb ðτb μb Þðr þ ζÞ sb rðθτc τb θbc μb Þ θac qa2 ðt0 Þμa þ sb μb ðθbc θc Þ
ðqa2 , qb2 , qc2 Þ :¼ , , ,
μa τb ðrð1 θτc Þ þ ζÞ rð1 θc Þ þ ζ
the limiting point in the loyalty region (i.e., the second Although Lemma 5 establishes an asymptotic
mode of operation) early in the time horizon. ordering of the customer base between modes of
operation, it does not imply that this relationship
Proposition 1 demonstrates that within each mode holds for any time t. In particular, at finite times
of operation, limiting points exist and can be written the dynamics of the system are governed by two
in closed-form. We note that although qb1 is piecewise competing exponential functions, as described in
defined, it does not affect our subsequent analysis. (9). Monotonicity guarantees are difficult to estab-
We select qc1 corresponding to the case of qb ðtÞ ≥ sb lish during this period, and we define q(t) to be in
because it describes the busiest time of operation in a state of disorder. As both exponential terms
the launch phase. If q(t) remained in a particular approach zero, the solution becomes identical to its
mode of operation, it would move towards the corre- non-homogeneous value. Thus, the third stage of
sponding limiting point. However, due to the time- our approach describes the duration of the disorder
varying nature of demand, q(t) may spend time in period and, again, focuses specifically on the launch
multiple modes of operation. As a result, we next and loyalty regions.
establish an ordering relationship amongst the limit-
ing points to provide a criterion for determining PROPOSITION 2 (Duration of Disorder Period). Let ψ k
which mode of operation is better. In particular, we be the smallest and χ k be the largest negative eigenvalues
focus on the size of the customer base as t → ∞ as this of the homogeneous system of ODEs corresponding to
represents the steady-state number of customers that (11) in mode k ∈ {1, 2}. If vðψ k Þ and vðχ k Þ represent the
the firm retains over the long term. first coordinates of the corresponding eigenvectors, then
the duration of the disorder period of qck ðtÞ is determined
LEMMA 5 (Asymptotic Monotonicity). Let QN1 , by the time required for the constant w k ¼ jwk jeχ k t0 to
QN2 and QN3 be queueing systems with identical exponentially reduce to 0 with a given degree of toler-
parameters but different initial conditions, that is, they ance, where
originate in the launch, loyalty, and lessening regions,
respectively; and neither system transitions into a higher vðψ k Þ½qc ðt0 Þ qck ðt0 Þ qb ðt0 Þ þ qbk ðt0 Þ
mode of operation. Then, for any combination of queueing wk ¼ for k ∈ f1, 2g,
eχ k t0 ½vðψ k Þ vðχ k Þ
parameters,
and t0 is the time the system state switches into operating
qc1 > qc2 > qc3 : mode k.
Lemma 5 demonstrates that when q(t) is in the COROLLARY 1. In the launch region, if qa1 ðtÞ ! qa1 , then
launch region, the limiting point that qc ðtÞ is moving 0 ¼ jw0 jeτa t0 ! 0 where
w
þ
λ1 δt0 λ0 þ sa ðτa μa Þ τa t0 ζ
w0 ¼ qa ðt0 Þ e e : sb ≥ 1 θbc sa , (14)
τa δ τa r
0 ðsa Þ ≥ 0,
w (15)
For the final stage of our approach, we combine the
above results to formulate the objective function for
our stationary staffing policy. This considers both the k ðsa , sb Þ ≥ 0, k∈ f1, 2g:
w (16)
limiting points and the duration of disorder results.
Our goal is to formulate an optimization problem to We note that the w k ðÞ terms implicitly depend on
determine sa and sb so as to maximize the time spent the staffing level through the initial conditions of
in the launch and loyalty regions. From Lemma 5, for the dynamical system. We make this relationship
sufficiently large t, it suffices to determine values sa explicit by denoting them as functions of sa and sb .
and sb such that the limiting points are as large as pos- As a result, EXP cannot be solved by an off-the-shelf
sible. Which points to choose follow from the modes solver tool. Nevertheless, in Section 4.3, we prove
of operation. For the launch region, we maximize qa1 several structural properties that lead to an efficient
and qc1 as the advertising campaign has just begun solution procedure.
and the focus is on the acquisition of new customers. Constraint (13) bounds the number of staff that can
For the loyalty region, we maximize qb2 as, at this be assigned to each customer class. Constraint (14)
point, the focus of the firm has shifted to retention defines a lower bound on the staffing level for return-
efforts because the rate of new arrivals has dwindled. ing customers. When the attrition rate from the orbit
As per Proposition 1, the limiting points qa1 , qc1 , and is lower than the rate at which base customers request
qb2 , are linear in the decision variables sa and sb . To service (ζ ≤ r), the firm benefits by sufficiently staffing
combine them, we apply Proposition 2. That is, we station Qb . The benefit can be quantified; it is equal to
multiply each limiting point by a convex combination the value of having a server dedicated to new clients
of the magnitude of the exponential terms governing weighted by the probability that a base customer
the length of the disorder period. Larger absolute val- returns to the orbit after service, that is, θbc . When the
ues of the exponential terms, (i.e., w 0, w
1, w
2 ) indicate attrition rate from the orbit is higher than the rate at
that dynamics of the corresponding queues are more which base customers request service (ζ > r) the value
time-varying in the vicinity of the initial conditions. of assigning a large number of servers sb to station Qb
To ensure that lengthy queues do not form over these diminishes. Finally, constraints (15) and (16) ensure
periods, we give more weight to state functions with that the system reaches the launch and loyalty
greater time-variability. As a result, with a slight regimes, respectively.
abuse of notation, we write down the objective of our
optimization problem showing the dependency of 4.2. Clinical Setting: Sinusoidal
each of the terms on the decision variables as follows: Assuming that patients arrive uniformly over time
w 0 ðsa Þ
is a severe limitation in the modeling of health care
zðsa , sb Þ ¼ ρ q ðs Þ systems. For instance, across a single day, many
0 ðsa Þ þ w
w 1 ðsa , sb Þ þ w 2 ðsa , sb Þjvðχ 2 Þj a a1 a
authors have found that the arrival rate in emer-
w 1 ðsa , sb Þ
þ ρ q ðs , s Þ gency wards is periodic, see for example, De Bruin
0 ðsa Þ þ w
w 1 ðsa , sb Þ þ w 2 ðsa , sb Þjvðχ 2 Þj b c1 a b
et al. (2007) and Green et al. (2007), (1991). model
w 2 ðsa , sb Þjvðχ 2 Þj this periodicity by assuming that the arrival rate
þ ρ q ðs Þ: (12)
0 ðsa Þ þ w
w 2 ðsa , sb Þjvðχ 2 Þj b b2 b
1 ðsa , sb Þ þ w
follows a sine function. This accounts for the cycli-
We note that the correction factor vðχ 2 Þ that multiplies cal pattern of demand in their study, however, its
2 comes from the general solution of the homoge-
w flexibility in mathematical modeling is particularly
neous system of ODEs. Furthermore, ρ is the relative attractive as a linear combination of sine functions
value associated with each customer class. Also, max- can be used to approximate any periodic and time-
imizing zðsa , sb Þ ensures that the service system varying arrival process (Yom-Tov and Mandelbaum
spends the least amount of time in the lessening 2014). Systems with periodic arrivals have also
phase, which is the least managerially desirable. The been analyzed by several researchers in the opera-
mixed-integer optimization problem is given by tions community (e.g., Liu and Whitt 2017, Whitt
2014) who use the form
max zðsa , sb Þ subject to (EXP)
sa ∈ ≥ 0 , sb ∈  ≥ 0
λðtÞ ≡ λ þ λσ sinðωt þ ϕÞ, (17)
sa þ sb ≤ s, (13) where λ > 0, σ ∈ [0, 1], ω > 0 and ϕ ≥ 0 are the

average arrival rate, relative amplitude, frequency,
and phase, respectively. We use a standard notation of our approach, we examine the limiting behavior
for the frequency of the sine function, that is, ω ¼ 2π
T, of (18) and find the extrema of these periodic orbits in
and assume the phase equals zero without loss of these two regions of interest.
generality.
LEMMA 6 (Asymptotic Dynamics). Suppose α j , β j
Substituting (17) into (8), we obtain the following
and γ j are real numbers for j ∈ {a, b, c} whose values
IVP.
are dependent on the initial conditions of (18). Then,
0 1 0 1
q_ a ðtÞ λ þ λσ sinðωtÞ μa ðq ðtÞ ^ sa Þ τa ðq ðtÞ sa Þþ
a a
B C B C
B q_ ðtÞ C ¼ B rqc ðtÞ μb ðqb ðtÞ^ sb Þ τb ðqb ðtÞ sb Þþ C
@ b A @ A (18)
þ
q_ c ðtÞ θac ðqa ðtÞ ^ sa Þμa þ θbc ðqb ðtÞ^ sb Þμb þ θτc τb ðqb ðtÞ sb Þ ðr þ ζÞqc ðtÞ,
q0 ¼ ðqa ðt0 Þ, qb ðt0 Þ, qc ðt0 ÞÞT :
Similar to Section 4.1, qa ðtÞ can be written in closed- 1. If min qa ðtÞ ≥ sa then there exists a set of
form while the solution to qb ðtÞ and qc ðtÞ remains as equilibrium points
8
>
> rθac sa μa þ sb ðτb μb Þðr þ ζÞ sb rðθτc τb θbc μb Þ θac sa μa þ sb μb ðθbc θc Þ
>
< , , qb ðtÞ ≥ sb
τb ðrð1 θτc Þ þ ζÞ rð1 θc Þ þ ζ
ðqb , qc Þ :¼ (19)
>
> rθac sa μa θac sa μa
>
: , , qb ðtÞ<sb ,
μb ðrð1 θbc Þ þ ζÞ rð1 θbc Þ þ ζ
in (9). In contrast, however, the dynamical system while qa ðtÞ converges to a periodic orbit of the form
described by (18) is bounded as per Lemma 4 and the
asymptotic solutions may not converge to limiting
qa ðtÞ ¼ αa þ βa sinðωtÞ þ γ a cosðωtÞ:
points. Thus, we apply the four-stage approach,
except now, we examine a system that converges to a
periodic function. 2. If max qa ðtÞ ≤ sa then q(t) converges to a set of
For the first stage, we again argue that characterizing periodic orbits of the form
the behavior of the system in only a few regions is suffi-
cient for developing a systematic approach to station- qj ðtÞ ¼ α j þ β j sinðωtÞ þ γ j cosðωtÞ for j ¼ fa; b; cg:
ary staffing in this time-varying environment. That is,
we partition the domain into three regions character-
ized by distinct modes of operation. The first mode has Notice that Lemma 6 describes the asymptotic
a high rate of employee utilization. The second mode is behavior of the system in two extreme cases. How-
the opposite, employees are underutilized. Finally, the ever, if max qa ðtÞ > sa and min qa ðtÞ < sa , the dynamics
third mode represents all other regions. cannot be so easily characterized. Figure 2 illustrates
Mode 1 - Overloaded Region: qa ðtÞ > sa and qb ðtÞ > sb ,
Mode 2 - Underloaded Region: qa ðtÞ ≤ sa and qb ðtÞ ≤ sb .
Mode 3 - Mixed Region: qa ðtÞ > sa and qb ðtÞ ≤ sb ,
Figure 2 Dynamics of ρqðtÞ when max q a ðtÞ > s a and min q a ðtÞ < s a
qa ðtÞ ≤ sa and qb ðtÞ > sb .
The institution’s goal is to provide service to as many
customers as possible while ensuring servers are suffi-
ciently utilized. Given the partition of the domain into
modes of operation, the highest utilization of servers is
achieved in the first mode. Thus, it makes sense to
investigate how the institution can remain in the over-
loaded region while avoiding the second mode, that is,
the underloaded region. As a result, in the second stage
one such example; qb ðtÞ and qc ðtÞ do not converge to a t → ∞ and q(t) converges to its non-homogeneous
sinusoid, rather, they are more complex periodic term, which by Lemma 6, is a periodic function or,
functions. Nevertheless, in all cases, including those in some cases, may be an equilibrium point.
where the dynamics cannot easily be characterized, Because the duration of the disorder period is
the extrema of the periodic orbits can be derived. determined by the homogeneous part of the gen-
eral solution in (9), Proposition 2 still holds, that is,
LEMMA 7 (Asymptotic Bounds). The limiting solu- the homogeneous systems corresponding to (11)
tions of q(t) obey the following relations and (18) are identical.
Thus, the final stage of our approach combines
λ λσ h μ i h μ i
qa ðtÞ ≥ q ¼ 2 μa sin arctan a þ ωcos arctan a , the results from stage two and three by assigning
a μa μa þ ω 2 ω ω
weights, derived in Proposition 2, to the asymptotic
β β
qb ðtÞ ≥ q ¼ αb þ βb sin arctan b þ ρb γ b cos arctan b , bounds from Lemma 7. The goal is to formulate an
b γb γb
optimization problem to determine sa and sb so as
qc ðtÞ ≤ qc :
to maximize the time spent in the overloaded
region and minimize the time spent in the under-
Upper bounds for qa ðtÞ and qb ðtÞ, and a lower loaded region. Using similar logic as in Section 4.1,
bound for qc ðtÞ, can also be derived although we omit we develop an objective function using q and qc
b
the results as they do not play a role in subsequent (i.e., the asymptotic bounds) multiplied by a con-
theory. Note that in order to describe when the vex combination of weights w 2 ðsa , sb Þ and w
1 ðsa , sb Þ.
bounds become tight, the second stage of the model- Because qc is reached when qa ðtÞ is overloaded, we
ing approach also requires that the time points corre- multiply qc by the proportion of time per period
sponding to when qa ðtÞ is underloaded or overloaded that qa ðtÞ spends in the overloaded state, (1 − κ).
be known. Similarly, we multiply q by κ to capture the
b
amount of time per period qa ðtÞ spends in the
LEMMA 8 (Crossing Points). The state function qa ðtÞ underloaded state. With a slight abuse of notation,
crosses the boundary Σ at times the objective is
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
β
2
βa ðsa αa þγ a Þðsa αa γ a Þ 1 ðsa , sb Þ
w
2arctan sa αa þγ a zðsa , sb Þ ¼ ð1 κρ ðsa ÞÞ ρ q ðs , s Þ
t¼
2π
þ n, where n ¼ 0, 1, 2 .. .: 1 ðsa , sb Þ þ w
w 2 ðsa , sb Þ b c a b
ω ω
2 ðsa , sb Þ
w
þ κ ρ ðsa Þ ρ q :
Although similar expressions for qb ðtÞ and qc ðtÞ can 1 ðsa , sb Þ þ w
w 2 ðsa , sb Þ b b
be derived, only the crossing points of qa ðtÞ are impor-
tant as they drive the asymptotic behavior of the sys- Notice that maximizing zðsa , sb Þ is akin to serving
tem. An implication of Lemma 8 is that the as many customers as possible while ensuring that
proportion of time qa ðtÞ spends underloaded or over- servers are highly utilized. That is, maximizing the
loaded per period can be obtained. Specifically, if first term increases the number of customers in the
t1 < t2 are two adjacent time points when qa ðtÞ crosses orbit while maximizing the second reduces the num-
Σ from above and below, respectively, κ :¼ ðt2 t1 Þ=T ber of idle servers by increasing the load of the sys-
is the proportion of time qa ðtÞ is underloaded. Thus, tem in the underloaded state. The mixed-integer
q(t) spends κT time units per period of qa ðtÞ in regions optimization model is
where the lower bounds in Lemma 7 are tight. Simi-
larly, q(t) spends (1 − κ)T time units per period in max zðsa , sb Þ subject to (SINE)
sa ∈  ≥ 0 , sb ∈  ≥ 0
regions where the upper bound is tight. Note that to
compute the crossing points for the value of the new
sa þ sb ≤ s, (20)
customers in the system, ρa qa ðtÞ, we need only multi-
ply the staffing level sa by ρ1 in Lemma 8. We denote
ζ þ
a
the proportion of time that ρqa ðtÞ is underloaded each
sb ≥ 1 θbc sa , (21)
period by κρ . r
The third stage of our approach accounts for the
behavior of q(t) at finite times. As in Section 4.1, k ðsa , sb Þ ≥ 0, k∈ f1, 2g,
w (22)
the dynamics of the system are initially governed
by two competing exponential functions and dur- where (20) and (21) remain as in Section 4.1 and (22)
ing this period, q(t) is said to be in a state of disor- is the analog of the constraints (15) and (16). We
der. The exponential functions approach zero as k ðsa , sb Þ terms implicitly depend
again note that the w
on the staffing level through the initial conditions of then qa ðtÞ will be overloaded indefinitely. According
the dynamical system. to the following result, if λ increases as fast as σ
decreases, the number of candidate solutions
4.3. Solution Approach decreases, that is, Ssine becomes empty.
In both EXP and SINE, the objective functions are
convex combinations of terms associated with the PROPOSITION 4. The optimal staffing level sa for SINE is
length of the disorder period. Unfortunately, these contained in the set
terms are implicit functions of the decision variables n o
and thus, a simple mixed-integer linear program can- Ssine :¼ sa ∈  ≥ 0 jsa ≥ min q ðtÞ, sa ≤ max q ðtÞ ,
a a
not be formulated. Nevertheless, in this section, we
show that the problems exhibit structural properties where q ðtÞ corresponds to qa ðtÞ in the underloaded state. If
a
that are conducive to formulating a solution proce- σ → 0 as fast as λ ! ∞, jSsine j ! 0.
dure that efficiently determines the optimal staffing
levels. Proposition 3 and 4 represent necessary conditions
First, the objective functions are neither convex nor for determining the optimal solutions to EXP and
concave. Thus, a straightforward relationship SINE. They also suggest that as the total number of
between the initial conditions of the dynamical sys- servers gets large and the requirements of the propo-
tem and the associated weights in the objective func- sitions are satisfied, the number of candidate solu-
tion cannot be exploited in an efficient solution tions decreases to 0. Thus, interestingly, as the system
approach. Instead, we prove that there exists a small size grows, relatively fewer points need to be evalu-
interval within the feasible region that contains the ated to find the optimal solution.
optimal solution. As a result, we only need to perform Our solution approach begins by computing the
an exhaustive search within this small interval to find upper and lower bounds of the sets Sexp and Ssine .
globally optimal solutions. Note that because assigning less than s servers is sub-
For EXP, notice that in Proposition 3, by requiring optimal, constraint (13) and (20) are always binding.
that qa ðtÞ transitions into the overloaded regime, we Then, starting from the upper bound, we iteratively
ensure that the optimal staffing level sa cannot be decrease sa and simultaneously increase sb until con-
greater than the maximum of qa2 ðtÞ which corre- straint (15) and (16) for EXP and constraint (22) for
sponds to the loyalty phase. Thus, we obtain the fol- SINE is no longer satisfied or we reach the lower
lowing result. bound of the sets Sexp and Ssine . Constraint (14) and
(21) may also reduce the upper bound of the feasible
PROPOSITION 3. The optimal staffing level sa for EXP is solution region if they are violated. During each itera-
contained in the set tion, we solve (11) for EXP and (18) for SINE using the
given staffing level. We then compute the objective
λa qa2 ðt Þ, if t ≥ 0, function value and determine whether it is the maxi-
Sexp :¼ sa ∈  ≥ 0 sa ≥ , sa ≤ ,
μa qa ðt0 Þ, otherwise: mum value observed so far. If EXP or SINE are infea-
sible, then we assign the value of the upper bound
where Sexp or Ssine to sa .
h i
log μa qa2 ðt0 Þ μλ0 μ λδ
1
eδt0 eμa t0 log μδλδ
1
4.4. Model Discussion
t ¼ :
a a a
μa δ The proposed four-stage procedure only requires that

the arrival process be modeled by a smooth function.
Furthermore, if λ0 ! ∞ as λ1 ! ∞, then jSexp j ! 0. Thus, the approach is general, and its implementation
is straightforward: to derive a staffing policy, it suf-
Proposition 3 provides a range of values where the fices to analyze the asymptotic behavior of (8), weight
optimal solution is guaranteed to be found. Notice its limiting points and solve an optimization problem.
that this set is much smaller than the size of the feasi- Furthermore, notice that we do not assign waiting
ble region corresponding to EXP. It also indicates that costs or penalties to customers that abandon: doing so
as λ0 and λ1 increase, the number of candidate solu- results in complex optimal control problems (e.g.,
tions decreases, that is, Sexp becomes empty. Similarly Anderson Jr et al. 2006) with limited tractability in
for SINE, notice that in Proposition 4 (see below), we cases with piecewise smooth systems. Massey and
require that qa ðtÞ must periodically reach an over- Pender (2018), for instance, propose a Lagrangian
loaded state. As a result, if sa is greater than the upper profit maximization based stationary staffing proce-
bound of the underloaded queue qa ðtÞ, then the over- dure for Erlang-A systems. Instead, we identify
loaded state is never reached. Furthermore, if sa is less preferable operating regimes (e.g., higher values of qc ,
than the lower bound of the underloaded queue qa ðtÞ, qc and q ) and determine the staffing policies that
b
induce them. The technique is similar to identifying high-likelihood of retention after a successful service
parameter values that result in desirable steady-state completion while a substantial chance that we will be
dynamics (see, e.g., Pender et al. 2017). The inclusion unable to retain clients in a congested system; our
of weights associated with the profitability of a cus- results are qualitatively similar as θic , i ∈ {a, b, τ}, is
tomer class generalizes the approach so that it remains varied between 0.5 and 1 (see Afèche et al. 2017, for
valid for environments where the value of serving a instance). Relative to the attrition rate ζ, we introduce
particular customer type may differ. For example, if fast, intermediate, and slowly varying forms of the
base customers are more profitable (i.e., spend more exponential (Section 4.1) and sinusoidal (Section 4.2)
time per visit or pay a subscription fee), it might be bet- arrival functions (see Appendix A “Simulation Parame-
ter for the firm to focus more on their abandonment. ters” for more details). The parameters are selected to
Conversely, if their switching cost is high, and thus ensure that the maximum intensity of arriving cus-
their loyalty, the firm can focus more on serving new tomers is significantly larger than the attrition rate.
customers. We choose a pool of twenty servers (s = 20) and
We assign larger weights to the limiting points of assign them to new and returning customers. This
state functions with greater variability during the experimental setup is similar to Chevalier and
early stages of the time horizon. That is, the weights Tabordon (2003), for instance, who consider a pool
in (9) characterize the deviation of the solution from of six specialized servers. We also conduct simula-
its steady-state caused by the initial conditions: lar- tions of larger systems (s = 40 and s = 60) to study
ger weights correspond to more variation prior to how performance scales with the size of the system.
the solution approaching its limiting behavior. The We vary the service rate (μi ) and abandonment rate
result is that queues with relatively high variation (τi ) for i = {a, b} and the rate that base customers
will be prioritized by EXP and SINE when assigning request service from the orbit (r). The rates are cho-
servers. Weighting the points in this fashion is simi- sen so that higher values are larger than the attri-
lar to keeping safety stock in inventory systems with tion rate and the smallest values remain close to it.
demand uncertainty, that is, products with higher We consider scenarios of high and low abandon-
demand variability have a larger amount of safety ment, service and arrival rate of returning clients
stock. (see Appendix A “Simulation Parameters”). We
repeat these trials for two time horizons; long
(t = 25) and short (t = 10). This dichotomy is con-
5. Numerical Results sistent with the literature (see Liu and Xie 2018, for
In this section, we present a comprehensive numerical instance). Each experiment begins with a burn-in
study of the staffing policies generated by optimiza- period of t = 5 time units after which we record
tion problems EXP and SINE. We compare their per- performance measures associated with service qual-
formance to modifications of the tail probability of ity. This includes the average number of new and
delay (TPoD) staffing policy developed in Liu (2018), a returning customers served (throughput) and the
generalized version of the square root staffing (SRS) average total number of customers who abandon.
policy, by varying customer patience levels, the speed Both metrics are commonly used in the literature to
of service, and the frequency of service requests. We evaluate the performance of fluid (deterministic)
also evaluate the magnitude of the deviation of EXP queueing systems (e.g., Bassamboo and Randhawa
and SINE from two benchmarks: (i) the optimal staff- 2015, Buzacott and Shanthikumar 1993). All
ing policy (OPT) determined by simulating all valid reported values represent the average over 10
staffing levels in the stochastic regime and choosing stochastic simulations. The staffing policies and
the policy that maximizes the throughput or mini- discrete-event simulation were programmed in Mat-
mizes the total number of class a and b clients that lab R2015a.
abandon the system; and (ii) a First-Come-First- We compare our stationary staffing policies, EXP
Served (FCFS) policy which makes the entire capacity and SINE, to four variants of the SRS rule that are sta-
available to new or returning customers based on a tionary or near-stationary. The first, GSRS0 , is a gener-
FCFS basis, that is, it assumes servers are fully flexible. alized square root staffing policy that does not allow
Finally, for different values of ρa , we identify settings any reassignment of staff over the time horizon. The
where EXP and SINE outperform the SRS-based second, GSRSt , is similar except that it allows staff to
benchmarks and discuss the implications for manage- be reassigned a fixed number of times. Both are modi-
ment theory. fications of the TPoD staffing policy. More specifi-
We fix the attrition rate of base customers from the cally, for the stationary benchmark, we compute
orbit to ζ = 1 and set the probabilities of congestion staffing levels according to
and fitness abandonment to θc ¼ 0:5 and θic ¼ 0:9 for pffiffiffiffi
sa :¼ sh þ ξ sh , sb :¼ s sa , (23)
i = {a, b}, respectively. This ensures that there is a
and performance (for additional results, see Appendix A

rffiffiffiffiffi
τa ρ λa “Summary of Results for EXP, s = 20”).
ξ :¼ zχ , sh :¼ a eτa h ,
μa μa As the system becomes less congested, the accuracy
of the fluid approximation naturally decreases which
where λa is the average arrival rate to Qa , μa is the reduces the performance gap among the policies. This
average service rate of new customers, and ξ is the explains why their throughput, for example, equal-
desired level of service. We compute zχ by inverting izes in the “slow server” setting. In systems with fas-
the standard normal distribution for a specified ter servers, the traffic intensity is significantly higher
probability of delay χ. Following Gans et al. (2003), due to the small number of servers assigned to new
Brown et al. (2005), Aksin et al. (2007), and Liu clients.
(2018), we adhere to the rule that 80 percent of ser- We note that as the system becomes more station-
vice requests are served within 1/3 of a time unit. ary (i.e., δ decreases) the performance of the bench-
Thus, we fix the probability of delay χ and delay marks improve and the gap between all staffing
parameter h to 0.2 and 1/3, respectively. We then policies decrease. We demonstrate this in Figure 3
implement (23) by computing λa over the entire time where the performance of the staffing policies for the
horizon. larger systems are presented. In both settings, s = 40
For our non-stationary benchmark, the staffing and s = 60, the throughput equalizes as the value of δ
levels follow the equations in (23) once again. How- becomes small whereas the gap widens as δ increases.
ever, for the advertising campaign, λa is computed by Moreover, the gap amongst all policies increase in the
splitting the time horizon into two equal intervals. For system size, that is, the performance of EXP improves
the clinical setting, we compute λa at times when λ(t) as the system becomes larger. Finally, we observe that
is above and below its symmetry line. This ensures the performance of EXP is close to OPT and the gap
that the staffing policy switches once per period of remains small as the system size increases.
λ(t). For both the advertising campaign and the clini- Because Figure 3 displays the typical performance
cal setting, our non-stationary benchmarks re-allocate of the policies, it also demonstrates that adjusting a
the staff over the intervals characterized by a higher staffing policy partway through the time horizon does
and lower intensity of arrivals of new customers, not increase throughput. In fact, doing so decreases
accordingly. We implement two versions of GSRS0 throughput, that is, it reduces service accessibility.
and GSRSt , with τa ¼ 0 and τa ≠ 0, corresponding to This is because, as observed in our simulations, EXP
the Erlang-C and Erlang-A versions of the TPoD, does not singularly focus on new customers. Instead,
respectively. it attempts to strike an equitable balance between the
acquisition of new customers and the retention of
5.1. Advertising Campaign: Results existing ones. Similar results are observed across all
In this setting, the amount that demand varies with simulation experiments which indicates that, in con-
time is determined solely by the decay rate of the trast to the GSRS benchmarks, EXP ensures high
exponential (i.e., δ) in (10). Systems that slowly con- levels of service quality for all customer types.
verge to a stationary state have small values of δ To demonstrate the effect that varying ρa has on
and are characterized by low levels of non- performance, we present a representative example
stationarity whereas systems that quickly reach a with fast servers and high variability in the exoge-
steady-state solution are more time-varying and nous demand (see Figure 4a). We find that EXP con-
have large values of δ. Because our results are quali- sistently outperforms the SRS benchmarks in
tatively similar for both short and long horizons, we virtually all settings. However, for large values of ρa ,
only discuss cases where t = 25. We first discuss set- serving new customers becomes the priority. In this
tings where new and returning customers are regime, all policies approach OPT as all capacity is
equally valuable. We then investigate the sensitivity assigned to new customers. In contrast, when ρa < 1,
of our results to changes in the relative value of the less capacity is allocated to new customers. Since
customer classes. returning customers were once new customers, EXP
With ρa ¼ 1, high levels of non-stationarity (δ ≥ 0.6) and the GSRS benchmarks struggle to determine the
and varying levels of traffic intensity, EXP outper- optimal allocation of server capacity which results in
forms GSRS0 and GSRSt , that is, throughput is better lower throughput as compared to OPT. Neverthe-
and total abandonment is similar or lower. For exam- less, even in this regime, there is a large
ple, when the abandonment rate is low (clients are difference in throughput between EXP and the GSRS
patient), the throughput of EXP is 1.3 and 1.25 times benchmark.
larger than the GSRS0 and GSRSt , respectively. Finally, in Figure 4b, we compare EXP with the
Remarkably, GSRS0 and GSRSt with τa ≠ 0 assign too FCFS policy as ρa is varied. Since FCFS assumes ser-
few servers to new clients which results in poor vers are fully flexible, the difference represents the
Figure 3 Total Throughput for EXP with t = 25 for the Systems where Base Customers Require Frequent Service: (a) s = 40; and (b) s = 60 [Color
figure can be viewed at wileyonlinelibrary.com]
Figure 4 Total Throughput as a New Customer’s Value ρa is Varied from 0 to 8 with t = 25 for Systems with Fast Servers and s = 40: (a) EXP vs.
OPT and the GSRS Benchmarks (b) EXP vs. OPT and FCFS [Color figure can be viewed at wileyonlinelibrary.com]
penalty (or decrease in throughput) that can be attrib- 5.2. Clinical Setting: Results
uted to specialization. We find that, on average, FCFS When demand is periodic, the degree of non-
generates approximately 6% and 16% more through- stationarity is determined by the cycle length T and
put than OPT and EXP, respectively. However, the the relative amplitude σ of the sinusoid function. Fol-
performance gap varies depending on the relative lowing Green et al. (1991), systems that exhibit high
value of new vs. returning customers. For instance, levels of time-varying behavior have short cycle
when ρa ¼ 0:25, returning customers are four times lengths and large amplitude levels.
more valuable than new customers and the difference Our results are qualitatively similar in this setting
in throughput between FCFS and EXP is 49%. As ρa as compared to Section 5.1. We again confirm that for
increases, this performance gap decreases. For exam- the most time-varying regimes, when values of σ and
ple, when new customers are twice as valuable as T are large and small, respectively, SINE outperforms
returning customers, the difference in throughput the benchmarks in all queueing regimes (for addi-
between FCFS and EXP is approximately 10% and this tional results, see Appendix A “Summary of Results
further reduces to approximately 4% when ρa ¼ 8. for SINE, T = 2, 8, s = 20”). This effect is demon-
The result provides insight as to the value that spe- strated in Figure 5. In this setting, for the systems
cialization should bring to the firm; the decrease in s = 40 and s = 60, the performance of SINE improves
throughput as predicted by our model should be bal- as σ increases. Once again, as the system becomes
anced by an increase in the probability that served more stationary, that is, σ decreases or T increases, the
customers (re-)join the customer base. performance of all the policies equalizes. According
to this figure, there is a gap between OPT and the rest

of policies but it does not grow as the system size
6. Discussion and Conclusion
increases. In addition, we observe that re-assigning We contribute to the operations research literature by
staff partway through the period of the arrival func- showing that our approach to staffing succeeds in
tion does not improve throughput and performs cases where standard near-stationary methods fail
worse than SINE. This result is consistent with our (see, e.g., Green and Kolesar 1991, Green et al. 1991).
findings from the previous section in that the perfor- That is, our policies perform as well as the bench-
mance of GSRSt is not better than GSRS0 . marks in more stationary settings and substantially
Similar to Section 5.1, we demonstrate the effect better in environments characterized by high levels of
that varying ρa has on performance by presenting a time variability and traffic intensity. In particular, our
representative example in a time-varying setting with methodology ensures that neither new nor returning
fast servers, T = 2, and σ = 0.9 (see Figure 6). customers have neglected—a common problem that
Although SINE outperforms the SRS-based policies in is present in the existing literature, for example, King
virtually all regimes, total throughput is affected by et al. (2016), Afèche et al. (2017)—while high through-
the relative value of new customers, that is, ρa . When put levels are maintained. Our method is also more
ρa < 1, although SINE outperforms the GSRS bench- time-efficient as compared to computer-based
marks, all polices struggle to determine an appropri- approaches that simulate multiple trajectories for each
ate assignment of capacity (as compared to OPT). staffing level and then use the results to derive the
Nevertheless, on average, the difference in through- optimal server split.
put between SINE and the best performing GSRS Our analysis demonstrates that stationary policies
benchmark is approximately 10%. As ρa exceeds one, may perform as well as time-dependent alternatives
SINE becomes substantially better at determining an when the time-varying dynamics of the system are
appropriate allocation of capacity. For instance, when accounted for. Specifically, the results from our
ρa ¼ 2, the performance gap between the policies numerical study suggest that there is little benefit to
widens to 18% as 20 better handles the time-varying re-assigning staff partway through an advertising
dynamics in the system (see Figure 6a). campaign. Furthermore, re-assigning staff partway
Finally, we again compare SINE with the FCFS pol- through a period, in a periodic environment such as a
icy as ρa is varied. As seen in Figure 6b, there is a per- clinic, is not useful either. In both cases, the bench-
formance gap as, on average, FCFS generates 8% and marks tend to severely overestimate or underestimate
25% more throughput than OPT and SINE, respec- the demand for either new or base clients and the
tively. When ρa ¼ 0:25, the difference in throughput quality of service deteriorates as a result. Obviously,
between FCFS and SINE is large (approximately if more adjustments are allowed, higher throughput
70%). However, this value quickly decreases as FCFS values can be achieved. However, frequently re-
generates only 13% and 3% more throughput than assigning staff is costly—both cognitively and finan-
SINE when ρa equals to 2 and 8, respectively. As pre- cially—and becomes less operationally feasible. Thus,
viously noted, this performance gap provides a quan- our work has important practical implications: man-
tifiable bound on the extra value that specialization agers who want to minimize the number of staffing
should bring to the firm. changes can be empowered to make well-performing
Figure 5 Total Throughput for SINE with t = 25 and T = 2 for the Systems where Base Customers Require Frequent Service: (a) s = 40; and (b)
s = 60 [Color figure can be viewed at wileyonlinelibrary.com]
Figure 6 Total Throughput as a New Customer’s Value ρa is Varied from 0 to 8 with t = 25 for Systems with Fast Servers and s = 40: (a) SINE vs.
OPT and the GSRS Benchmarks (b) SINE vs. OPT and FCFS [Color figure can be viewed at wileyonlinelibrary.com]
assignments using our stationary alternative while regimes for customer retention. Specifically, success-
still accounting for time-varying dynamics. ful services attract many customers and it is natural
We contribute to the literature on customer acquisi- that they, in operationally efficient organizations, will
tion and retention by providing new insights into the experience some effects of congestion (e.g., customers
field of customer relationship management (CRM). who must wait). According to Lu et al. (2013),
For example, King et al. (2016) suggest that compa- although periods of high congestion are correlated
nies with a larger customer base should begin to with high sales, a moderate increase in the number of
invest less in acquisition activities. This is because queued customers generates a sales reduction equiva-
retaining too many customers becomes expensive and lent to a 5% price increase. Campbell and Frei (2011)
acquiring new ones is no longer essential; growth is similarly show that longer waiting times affect cus-
not a priority for larger firms. We refine this argument tomer satisfaction and retention. We demonstrate that
by demonstrating that either customer acquisition or our approach gives customers more access to com-
retention efforts, if neglected, can reduce a customer’s pany services which reduces abandonment and as a
ability to access services. Thus, cost savings may come result, long wait times. Thus, it follows that our
at the expense of service quality, and this affects cus- approach to staffing, in addition to better balancing
tomers’ overall perception of the company’s value. acquisition and retention efforts, can lead directly to
This perception, in turn, may have implications for increased revenues.
the future: customers may begin to leave for no speci- Our approach is amenable to profitability consider-
fic reason (Gee et al. 2008, Griffin 2001) and the ability ations that are common in the CRM literature. In par-
to acquire new and retain base clients may be ticular, when all clients are equally valuable, our
hindered. staffing policy maximizes the common good, that is, it
Moreover, the need to serve both classes of cus- improves the service quality for all customers equally.
tomers at all times confirms a link between customer Conversely, if classes are not equally valuable, our
acquisition and retention efforts found in empirical method easily accounts for this by assigning greater
studies but not captured in the operations literature. weight to more profitable customers classes. We
According to marketing researchers, assuming that achieve this by circumventing classical approaches
these processes are independent leads to incorrect that use optimal control theory, which are notoriously
estimations of the duration of the relationship difficult to solve efficiently. Furthermore, we compare
between a customer and the firm, which is important our policy to a fully flexible, First-Come-First-Served
for retention analysis (Thomas 2001). Our approach benchmark and find that although there is a perfor-
explicitly models this dependence and suggests that mance loss when specialized training is required, the
focusing on acquisition or retention efforts alone is drop in throughput is negligible when new customers
not sufficient to maintain high levels of throughput are highly valued. Note that our results represent an
and service quality (e.g., Afèche et al. 2017, Araghi upper bound on the performance decrease as we do
2014). Instead, we provide a decision support tool for not account for the potential growth in customer
staffing that balances these competing priorities. retention as a result of the increased service quality
In addition, the high-traffic environments where provided by specialized employees. However, when
our staffing approach performs well are important the instantaneous value of an interaction with a base
customer is relatively high compared to the instanta- Bekker, R., A. M. de Bruin. 2010. Time-dependent analysis for
neous value of an interaction with a new customer, refused admissions in clinical wards. Ann. Oper. Res. 178(1):
45–65.
cross-training and dynamic reallocation of staffing
Berger, P. D., N. I. Nasr. 1998. Customer lifetime value: Marketing
between base and new customers may be particularly models and applications. J. Interact. Mark. 12(1): 17–30.
valuable. Bhandari, A., A. Scheller-Wolf, M. Harchol-Balter. 2008. An exact
Finally, we note that in cases where managers are and efficient algorithm for the constrained dynamic operator
only concerned with the long-term behavior of the staffing problem for call centers. Management Sci. 54(2): 339–
353.
system, our methodology may require some adjust-
Blattberg, R. C., J. Deighton. 1996. Manage marketing by the cus-
ment. In particular, the weights in Proposition 2 are
tomer equity test. Harv. Bus. Rev. 74(4): 136.
no longer necessary and an unweighted optimization
Borovkov, A. 1967. On limit laws for service processes in multi-
model can prove to be more useful and efficient. Fur- channel systems. Sib. Math. J. 8(5): 746–763.
thermore, although we investigate two arrival pro- Borst, S., A. Mandelbaum, M. I. Reiman. 2004. Dimensioning large
cesses motivated by common scenarios in the extant call centers. Oper. Res. 52(1): 17–34.
literature, that is, an advertising campaign (exponen- Boushey, H., S. J. Glynn. 2012. There are significant business costs
tial decay) and clinical setting (periodic), it would be to replacing employees. Cent. Am. Stud. 16: 1–9.
interesting to explore how our approach can be Brown, L., N. Gans, A. Mandelbaum, A. Sakov, H. Shen, S. Zel-
tyn, L. Zhao. 2005. Statistical analysis of a telephone call cen-
adapted to systems that are driven by different, more ter: A queueing-science perspective. J. Am. Statist. Ass. 100
complex, time-varying arrival functions. (469): 36–50.
Buttle, F. 2004. Customer Relationship Management. Routledge,
Abingdon-on-Thames.
References
Buzacott, J. A., J. G. Shanthikumar. 1993. Stochastic Models of Man-
Afèche, P., M. Araghi, O. Baron. 2017. Customer acquisition, ufacturing Systems, Vol. 4. Prentice Hall, Englewood Cliffs, NJ.
retention, and service access quality: Optimal advertising,
capacity level, and capacity allocation. Manuf. Serv. Oper. Campbell, D., F. Frei. 2011. Market heterogeneity and local
Manag. 19(4): 674–691. capacity decisions in services. Manuf. Serv. Oper. Manag. 13
(1): 2–19.
Akcali, E., M. J. Côtè, C. Lin. 2006. A network flow approach to
optimizing hospital bed capacity decisions. Health Care Manag. Chan, C. W. & V. Sarhangian 2018. Dynamic server assignment in
Sci. 9(4): 391–404. multiclass queues with shifts, with application to nurse staff-
ing in emergency. Retrieved from https://sarhangian.mie.
Aksin, Z., M. Armony, V. Mehrotra. 2007. The modern call center: utoronto.ca/wp-content/uploads/shiftscheduling_2020.pdf
A multi-disciplinary perspective on operations management (accessed date January 9, 2019).
research. Prod. Oper. Manag. 16(6): 665–688.
Chan, C. W., G. Yom-Tov, G. Escobar. 2014. When to use
Anderson, Jr E. G., D. J. Morrice, G. Lundeen. 2006. Stochastic speedup: An examination of service systems with returns.
optimal control for staffing and backlog policies in a two- Oper. Res. 62(2): 462–482.
stage customized service supply chain. Prod. Oper. Manag. 15
(2): 262–278. Chen, I. J., K. Popovich. 2003. Understanding customer relation-
ship management (CRM) people, process and technology.
Andradóttir, S., H. Ayhan, D. G. Down. 2001. Server assignment Bus. Process Manag. J. 9(5): 672–688.
policies for maximizing the steady-state throughput of finite
queueing systems. Management Sci. 47(10): 1421–1439. Chevalier, P., N. Tabordon. 2003. Over flow analysis and cross-
trained servers. Int. J. Prod. Econ. 85(1): 47–60.
Andradóttir, S., H. Ayhan, D. G. Down. 2003. Dynamic server
allocation for queueing networks with flexible servers. Oper. Columbus, L. 2016. 2015 gartner crm market share analysis shows
Res. 51(6): 952–968. salesforce in the lead, growing faster than market. Retrieved
from https://www.forbes.com/sites/louiscolumbus/2016/
Andradóttir, S, H. Ayhan, D. G. Down. 2007. Dynamic assignment 05/28/2015-gartner-crm-market-share-analysis-shows-salesf
of dedicated and flexible servers in tandem lines. Probab. Eng. orce-in-the-lead-growing-faster-than-market/ (accessed date
Inf. Sci. 21(4): 497–538. January 8, 2019).
Araghi, M. 2014 Problems in Service Operations with Heterogeneous Dant, R. P., P. D. Berger. 1996. Modelling cooperative advertising
Customers. University of Toronto, Canada. decisions in franchising. J. Oper. Res. Soc. 47(9): 1120–1136.
Bassamboo, A., R. S. Randhawa. 2010. On the accuracy of fluid Daw, A., J. Pender. 2018. Queues driven by hawkes processes.
models for capacity sizing in queueing systems with impa- Stoch. Syst. 8(3): 192–229.
tient customers. Oper. Res. 58(5): 1398–1413.
Daw, A., J. Pender. 2019. New perspectives on the erlang-a queue.
Bassamboo, A., R. S. Randhawa. 2015. Scheduling homogeneous Adv. Appl. Probab. 51(1): 268–299.
impatient customers. Management Sci. 62(7): 2129–2147.
Daw, A., A. Castellanos, G. B. Yom-Tov, J. Pender, L. Gru-
Bassamboo, A., A. Zeevi. 2009. On a data-driven method for staff- endlinger. 2020. The co-production of service: Modeling ser-
ing large call centers. Oper. Res. 57(3): 714–726. vice times in contact centers using hawkes processes. arXiv
Bassamboo, A., J. M. Harrison, A. Zeevi. 2006. Design and control preprint arXiv:2004.07861.
of a large call center: Asymptotic analysis of an lp-based De Bruin A. M., A. Van Rossum, M. Visser, G. Koole. 2007.
method. Oper. Res. 54(3): 419–435. Modeling the emergency cardiac in-patient flow: An applica-
Batta, R., O. Berman, Q. Wang. 2007. Balancing staffing and tion of queuing theory. Health Care Manag. Sci. 10(2): 125–137.
switching costs in a service center with flexible servers. Eur. J. Dean, A. M. 2002. Service quality in call centres: Implications for
Oper. Res. 177(2): 924–938. customer loyalty. Manag. Serv. Qual. Int. J. Res. 12(6): 414–423.
BCE 2016. Annual report. Retrieved from http://www.bce.ca/ Defraeye, M., I. Van Nieuwenhuyse. 2013. Controlling excessive
investors/AR-2016/2016-bce-annual-report.pdf (accessed date waiting times in small service systems with time-varying
February 2, 2018).
demand: An extension of the ISA algorithm. Decis. Support centres. Technical report, Technical Report, Department of
Syst. 54(4): 1558–1567. Engineering Science, University of Auckland.
Defraeye, M., I. Van Nieuwenhuyse. 2016. Staffing and scheduling Hillier, F. S., K. C. So. 1989. The assignment of extra servers to
under nonstationary demand for service: A literature review. stations in tandem queueing systems with small or no buffers.
Omega 58: 4–25. Perform. Eval. 10(3): 219–231.
Del Boca, A., P. Rota. 1998. How much does hiring and firing Holman, D., R. Batt, U. Holtgrewe. 2007. The global call center
cost? Survey evidence from a sample of Italian firms. Labour report: International perspectives on management and
12(3): 427–449. employment.
Ding, S., M. Remerova, R. D. van der Mei, B. Zwart. 2015. Fluid Hoyer-Leitzel, A., A. Nadeau, A. Roberts & A. Steyer 2017.
approximation of a call center model with redials and recon- Detecting transient rate-tipping using steklov averages and
nects. Perform. Eval. Rev. 92: 24–39. lyapunov vectors. arXiv preprint arXiv:1702.02955.
Dong, Y., Y. Yao, T. H. Cui. 2011. When acquisition spoils reten- Hu, K., G. Allon, A. Bassamboo. 2016. Understanding customers
tion: Direct selling vs. delegation under CRM. Management retrial in call centers: Preferences for service quality and ser-
Sci. 57(7): 1288–1299. vice speed. Available at SSRN 2838998.
Duan, Y., Y. Jin, Y. Ding, M. Nagarajan, G. Hunte. 2020. The cost Huang, J., B. Carmeli, A. Mandelbaum. 2015. Control of patient
of task switching: Evidence from the emergency department. flow in emergency departments, or multiclass queues with
Available at SSRN 3756677. deadlines and feedback. Oper. Res. 63(4): 892–908.
Feinberg, R. A., L. Hokama, R. Kadam, I. Kim. 2002. Operational Iglehart, D. L. 1965. Limiting diffusion approximations for the
determinants of caller satisfaction in the banking/financial many server queue and the repairman problem. J. Appl. Pro-
services call center. Int. J. Bank Mark. 20(4): 174–180. bab. 2(2): 429–441.
Feldman, Z., A. Mandelbaum, W. A. Massey, W. Whitt. 2008. Iglehart, D. L. 1973a. Weak convergence in queueing theory. Adv.
Staffing of time-varying queues to achieve timestable perfor- Appl. Probab. 5(3): 570–594.
mance. Management Sci. 54(2): 324–338. Iglehart, D. L. 1973b. Weak convergence of compound stochastic
Fruchter, G. E., Z. J. Zhang. 2004. Dynamic targeted promotions: process, I. Stoch. Proc. Appl. 1(1): 11–31.
A customer retention and acquisition perspective. J. Serv. Res. Jacobson, E. U., N. T. Argon, S. Ziya. 2012. Priority assignment in
7(1): 3–19. emergency response. Oper. Res. 60(4): 813–832.
Futamura, K. 2000. The multiple server effect: Optimal allocation Jain, D., S. S. Singh. 2002. Customer lifetime value research in
of servers to stations with different servicetime distributions marketing: A review and future directions. J. Interact. Mark.
in tandem queueing networks. Ann. Oper. Res. 93(1): 71–90. 16(2): 34–46.
Gans, N., G. Koole, A. Mandelbaum. 2003. Telephone call centers: Janssen, A., J. S. Van Leeuwaarden, B. Zwart. 2011. Refining
Tutorial, review, and research prospects. Manuf. Serv. Oper. square-root safety staffing by expanding erlang C. Oper. Res.
Manag. 5(2): 79–141. 59(6): 1512–1522.
Gee, R., G. Coates, M. Nicholson. 2008. Understanding and prof- Jouini, O., G. Koole, A. Roubos. 2013. Performance indicators for
itably managing customer loyalty. Mark. Intell. Plan. 26(4): call centers with impatient customers. IIE Trans. 45(3): 341–
359–374. 354.
Green, L., P. Kolesar. 1991. The pointwise stationary approxima- Kang, W., K. Ramanan, et al. 2010. Fluid limits of many-server
tion for queues with nonstationary arrivals. Management Sci. queues with reneging. Ann. Appl. Probab. 20(6): 2204–2260.
37(1): 84–97.
Keaveney, S. M. 1995. Customer switching behavior in service
Green, L., P. Kolesar, A. Svoronos. 1991. Some effects of nonsta- industries: An exploratory study. J. Mark. 59, 71–82.
tionarity on multiserver markovian queueing systems. Oper.
King, G. J., X. Chao, I. Duenyas. 2016. Dynamic customer acquisi-
Res. 39(3): 502–511.
tion and retention management. Prod. Oper. Manag. 25(8):
Green, L. V., P. J. Kolesar, W. Whitt. 2007. Coping with time- 1332–1343.
varying demand when setting staffing requirements for a ser-
Kitaev, M. Y., R. F. Serfozo. 1999. M/m/1 queues with switching
vice system. Prod. Oper. Manag. 16(1): 13–39.
costs and hysteretic optimal control. Oper. Res. 47(2): 310–312.
Griffin, J. 2001. Winning customers back. Bus. Econ. Rev. 48(1): 8.
Koole, G. 1997. Assigning a single server to inhomogeneous
Hahn, J. 2016. Hopf bifurcations in fast/slow systems with rate- queues with switching costs. Theoret. Comput. Sci. 182(1-2):
dependent tipping. arXiv preprint arXiv:1610.09418. 203–216.
Halfin, S., W. Whitt. 1981. Heavy-traffic limits for queues with L’Ecuyer, P., K. Gustavsson, L. Olsson. 2018. Modeling bursts in
many exponential servers. Oper. Res. 29(3): 567–588. the arrival process to an emergency call center. 2018 Winter
Hampshire, R. C., O. B. Jennings, W. A. Massey. 2009. A time- Simulation Conference (WSC), IEEE, pp. 525–536.
varying call center design via lagrangian mechanics. Probab. Lee, N., V. G. Kulkarni, Y. Hirasawa. 2014. Optimal static assign-
Eng. Inf. Sci. 23(2): 231–259. ment and routing policies for service centers with correlated
Hampshire, R. C., S. Bao, W. S. Lasecki, A. Daw, J. Pender. 2020. traffic. Probab. Eng. Inf. Sci. 28(3): 279–311.
Beyond safety drivers: Applying air traffic control principles Liu, Y. 2018. Staffing to stabilize the tail probability of delay in
to support the deployment of driverless vehicles. PLoS ONE service systems with time-varying demand. Oper. Res. 66(2):
15(5): e0232837. 514–534.
Harrison, J. M., A. Zeevi. 2004. Dynamic scheduling of a multi- Liu, Y., W. Whitt. 2017. Stabilizing performance in a service sys-
class queue in the halfin-whitt heavy traffic regime. Oper. Res. tem with time-varying arrivals and customer feedback. Eur. J.
52(2): 243–257. Oper. Res. 256(2): 473–486.
Harrison, J. M., A. Zeevi. 2005. A method for staffing large call Liu, R., X. Xie. 2018. Physician staffing for emergency departments
centers based on stochastic UID models. Manuf. Serv. Oper. with time-varying demand. INFORMS J. Comput. 30(3): 588–607.
Manag. 7(1): 20–36.
Lu, S. F., L. X. Lu. 2017. Do mandatory overtime laws improve
Henderson, S., A. Mason, I. Ziedins, R. Thomson. 1999. A heuris- quality? staffing decisions and operational flexibility of nurs-
tic for determining efficient staffing requirements for call ing homes. Management Sci. 63(11): 3566–3585.
Lu, Y., A. Musalem, M. Olivares, A. Schilkrut. 2013. Measuring Rodrı́guez-Gutiérrez, C. 2007. Effects of temporary hiring on the
the effect of queues on customer purchases. Management Sci. profits of Spanish manufacturing firms. International Journal
59(8): 1743–1763. of Manpower.
Madiedo, J. P., A. Chandrasekaran, F. Salvador. 2020. Capturing Rogers, R. D., S. Monsell. 1995. Costs of a predictible switch
the benefits of worker specialization: Effects of managerial between simple cognitive tasks. J. Exp. Psychol. Gen. 124(2):
and organizational task experience. Prod. Oper. Manag. 29(4): 207.
973–994. Rubinstein, J. S., D. E. Meyer, J. E. Evans. 2001. Executive control
Malhotra, N. K. 2007. Review of marketing research. Review of Mar- of cognitive processes in task switching. J. Exp. Psychol. Hum.
keting Research, v–v. Emerald Group Publishing Limited, Bing- Percept. Perform. 27(4): 763.
ley. Smith, J. M., R. Barnes. 2015. Optimal server allocation in closed
Mandelbaum, A., W. A. Massey, M. I. Reiman. 1998. Strong finite queueing networks. Flex. Serv. Manuf. J. 27(1): 58–85.
approximations for markovian service networks. Queueing Smith, J. M., F. Cruz, T. van Woensel. 2010. Optimal server alloca-
Syst. 30(1-2): 149–201. tion in general, finite, multi-server queueing networks. Appl.
Massey, W. A., J. Pender. 2018. Dynamic rate erlang-a queues. Stoch. Models Bus. Ind. 26(6): 705–736.
Queueing Syst. 89(1-2): 127–164. Taylor, M. 2018. 18 CRM statistics you need to know for 2018.
Milovic, B. 2012. Application of customer relationship manage- Retrieved from https://www.superoffice.com/blog/crm-sof
ment strategy (CRM) in different business areas. Fu. Econ. tware-statistics/ (accessed date January 9, 2019).
Org. 9(3): 341–354. Thomas, J. S. 2001. A methodology for linking customer acquisi-
Moon, K., P. K. Loyalka, P. Bergemann & J. Cohen 2020. The hid- tion to customer retention. J. Mark. Res. 38(2): 262–268.
den cost of worker turnover: Attributing product reliability to Thompson, G. L., J. T. Teng. 1984. Optimal pricing and advertis-
the turnover of factory workers. Available at: https://fsi.stanf ing policies for new product oligopoly models. Mark. Sci. 3(2):
ord.edu/publication/hidden-cost-worker-turnover-attributing- 148–168.
product-reliability-turnover-factory-workers (accessed date
Tsai, Y. C., N. T. Argon. 2008. Dynamic server assignment policies
January 9, 2019).
for assembly-type queues with flexible servers. Nav. Res.
Niyirora, J., J. Pender. 2016. Optimal staffing in nonstationary ser- Logist. 55(3): 234–251.
vice centers with constraints. Nav. Res. Logist. 63(8): 615–630.
de Véricourt F, O. B. Jennings. 2011. Nurse staffing in medical
Oracle 2011. Customer experience impact report. Retrieved from units: A queueing perspective. Oper. Res. 59(6): 1320–1331.
http://www.oracle.com/us/products/applications/cust-exp-
Whitt, W. 2006. Fluid models for multiserver queues with aban-
impact-report-epss-1560493.pdf (accessed date February 2,
donments. Oper. Res. 54(1): 37–54.
2018).
Whitt, W. 2014. The steady-state distribution of the mt/m/∞
Park, P. S., P. M. Bobrowski. 1989. Job release and labor flexibility
queue with a sinusoidal arrival rate function. Oper. Res. Lett.
in a dual resource constrained job shop. J. Oper. Manag. 8(3):
42(5): 311–318.
230–249.
Whitt, W. 2016. Queues with time-varying arrival rates: A bibliogra-
Parsons, L. J. 1975. The product life cycle and time-varying adver-
phy. Technical report, Working paper.
tising elasticities. J. Mark. Res. 12(4): 476–480.
Whitt, W. 2018. Time-varying queues. Queueing Models Serv.
Pender, J., R. H. Rand, E. Wesson. 2017. Queues with choice via
Manag. 1(2): 15–67.
delay differential equations. Int. J. Bifurc. Chaos 27(04):
1730016. Yom-Tov, G. B., A. Mandelbaum. 2014. Erlang-r: A time-varying
queue with reentrant customers, in support of healthcare
Reichheld, F. F., J. W. Sasser. 1990. Zero defections: Quality comes
staffing. Manuf. Serv. Oper. Manag. 16(2): 283–299.
to services. Harv. Bus. Rev. 68(5): 105–111.
Reichheld, F. F., P. Schefter. 2000. E-loyalty: Your secret weapon
on the web. Harv. Bus. Rev. 78(4): 105–113.
Supporting Information
Reinartz, W., J. S. Thomas, V. Kumar. 2005. Balancing acquisition Additional supporting information may be found online
and retention resources to maximize customer profitability.
J. Mark. 69(1): 63–79.
in the Supporting Information section at the end of the
article.
Robbins, T. R., T. P. Harrison. 2010. A stochastic programming
model for scheduling call centers with global service level
agreements. Eur. J. Oper. Res. 207(3): 1608–1619. Appendix and Proofs of Statements.

Customer Acquisition and Retention: A Fluid Approach For Staffing

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Customer Acquisition and Retention: A Fluid Approach For Staffing

Uploaded by

Copyright:

Available Formats

Vol. 30, No. 11, November 2021, pp. 4236–4257 DOI 10.1111/poms.