See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4053730

Call center simulation modeling: Methods, challenges, and opportunities

Conference Paper  in  Proceedings - Winter Simulation Conference · January 2004

DOI: 10.1109/WSC.2003.1261416 · Source: IEEE Xplore

CITATIONS READS

73 5,965

2 authors, including:

Vijay Mehrotra
University of San Francisco
25 PUBLICATIONS   903 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Vijay Mehrotra on 17 May 2014.

The user has requested enhancement of the downloaded file.

Proceedings of the 2003 Winter Simulation Conference
S. Chick, P. J. Sánchez, D. Ferrin, and D. J. Morrice, eds.

CALL CENTER SIMULATION MODELING: METHODS, CHALLENGES, AND OPPORTUNITIES

Vijay Mehrotra Jason Fama

Department of Decision Sciences Engineering Group

College of Business Blue Pumpkin Software Inc.
San Francisco State University 884 Hermosa Court
1600 Holloway Avenue Sunnyvale, CA 94085, U.S.A.
San Francisco, CA 94123, U.S.A.

ABSTRACT an estimated 3% of the United States population works in

this industry. Most recent: an explosion of outsourced call
Using stochastic models to plan call center operations, centers springing up in India, the Philippines, the Carib-
schedule call center staff efficiently, and analyze projected bean, and Latin America, serving overseas customers in the
performance is not a new phenomenon, dating back to Er- United States and Western Europe as well as growing do-
lang's work in the early twentieth century. However, sev- mestic market needs.
eral factors have recently conspired to increase demand for From a mathematical perspective, call centers are in-
call center simulation analysis. teresting for a variety of reasons:
• Increasing complexity in call traffic, coupled with • Call centers typically handle more than one type
the almost ubiquitous use of Skill-Based Routing. of call, with each distinct call type referred to as a
• Rapid change in operations due to increased “queue” (as discussed below, this usage is not
merger and acquisition activity, business volatil- consistent with our normal definition of a queue).
ity, outsourcing options, and multiple customer • Inbound calls within each queue arrive at random
channels (inbound phone, outbound phone, email, over the course of time.
web, chat) to support. • In many centers, agents make outbound calls to
• Cheaper, faster desktop computing, combined customers, either proactively (typically for tele-
with specialized call center simulation applica- marketing or collections activities) or as a follow-
tions that are now commercially available. up to previous inbound calls.
In this tutorial, we will provide an overview of call center • Each call is of a random duration, as is the work
simulation models, highlighting typical inputs and data (data entry, documentation, research, etc.) that
sources, modeling challenges, and key model outputs. In agents must do after completing the phone call.
the process, we will also present an interesting “real- • Through Automatic Call Distribution (“ACD”)
world” example of effective use of call center simulation. and Computer Telephony Interaction (“CTI”) de-
vices, inbound calls can be routed to agents,
1 INTRODUCTION: “WHY CALL CENTERS?” groups, and/or locations, with advancements in
these routing technology supporting more and
The trend in our economy from manufacturing towards more sophisticated logic over time.
services is well documented. One notable facet of this • Individual agents can be skilled to handle one
transition towards services has been the explosion of the type of call, several types of calls, or all types of
call center industry. calls, with different priorities and preferences
Mehrotra (1997) defines call centers as “Any group specified in the routing logic.
whose principal business is talking on the telephone to cus- Thus, call centers can be thought of as stochastic sys-
tomers or prospects.” In this paper, we will refer to the tems with multiple queues and multiple customer types.
individuals who talk on the phone with customers as As we discuss below, there are great challenges associated
“agents.” with managing these systems effectively.
While the size of the industry is difficult to accurately To summarize, call centers are of interest both because
determine, a plethora of statistics from diverse sources re- of the sheer size of the industry, both in the United States
flect that fact that this is a huge and growing global indus- and overseas, and because of the operational and mathe-
try. Most stunning: Mandelbaum (2001) cites a study that matical complexity associated with these operations, which

135
 Mehrotra and Fama

makes it difficult for decision makers to understand system On a day-to-day basis, while simultaneously keeping
dynamics without effective modeling. costs, service quality, and employee satisfaction, these ex-
The remainder of this tutorial is organized as follows. ecutives and managers must (implicitly or explicitly) an-
In Section 2, we motivate the need for and value of simula- swer a number of important questions for which decision
tion in the context of effective call center management. In support models are valuable:
Section 3, we discuss how call centers make use of simula- • How many agents should we have on staff with
tion models, focusing on the key output statistics that are which particular skills? How should we schedule
used for system performance evaluation. In Section 4, we these agents’ shifts, breaks, lunches, training,
provide a modeling framework for call center simulation, meetings and other activities?
and discuss the key inputs associated with simulation mod- • How many calls of which type do we expect at
els, introducing the concepts through the formulation of a which times?
simulation model. In Section 5, we discuss business deci- • How quickly do we want to respond to each type
sions associated with this model and explore some of the of inbound call?
results of our analysis. Finally, in Section 6, we propose • How should we cross-train our agents? How
likely future directions for call center simulation. should we route our calls to make the best use of
Note: Throughout this paper, we will use the term these resources?
“call center” and focus our discussion on centers that are • Given a forecast, a routing design, and an agent
processing only phone calls (either inbound, outbound, or schedule, how well will our system perform?
both). Another common term in this industry is “contact • What is our overall capacity? How will a spike in
center,” which refers to centers handling not only phone call volumes impact our overall performance?
calls but other types of customer contacts such as email,
• How is our center doing right now? What has
fax, paper, and/or chat sessions. We have chosen to focus
changed since we did our last forecast and pub-
on call centers here for clarity of exposition. However, lished our schedules? If the changes are signifi-
leveraging the ideas presented here from phone-only call
cant, what can I do to respond to minimize the
centers to multi-channel contact centers is a straightfor-
impact on the rest of the day or week?
ward extension that we have also engaged in extensively. There are a variety of mathematical methods (see
Grossman et al. 2001 and Mandelbaum 2001 for more dis-
2 CALL CENTER MANAGEMENT cussion of this) and associated software to help call center
CHALLENGES AND THE NEED personnel as they try to address these types of questions,
FOR MODELS most notably workload forecasting models based on time
series and agent scheduling optimization solutions.
Those responsible for managing call centers face a very However, over the past several years, simulation has
difficult set of challenges. At a high level, they must strike emerged to play an important role in the call center design
a balance between three powerful competing interests, as and management arena.
shown in Figure 1 below.
3 HOW CALL CENTERS USE SIMULATION

Costs There are three major ways that simulation is utilized

within the call center industry:
1. Traditional Simulation Analysis: A simulation
model is built to analyze a specific operation, with
inputs obtained from a variety of data sources, as
discussed in Section 4 below.
Service 2. Embedded Application – ACD/CTI Routing:
Many of the leading ACD and CTI applications
Quality include a routing simulation to provide insights to
routing design engineers about the impact of dif-
ferent decisions.
3. Embedded Application – Agent Scheduling: Al-
ready a complex scheduling problem (see Andrews
and Parsons 1989 for a more detailed discussion),
Employee optimal call center agent scheduling is even more
complex when both calls and agents are non-
Satisfaction homogeneous. Many commercial scheduling soft-
Figure 1: The Call Center Management Balancing Act ware applications, including the one developed by

136
 Mehrotra and Fama

the authors’ company, make use of simulation as

part of their overall optimization engine.
Abandonment
In each of these cases, the key output statistics typi- Queue Defs, Model and
cally include the some or all of the following metrics: Time Period, Parameters
• Queue Statistics: The two dominant queue statis- Routing Logic
tics for inbound queues and call centers are Aver-
age Speed of Answer (“ASA”) and Percent of Agent: Agent Call FCs
Skill Schedules
Calls Answered with a queue time of less than Definitions (Volumes,
some defined value (“PCA” or, more commonly, (Shifts and Patterns,
“Service Level).” Note that for each queue this Activities) AHTs)
statistic is interesting at the interval level (typi-
cally 15 minutes, 30 minutes, or one hour) and
also at the aggregate daily and weekly levels; ad-
ditionally, management is interested in the overall
performance across a collection of queues that Call Center
draw upon a common pool of agent resources.
• Abandonment Statistics: For most inbound call Simulation Modeling
centers, particularly those focused on customer ser- Engine
vice and/or sales, a great deal of attention is paid to
the overall number of customers who abandon (that
is, hang up and thus leave the queue before being
served). This is known to be a significant indicator
of customer satisfaction (see Feinberg et al. 2000 System
Performance
for a recent published study on this). Many centers Predictions
will look at the more restrictive metric of number
of customers abandoning beyond the target Service
Level parameter, based on the rationale that a cer- Figure 2: Call Center Simulation Modeling Framework
tain waiting time in queue (as defined by the Ser-
vice Level parameter, which ranges from 5 seconds
to several minutes across companies and industries) sources, and as with all simulation designs, there are deci-
is inevitable. sions to be made about the level of detail to include in the
• Volume Statistics: For outbound queues and call model.
In the sections below, we discuss these key input areas
centers, the real statistic of interest is Right Party
in more detail. In the process, we will use our example
Connects (“RPCs”). That is, for all of the at-
tempted calls that were made, what percentage of model to illustrate these modeling concepts.
these calls reached the targeted individual (as op-
posed to no answer, answering machine, or some 4.2 Key Inputs: Queue Definitions, Time
other human being). Outbound contact center Period, and Routing Logic
managers are typically interested in RPC on both
an absolute and a percentage basis. For inbound The basic building blocks of a call center simulation model
queues, the Calls Handled statistic is of interest, are the calls, the agents, and the time period during which
and is easily derived by subtracting Abandoned the call center is open. In turn, the basic routing logic
Calls from the total number of incoming calls (re- connects the way that the calls interact with the people dur-
ferred to as “Offered Calls”). ing that time period.
Typical call center simulation models contain more
4 CALL CENTER SIMULATION MODELLING than one queue (as single queue models are ordinarily dealt
with analytically) and run for a period of one day, one
week, or multiple weeks.
4.1 Framework
Our example model is for a Collections call center. As
is typical for call centers, this operation is part of a larger
The biggest challenge of call center simulation modeling is
business context, in which creditors’ records are being
the definition and organization of model inputs. Figure 2
monitored on a regular basis for potential delinquency.
below illustrates our framework for call center simulation
Once a customer falls into delinquency, several things
model definition and key inputs.
happen: (a) the information on the account is added to a list
As reflected in Figure 2, call center simulation models
of prospects for an Outbound collections call; (b) they are
feature a diverse range of inputs from multiple data

137
 Mehrotra and Fama

notified about the state of their credit by mail; and (c) addi- val, it has been customary to translate call volume forecasts
tional limitations may be placed on the account. into λ values for Poisson arrivals and AHT forecasts into µ
The call center itself features two queues: an Inbound values for Exponential service times.
queue and an Outbound queue. The time period that we A great deal of research has been conducted on call
are using for the analysis is one week. volume forecasting models, and the interested reader is re-
The two agent groups and the basic routing logic are ferred to Mabert (1985) and Andrews and Cunningham
illustrated in Figure 3. Calls from the Inbound queue will (1995) for valuable discussions on this topic.
arrive and be served by an agent from Group #1, the In- Forecasts must be created for each queue for each time
bound-Only group. If no agent from this group is avail- interval in the simulation period.
able, the calls will wait in queue. If, after some pre- The most common call center forecasting approach is
defined period of time, the Inbound call has not yet been to create weighted averages of historical data for specific
served, it will then also queue for an agent from Group #2, time intervals over the course of a week. For example, the
the Cross-Trained Outbound group. initial call volume forecast for 8:15 a.m. - 8:30 a.m. next
Tuesday might be computed as the average of call volume
for the 8:15 a.m. - 8:30 a.m. period on the past several
Abandonment Tuesdays. From here, alterations may be (or more com-
Outbound Call
Prospects
monly, should be!) made based on additional information
(Effectively (e.g. specific marketing activities for a sales center or
Inbound Calls:
Unlimited Pool) Arriving emerging product issues for a technical support center) that
Randomly may cause volume to differ substantially from previous
Based on FCs patterns.

Agent 4.3.1 Average Handle Time Forecasts

Group #3:
Outbound
Only Agent As mentioned earlier, most call center models assume that
Agent
Group #2: call handling times are exponentially distributed. We
Group #1:
Cross- would recommend using more accurate distributional in-
Inbound
Trained
Outbound
Only formation about call handling times whenever possible.
For example, it is common to find technical support call
center for which call handling times are bi-modal (easy
Figure 3: Queues, Agent Groups, and Routing Logic for cases with a shorter mean, harder cases with a longer one).
Example Model However, the primary reason that the call center indus-
try accepts the assumption of exponential handling times is
Meanwhile, Group #2 and Group #3 agents will be because the ACD and CTI devices (the primary source for
logged into a Predictive Dialer, which places Outbound historical call volume data) store only average handling
calls to prospects from a list. When an answer is detected times at the interval level. With a dearth of consistent sec-
by the Predictive Dialer, the call is automatically routed to ond moment information available, we have thus accepted
one of the Outbound Specialists or one of the available this assumption far more often than we would like; in par-
Cross-Trained Outbound agents. The agent then engages ticular, we have modeled exponential handling times in the
in a collections discussion (if they have reached the Right numerical example presented in Section 5.
Party) or leave a message (if they have reached an answer- Note: in this paper, we refer to Average Handling
ing machine or another party on the same phone number). Time, or AHT. However, when obtaining data from ACD
Along with mailings to delinquent customers, these mes- reports, it is not uncommon to find two fields that are then
sages will generate some calls to the Inbound queue. summed together to compute AHT: Average Talk Time
(“ATT”) and After Call Work (“ACW”).
4.3 Key Inputs: Call Forecasts
4.4 Key Inputs: Agent Schedules
Call forecasts are typically driven by a combination of his-
torical data, time series model, and expert judgment. There Agent schedules can be thought of as a series of activities
are two major types of call forecasts: call volumes and taking place over the course of a day. For example, an
Average Handling Time. Both are required for any basic agent who comes to work at 8:00 am for a nine-hour shift
call center simulation. may have a 15-minute break at 9:45 am, lunch at 11:30 am,
Due to the telecommunication and call center indus- an on-line training course from 1:00 – 2:00 pm, and a
tries’ history of using steady state M/M/n queue formulas break at 3:15 pm before leaving work at 5:00 pm.
to derive the number of agents needed for each time inter-

138
 Mehrotra and Fama

From a simulation perspective, each agent is viewed as by Garnett et al. 2002). We refer to the mean of this distri-
a resource to perform certain types of work. Note that in bution as “the patience factor.”
the call center context, agents are actually productive only Given this modeling choice, we must still with the
during the interval in which the agents are scheduled to be challenge of selecting the patience factor, which we esti-
actually handling phone calls. mate from historical data about callers’ time in queue.
In addition, it is conventional to model agents as com- We do not include caller retrial in the example model.
pleting the task that they are engaged, even if it extends
past the time at which they are to switch activities. That is, 4.6 Key Inputs: Agent Skills
an agent within our simulation will be modeled as complet-
ing the phone call that he is working on before leaving for Our definition of “Agent Skills” is comprised of three ma-
a break or a lunch. jor types of inputs for each agent or group of agents:
A common step in call center simulation is to translate a 1. What calls is the agent capable of handling?
set of individual agent schedules into a matrix of resources, 2. Given a choice of multiple calls waiting, which
where the dimensions of the matrix are defined by the num- will the agent handle (“call priority”)
ber of Agent Groups and the number of Time Intervals. 3. How fast will the agent be able to handle each
In our example, we have leveraged the fact that our type of call, and how often will the agent resolve
schedules are at a 15-minute level of granularity, and there- the issue successfully (“call proficiency”)
fore prior to running the simulation we have converted When combined with routing logic and call forecasts,
these schedules into a number of on-phone agents for each these attributes fully specify the queueing model to be
group for each 15-minute interval. simulated.
In our example, we have three distinct groups of
4.5 Key Inputs: Abandonment agents, each with different skills:
Model and Parameters • Agent Group #1 (Inbound Only) handle only In-
bound calls on a First-Come-First-Served basis.
Abandonment is one of the most hotly debated topics in These agents have a call proficiency of 1.0 for In-
call center management and research. There are two basic bound Calls, so that their AHT is equal to the
questions that must be answered in order to effectively forecasted AHT for Inbound Calls.
model customer abandonment behavior: • Agent Group #3 (Outbound Specialists) handle
1. What is the customer’s tolerance for waiting, and only Outbound calls. These agents have a call
at what point will this customer hang up and proficiency of 1.0 for Outbound Calls, meaning
thereby leave the queue? that their AHT is equal to the forecasted AHT for
2. How likely is the customer to call back, and after Outbound Calls.
how long? • Agent Group #2 (Cross-Trained Outbound) handle
Many researchers (e.g. Hoffman and Harris 1986, An- both Inbound and Outbound calls. These agents
drews and Parsons 1993) have examined the challenge of have a call proficiency of 1.0 for Outbound Calls,
modeling these problems from both an empirical point of meaning that their AHT is equal to the forecasted
view and from an analytic perspective. AHT for Outbound Calls. However, these agents
From our experience, these questions are difficult to will give priority to Inbound Callers if there are any
answer not only because of the mathematical complexity of waiting in queue, and have a call proficiency of 2.0
the queue dynamics but also because of a lack of observ- for Inbound calls, reflecting the relative ineffi-
able data about customer abandonment and retrial. While ciency of cross-training (see Pinker and Shumsky
many surveys have been done, we have observed great dif- 2000 for more discussion of this phenomenon both
ferences in customer behavior across different industries in and out of the contact center).
and different companies’ operations. In addition, informa-
tion provided to callers about expected waiting time and/or 4.7 Other Modeling Considerations
position in queue can have a marked impact on abandon-
ment behavior. 4.7.1 Shrinkage
In our example model, simulated customers arrive at the
call center and are served by an agent if one is available. If It is well known that a certain amount of agent time will be
not, they join the queue, at which point they are also as- lost, either in large blocks (unanticipated shift cancella-
signed a “life span” which is drawn from an exponential dis- tions, partial day absences for personal reasons) or in small
tribution. If a customer’s life span expires while they are blocks (late arrivals to the call center, extra-long breaks,
still waiting in queue, they then abandon the queue. trips to the bathroom).
That is, we represent customers’ tolerance for waiting There is an important distinction between two differ-
in queue as an exponential random variable (as suggested ent kinds of lost agent time. On one hand, agent time that

139
 Mehrotra and Fama

is known to be lost prior to the creation and publication of (Group #2 and Group #3). We treat agent schedules for
a schedule has essentially no additional impact on the each of the three agent groups, as well as call forecasts (a
simulation model beyond the fact that this particular agent total of about 20,000 calls for the week) for the Inbound
is not included in the schedule. calls, as fixed inputs for this simulation model. In addi-
On the other hand, scheduled time that is not worked, tion, we assume that there is an effectively unlimited pool
either because of unexpected absences or because of lack of customers to contact with Outbound calls.
of rigorous adherence with agent schedules, is time that The operational problem facing the management of
should be accounted for in the simulation if this represents this call center is focused on call routing and agent skilling.
a known phenomenon (e.g. higher absenteeism on Mon- Underlying this problem is the classical tension between
days). In the call center industry, this is known as “shrink- specialization and cost.
age” and it is a major management problem as well as sig- In terms of specialization, Inbound agents are far more
nificant modeling challenge. effective in handling Inbound calls than Outbound agents,
Most call centers have significant levels of shrinkage – while Inbound calls disrupt the rhythm and effectiveness of
we have seen many sites with over 30% overall. We have Outbound agents; for both of these reasons, it would be far
included a shrinkage level of 10% in our example model. better to have specialized agents for Inbound and Out-
bound calls respectively.
4.7.2 Additional Detail for Outbound Queues In terms of cost, there is a management objective of
handling 80% of Inbound calls within 60 seconds for each
As we have discussed earlier, the workflow associated with interval of the day. With dedicated agents, this translates
Outbound calls is very different than the logic for Inbound into a larger amount of Inbound agents required than are
queues. At heart, this modeling difference stems from the actually available. Current staffing levels, therefore, will
fact that inbound calls are characterized by a random arri- result in longer than desirable waiting times, which in turn
val pattern; in contrast, the outbound dialing pattern can be is correlated with higher abandonment rates.
scheduled but each call features a random outcome (right Specifically, the business decisions to be addressed are
party connect, wrong party connect, no answer). as follows:
In addition, as discussed in Section 3 above, the per- • Of the 150 Outbound-skilled agents, how many of
formance metrics associated with Outbound queues are them should be enabled to handle Inbound calls
quite different (overall RPCs achieved, rather than the and included in the Cross-Trained Outbound
queue and abandonment statistics that are typically used to group?
evaluate Inbound queues). In order to effectively estimate • If no Inbound agents are available, how quickly
the number and pattern of RPCs, simulation models require should an Inbound call be offered to the Cross-
information about the probabilities that a given dial Trained Outbound group?
achieves an RPC, which typically varies by time of day, as In an ideal world, there would be an empirical “right
well as the AHT associated with an RPC. answer” to these questions, a mathematically optimal solu-
To model one level deeper, one might consider actu- tion that could be determined through sequential simula-
ally representing the detailed logic of the predictive dialer tion runs.
(see Samuelson 1999 for more on predictive dialer logic). In practice, however, such decisions typically involve
However, this level of detail was not necessary for the substantial trade-offs that are difficult to value in relation
types of business decisions being addressed by our exam- to one another, and simulation’s role is to quantify the im-
ple model, and so we have not included detailed dialer pact of different possible decisions.
logic in it. The key output metrics for these simulations are:
1. Phone Service Levels (% of Inbound calls handled
5 EXAMPLE: ROUTING STRATEGIES FOR within 60 seconds).
A COLLECTIONS CALL CENTER 2. Abandonment Levels (% of Inbound callers
hanging up prior to receiving service).
5.1 Operational Problem and Business Decisions 3. Right Party Connects (total number of Outbound
calls completed to the correct individuals).
Throughout Section 4, we have described parts the simula- 4. Number of Overflows (of Inbound calls to Cross-
tion model associated with this example. The call center of Trained Outbound group).
interest is illustrated in Figure 3, and the formulation was
motivated by discussions with several blended inbound-
outbound centers about optimal system design.
In our example, the call center is open Monday - Fri-
day from 7:00 am to 6:00 pm. There are 50 Inbound
agents (Group #1) and a total of 150 Outbound agents

140
 Mehrotra and Fama

5.2 Numerical Results impact of this cross training on the Outbound call statistics.
These trade-offs are evident in Table 2 below.
5.2.1 Determining the Number of Replications Based on these preliminary simulations, we chose to
focus on cross-training a total of 30-40 Outbound agents.
For each of the individual scenarios that are discussed be- From here, we turned our attention to defining parameter
low, we ran multiple replications of the simulation model for how long Inbound calls should wait before being made
and computed estimates for performance measures based available to Cross-Trained Outbound agents.
on the average of the run length.
For purposes of determining the number of runs for 5.2.4 Varying the Wait Time Parameter for
each scenario, we focused on average weekly Service Overflowing Inbound Calls
Level for the Inbound queue as the statistic of interest. Af-
ter each run, we would examine overall standard deviation Results for different scenarios associated with 30 and 40
of this statistic across all runs to date. We continued to run Cross-Trained Outbound agents are shown in Tables 2
additional iterations until this overall standard deviation and 3.
was under 2.5%, which we had set arbitrarily as our confi-
dence threshold. Table 2: Simulation Results - 30 Cross-Trained Agents
Scenario Wait Until Abandon Number RPCs for
5.2.2 Base Case # Overflow SL % % Interrupts Week
4 0 seconds 86.7 4 2816 2339
Our baseline scenario is one with no Outbound Cross- 9. 15 seconds 85.9 4.2 2592 2360.3
Trained agents. This base case is listed as Scenario 1 in 10. 30 seconds 82.6 4.7 2403 2380.6
Table 1 below. 11. 45 seconds 78.3 5.7 2173 2414.3
12. 60 seconds 67.9 7.6 1769 2453.8
Table 1: Simulation Results For Base Case and Initial
Cross-Training Scenarios
Scenario Number Number Total Table 3: Simulation Results - 40 Cross-Trained Agents
# Cross- Inter- RPCs for Scenario Wait Until Abandon Number RPCs for
Trained SL % Abandon % rupts Week # Overflow SL % % Interrupts Week
1. 5 0 seconds 92.1 2.3 3242 2301.7
0 56.3 16 0 2621.8
2. 13. 15 seconds 90.2 2.7 3057 2324.7
10 70.1 10 1288 2494.4
3. 20 80 6.3 2211 2400.5 14. 30 seconds 87.7 3.3 2727 2347.2
4. 30 86.7 4 2816 2339 15. 45 seconds 81.9 4.3 2428 2377.9
5. 40 92.1 2.3 3242 2301.7 16. 60 seconds 70.1 6.3 1998 2427.2
6. 50 93.5 1.7 3374 2284
7. 75 97.1 0.8 3715 2255.1 5.2.5 Summary
8. 100 98 0.6 3726 2250.4
The different scenarios that we have simulated have en-
From this base case, it was clear that the Inbound abled us to (a) hone in on the right levels of cross training
Agent Group alone cannot deliver the desired Service Lev- to meet the Service Level goals with the current staffing
el (80% within 60 seconds), and that the Abandonment levels and (b) examine trade-offs between different scenar-
Rate is also much higher than desired. ios in terms of the key model outputs.
For example, consider Scenarios 3, 10, 14, and 15, all
5.2.3 Varying Cross-Training Levels of which deliver SLs at or above the 80% target. The an-
swer to which of these is the “best” choice will of course
We then began to vary the number of Outbound-Skilled depend on the relative value of RPCs, Service Levels, and
Agents who were included in the Cross-Trained Outbound Abandoned customers. However, it is interesting to note
group, assuming for these initial experiments that Inbound that Scenario 3 produces essentially the same SL and RPC
calls would immediately overflow to Cross-Trained Out- values as Scenarios 10 and 15 – but with a substantially
bound agents whenever all Inbound Only agents were higher abandonment rate. In turn, the tangible difference
busy. The impact of this cross training on the population between Scenario 14 and 15 enables managers to explicitly
of Inbound callers is dramatic, as even limited cross train- quantify the level of increased Service Level and decreased
ing has a big impact on Service Levels and Abandonment abandonment rates against the decreased number of RPCs.
Rates. In addition, there is an equally obvious negative Finally, it is worth mentioning that while we have
shown summary statistics for sixteen scenarios here, it is
relatively easy for us to produce more detailed statistics

141
 Mehrotra and Fama

and also to vary different parameters to examine any num- REFERENCES

ber of other cases. This flexibility, in turn, enables manag-
ers and analysts to develop a sense for system dynamics Andrews, B. H. and S. M. Cunningham. 1995. L.L. Bean
and also to proactively answer common senior manage- Improves Call Center Forecasting. Interfaces 25:1-13.
ment questions such as “what would a 10% increase in call Andrews, B. H. and H. L. Parsons. 1989. L.L. Bean Chooses
volume next week do to us?” or “what is the value of add- an Agent Scheduling System. Interfaces 19:1 – 9.
ing an outsourcer to help us during our peak months?” Andrews, B. H. and H. L. Parsons. 1993. Establishing
Telephone-Agent Staffing Levels Through Economic
6 WHAT THE FUTURE HOLDS FOR CALL Optimization. Interfaces 23:14-20.
CENTER SIMULATION Feinberg, R. A., I. Kim, B. Hokama, K. Ruyter, and C.
Keen. Operational Determinants of Caller Satisfaction
Looking out into the future, we see two major trends im- in the Call Center. International Journal of Service
pacting call center simulation. First of all, operational Industry Management 11:131-141.
complexity will continue to grow: more queues, more dif- Garnett, O., A. Mandelbaum, and M. L. Reimann. 2002.
ferent agent schedules, more diverse skilling combinations Designing a Call Center With Impatient Customers.
and routing rules. This will put pressure on analysts to not Manufacturing and Service Operations Management
only build richer models, but also to define output metrics 4:208-227.
that enable them – and their management – to understand Grossman, T. A., D. A. Samuelson, S. L. Oh, and T. R.
the bigger picture as well as the more minute statistics. Rohleder. 2001. Call Centers. In Encyclopedia of Op-
Even in the very simple numerical example above, it is erations Research, ed. S. L. Gass and T. M. Harris,
easy to see how one can become overwhelmed with the 73-76. Norwell: Kluwer Academic Publishers.
sheer volume of numbers that simulation can produce. Hoffman, K. L. and C. M. Harris. 1986. Estimation of a
In addition, as executives begin to understand that the Caller Retrial Rate for a Telephone Information Sys-
call center is a key component in their customer value de- tem. European Journal of Operational Research
livery chain, we foresee an increased desire to understand 27:207-214.
the risks inherent in any particular operational configura- Mabert, V. A. 1985. Short-Interval Forecasting of Emer-
tion. In particular, we see interesting and important oppor- gency (911) Workloads. Journal of Operations Man-
tunities in randomizing not only call arrival patterns and agement 5:259-271.
handling times but also overall call volumes, and using Mandelbaum, A. 2001. Call Center Research Bibliography
techniques from risk analysis and experimental design with Abstracts, Technical Report, Technion, Israel In-
along with simulation models to quantify system capacity stitute of Technology.
and delivery risks. Mandelbaum, A. and N. Shimkin. 2000. A Model for Ra-
Finally, we hope for and expect improvements in the tional Abandonments from Invisible Queues. Queue-
quality of data provided for quantitative analysis. In par- ing Systems, Theory, and Application 36:141-173.
ticular, increased accuracy and detail associated with han- Mehrotra, V. 1997. Ringing Up Big Business. OR/MS
dle time distributions, waiting time distributions, and Today 24:18-24.
abandonment time distributions will lead to better model Pinker, E. and R. Shumsky. 2000. The Efficiency-Quality
inputs and more robust results. Tradeoff of Crosstrained Workers. Manufacturing and
Service Operations Management 2:32-48.
ACKNOWLEDGMENTS Saltzman, R. and V. Mehrotra. 2001. A call center uses
simulation to drive strategic change. Interfaces
The authors would like to thank the call center directors, 31:87-101.
managers, and executives that we have had the chance to Samuelson, D. A. 1999. Predictive Dialing For Outbound
work with over the past several years. Through our profes- Telephone Call Centers. Interfaces 29:66-81.
sional and personal interactions with these overworked and
underappreciated individuals, we have learned a great deal AUTHOR BIOGRAPHIES
about call center operations, management, and data
sources, all of which has contributed greatly to our ability VIJAY MEHROTRA is an Assistant Professor in the
to model and analyze these types of systems. We have College of Business at San Francisco State University.
also gotten a first-hand sense of the pressures that these in- Vijay joined the SFSU faculty in the fall of 2003 after
dividuals work under, and hope that our experience and our over ten years in the operations management consulting
models can help provide insight and support to them. field. Most recently, he was a Vice President with the Solu-
tions Group at Blue Pumpkin Software. Prior to joining
Blue Pumpkin, he was co-founder and CEO of Onward Inc.,
an operations management consulting firm based in Moun-

142
 Mehrotra and Fama

tain View, CA that focuses on the successful application of

Operations Research techniques to business applications.
Vijay’s research interests include applications of sto-
chastic processes and optimization, queueing networks, and
the adoption of models and information technology by indi-
viduals and organizations. Within the simulation area, he has
been actively involved with modeling semiconductor manu-
facturing facilities, electric power production systems, con-
tainer ship traffic, as well as call center operations.
Over the course of his consulting career, Vijay has
worked with a wide variety of clients in many industries,
including EDS, Intuit, Hewlett Packard, Charles Schwab,
AOL, Hewlett Packard, IBM, General Electric, Sykes,
CIGNA, Tyecin Systems, National Semiconductor, and
Remedy.
Vijay holds a B.A. degree in Mathematics, Economics,
and Statistics from St. Olaf College and an M.S. and Ph.D.
in Operations Research from Stanford University. He is the
past President of the Northern California INFORMS chap-
ter, and writes a regular column in OR/MS Today entitled
“Was It Something I Said?” His email address is
<[email protected]>.

JASON FAMA is an analyst and developer in the algo-

rithms group at Blue Pumpkin Software. While at Blue
Pumpkin, Jason has been actively involved in the devel-
opment of forecasting, queueing, and scheduling algo-
rithms, and with effectively embedding efficient simulation
models within Blue Pumpkin's overall workforce optimiza-
tion framework.
Prior to joining Blue Pumpkin, had worked as a re-
searcher Rockwell’s Palo Alto Laboratory. Jason holds a
B.S. degree in Economics and Computer Science from the
University of California at Berkeley. His email address is
<[email protected]>.

143

View publication stats