Call Center optimization - step by step

Erik Chromy, Matej Kavacky and Ivan Baronak

Faculty of Electrical Engineering and Information Technology STU in Bratislava
Bratislava, Slovakia
Email: [email protected], [email protected], [email protected]

Abstract—The paper deals with the contact center modeling parameters should be taken into consideration by call routing
with emphasis on the optimal number of agents. The contact algorithms:
center belongs to the queueing systems and its mathematical
model can be described by various important parameters. The • caller requirements for service,
Erlang C formula tends to be suitable tool for the modeling of • agent profile,
QoS parameters of contact centers. In our paper, we propose two • state of the agent (free or busy),
parameters: downtime and administrative task duration. These • state of the queue (number of callers, average waiting
parameters are useful for better determination of the optimal
time for service),
number of contact center agents. Based on these parameters
we propose a mathematical model for contact centers and a six • system load in real-time,
step estimation model to compute the optimal number of contact • alternative resource (in the case of overload).
center agents.
Keywords—contact center; Erlang C formula; optimization;
quality of service

I. I NTRODUCTION
A contact center is in common a typical example of a
queueing system [1]–[5]. In a contact center a given number
of calls occurs within a defined time interval. These calls
originate randomly and are independent from each other.
Queueing systems often have to deal with the predetermination
of Quality of Service [6]–[12]. For the description and dimen-
sioning of queueing systems various mathematical models are
used e.g. Markov models (with various number of servers,
Fig. 1. ACD routing.
with or without a queue), Erlang formulas, Jackson networks
and non-Markovian models [13]–[15].
A contact center system consists of various components Although contact centers are effectively utilizing resources
where IVR (Interactive Voice Response) system and ACD we have to assume a situation where no free agents are
(Automatic Call Distribution) [16]–[18] can be considered as available to the system. Hence we have to deal with queueing
a basis of contact centers. in the case of ACD. Private Branch Exchange (PBX) inserts all
The IVR system automates call processing and many of incoming calls into a queue. The required service is selected at
time-consuming tasks (performed by agents and supervisors first (selection is based on the IVR) and based on the selection
on daily basis). Such system allows: inside the IVR system the call is routed to the queue for a given
service. The call is enqueued according to FIFO algorithm.
• automatic announcements of a welcome message and
Special case will occur when the caller is marked as special.
estimated waiting time for the caller,
In this case the call is put on the beginning of the queue,
• automatic performs various actions based on the es-
instead of the end of the queue.
timated waiting time, e.g. identification of the caller
The calls remain in the queue:
through PIN (Personal Identification Number),
• a caller can leave a message for the agent with a callback • until any agent of the calling group will be free, or

inquiry, • until expiration of the time interval dedicated for waiting

• reading of selected text from the database, of the caller (parameter is set by the contact center
• selection of the communication language, etc. manager), or
• until the call is terminated by the caller.
ACD belongs to the basic software needed for contact
center realization (Figure 1). It offers various and sophisticated
II. Q UEUEING SYSTEM AND PERFORMANCE
call routing functions, queue creation and productivity control
MEASUREMENT
of agents handling these calls. Agents with the same or
similar profile form service group. Call routing interconnects a The contact center is typical example of the queueing
calling customer with a given agent. The following important system with the following characteristics:

978-1-5090-3982-1/17/$31.00 ©2017 IEEE 119 TSP 2017

 • at the contact center input the requests from customers • waiting time in the queue - represents the time that the
are randomly occurring in random time intervals, request spent in queue before handling (then response
• the handling time of each request is random a variable time = waiting time in the queue + service time),
(hence, it is necessary to determine the average time for • queue length - number of requests in the queue,
handling request by agent), • number of requests in the system - total number of
• contact center can have more service groups, each with requests in the queueing system (i.e. requests waiting and
custom random parameters (number of requests on input being processed).
and average handling time),
III. C ONTACT CENTER MODEL
• the output of the system is the request handled by an
agent. Our proposed model of the contact center consist of IVR
system and of n service groups.
The common model of queueing system consists of three
The parameters of the IVR system are:
basic parts: process of arrivals, storing and server (handling).
• r - number of calls from range <0,1> handled automati-
The most used model for the arrival process is Poisson arrival
process [19]. Handling time is a random variable and it is cally without need of the contact center agent interaction.,
• c - number of calls incoming to the contact center during
independent from arrivals. The system inserts customers into
the memory until servers become free. There are two extreme the busy hour,
• 1/μ1 - average handling time for one request in IVR
cases. The storage can be such large, that it can be regarded as
infinite. In the second case the storage is only for customers in system.
the process of handling. Queueing systems are used in many For the service groups the following parameters are used:
fields (services, industry, storage and maintenance, etc.), but • n - number of service groups,
their application is very important specially in the field of • 1/μ2,i - average handling time for one calling request in
telecommunications. service group i,
In order to judge individual queueing systems, it is nec- • PC - upper bound from range <0,1> for request enqueue,
essary to implement the performance measurement (which • λ - number of requests incoming to the given service
will describe such system). Because the queueing system group during busy hour from IVR system,
is a dynamic system, the performance measurement may • AWT - acceptable waiting time (converted to hours),
vary over time. It holds, that the system is in steady state, • GoS - request rate from range <0,1> , which the agent
when all transients in the system are finished, the system is must accept within defined AWT period,,
stabilized and the performance measurement values are time • Tp,i - downtime (represents time dedicated for example
independent. The system is in so called statistical equilibrium, for the hygienic break) of service group i in minutes,
i.e. the arrival rate of requests into the system is equal to the • ac,i - rate of calls from range <0,1>, which require
rate at which the requests leave the system. Such system is administrative tasks (activity which the agent has to
called stable system. perform after call) in service group i,
The most important parameters of performance measure- • Tac,i - average time of the administrative task duration
ment are: (in minutes) which the agent has to perform for one call
• probability of number of requests in the system - the in service group i.
system can be described through the probability vector In our model, we propose two new parameters: downtime
of number of requests in the system. The average value and administrative task duration. After the contact center
can be determined from such vector, model definition, two cases of requests distribution can be
• utilization ρ - if the queueing system consists of one considered and modeled:
server, then the utilization is the fraction of time in
which the server is active. If there are no limitations a) Output requests from IVR systems are uniformly
on the number of requests in one queue, then the server distributed among service groups. Then for number of
utilization is as follows: requests incoming to each service group we have following
arrivalrate λ equation:
ρ= = (1)
servicerate μ λ(1 − r)
λ= (3)
In the case when the queueing system consists of multiple n
servers m, then for ρ the following equation is valid: b) Output requests are non-uniformly distributed among
service groups. In this case we have to define probability
λ
ρ= (2) (p0,1 till p0,n ) by which the service group is selected. These
mμ probabilities are set by contact center manager before start of
• throughput - throughput of simple queueing system is operation based on the detailed analysis of expected interest
defined as average number of the handled requests (i.e. on services from callers (forecast). Later these probabilities
leaving rate from the system), should be set more precisely by measuring of real traffic in
• response time - total time the request spent in the system, contact center.

120
 As we have n service groups, then for particular probabili-
ties following condition must be valid: On the base of previous considerations, we propose to
n
use the following equation, that is valid for the average
 number of served callers during the busy hour (taking into
p0,i = 1 (4)
i=1
acount downtime and administrative task):
Then for number of the requests incoming into i-th service μ2,i (60 − Tp,i )
group following equation must be valid: μ(Tp +ac ),i = (10)
60 + μ2,i .ac,i .Tac,i

λi = λ.(1 − r)p0,i (5) After estimation of the number of agents of given service
group mi , another QoS parameters should be verified also:
For total average time the caller will spent in contact center
• load of agent in given i-th service group during busy hour
system (e.g. IVR + i-th service group) we have the following
is very interesting indicator:
equation:
λi
1 1 1 1 ρ= (11)
Ti = T1 + T2,i = + + W2,i = + + W2,i mi .μ(Tp +ac ),i
μ1 μ2,i μ1 μ(Tp +ac ),i
(6) • also the average time the customer will spent in queue of
In our mathematical model, we propose to use the parameter given i-th service group can be verified:
μ(Tp +ac ),i instead of parameter μ2,i . Because, this parameter
takes downtime and administrative task into acount. PC
W2,i = (12)
The value of W2,i parameter depends on used mathematical μ(Tp +ac ),i .(mi − λi
μ(Tp +ac ),i )
model: Erlang C equation or Markov model M/M/m/K.
Based on analyses in papers [20]–[22], the Erlang C • further, the average number of callers in queue of i-th
formula is a very suitable tool for modeling of important service group is interesting for contact center manager:
QoS parameters of contact centers. It also allows to calculate
the optimal number of agents for given i-th service group. λi
Q2,i = λi
.PC (13)
Consequently, after defining of all parameters for the μ(Tp +ac ),i .(mi − μ(Tp +ac ),i )
calculation of required number of agents we have:
• for the average time the caller will spent in contact center
a) equation (7): (IVR + i-th service group) following equation is valid:

mi (λ.X)mi
λ m!(mi −λ.X) 1 PC
PC (mi , )= (7) T2,i = + λi
(14)
μ(Tp +ac ),i m
j −1
(λ.X)j mi (λ.X)mi
μ(Tp +ac ),i μ(Tp +ac ),i .(mi − μ(Tp +ac ),i )
j! + m!(mi −λ.X)
j=0
• the contact center manager can be also interested in
where the parameter X is represented by equation (8): probability that in i-th service group is just given number
of callers k:
60 + μ2,i .ac,i .Tac,i
X= (8)
μ2,i (60 − Tp,i ) ⎧ λi
⎪
⎨ P . ( μ(Tp +ac ),i )
k

0,i k! 0 ≤ k ≤ mi
b) or equation (9) if we have GoS parameter defined: Pk,i = λ m
)k m i i
⎪
⎩ ( m .μ i
i (Tp +ac ),i

AW T.λ − ln( 1−GoS P0,i . k ≥ mi

PC )
mi !
mi = μ2,i (9)
AW T. X (15)
where the value of the parameter λ in equations (7) and
(9) depends on the assumed distribution of output requests where P0,i represents the probability that i-th service group
from IVR system - uniform distribution (3), or non-uniform is empty (without requests) and following equation is valid:
distribution (5). In the case of equation (7) the required
number of agents can be obtained by iteration. m
i −1 (
λi i
( μ(T λ+a ) mi
μ(Tp +ac ),i )
i −1
p c ),i 1
P0,i = + λi
For calculating of required number of agents according i=0
i! mi ! 1− mi .μ(Tp +ac ),i
to equation (7) the parameters λ, 1/μ1 , n, 1/μ2,i , PC must be (16)
known. For equation (8) we also need to know the parameters From the view of contact center manager - if any of the QoS
GoS and AWT. For other parameters, if they are not known, parameters render unacceptable values, the number of agents
they get value 0. should be increased.

121
IV. R ECOMMENDED METHOD OF DETERMINATION OF THE through other QoS parameters (e.g. average waiting time in the
OPTIMAL NUMBER OF CONTACT CENTER AGENTS – STEP queue, overall average time in the system) and then eventually
BY STEP to add another agent into the service group. Our recommended
Basing on the information in chapter III., we recommend six steps facilitates and accelerates the process of optimal
to use the following six steps for determination of the optimal number of the contact center agents determination.
number of contact center agents: The mathematical model of the contact center can be
extended with repeated calls. The contact center manager can
Step 1: Estimation of parameters λ, 1/μ1 , 1/μ2,i , PC be also interested in situation when a caller tries to call again.
or λ, 1/μ1 , n, 1/μ2,i , PC , GoS and AWT. Parameter 1/μ1 Such calls increase the number of incoming calls at the input
means that the contact center will consists also of IVR system. to the contact center and thus have an impact on the number
of required agents. Everything would depend on the setting
Step 2: Calculation of parameter λ. If the output requests of the percentage of callers who will try to call again.
from IVR system are equally distributed among the service
groups, then we use equation (3). If the distribution of calls Acknowledgement
is done non-uniformly, then we use equation (5). In this case
also the probabilities of transitions between states of IVR This article was created with the support of the Ministry
tree have to be defined. of Education, Science, Research and Sport of the Slovak
Republic within the KEGA agency project - 007STU-
Step 3: If we consider downtime and/or administrative 4/2016 Progressive educational methods in the field of
work in the contact center, the parameter 1/μ2,i must be telecommunications multiservice networks and VEGA agency
modified into a new parameter μ(Tp +ac ),i according to project - 1/0462/17 Modeling of qualitative parameters in
equation (8). In this case parameters Tp,i , ac,i and Tac,i must IMS networks.
be defined. If the outgoing calls from the IVR system will
R EFERENCES
be distributed non-iniformly, the calculation according to
equation (10) must be done for each service group individually. [1] G. Bolch, “Queueing Networks and Markov Chains - Modeling and
Performance Evaluation with Computer Science Applications,” 2nd ed.,
New Jersey: John Wiley & Sons, 2006.
Step 4: Calculation of required number of agents. We [2] A. Budhiraja, A. Ghosh and X. Liu, “Scheduling control for Markov-
can use equation (7), or if we have set the parameters GoS modulated single-server multiclass queueing systems in heavy traf-
fic,” Queueing Systems, vol. 78, no. 1, pp. 57–97, 2014.
and AWT, then we will use the equation (9). As a result we [3] B. Buke and H. Chen, “Stabilizing policies for probabilistic matching
have the required number of agents m for each service group. systems,” Queueing Systems, vol. 80, no. 1-2, pp. 35–69, 2015.
[4] V. Shah and G. de Veciana, “Asymptotic independence of servers
activity in queueing systems with limited resource pooling,” Queueing
Step 5: Checkout calculation of QoS parameters of contact Systems, vol. 83, no. 1, pp. 13–28, 2016.
center according to equations (11), (12), (13), (14) and (15) [5] M. Budhiraja, “Invariance of workload in queueing systems,” Queueing
for each service group. In the case of non-satisfied values of Systems, vol. 83, no. 1, pp. 181–192, 2016.
[6] A. Kovac, M. Halas and M. Orgon, “E-model MOS estimate improve-
parameters (e.g. too long waiting time in the waiting queue or ment through jitter buffer packet loss modelling,” Advances in Electrical
too many customers in the waiting queue, etc.) it is necessary and Electronic Engineering, vol. 9, no. 5, pp. 233–242, 2011.
to add another agent into given i-th service group and repeate [7] M. Halas, A. Kovac, M. Orgon and I. Bestak, “Computationally efficient
this step 5. E-model improvement of MOS estimate including jitter and buffer
losses,” in 35th International Conference on Telecommunications and
Signal Processing, TSP 2012, Prague, 2012, pp. 86–90.
Step 6: As a result we have the optimal number of [8] M. Voznak, J. Rozhon, F. Rezac and E. Gresak, “VoIP Call Quality
agents of the contact center for n service groups. Assessment based on RPROP Neural Networks,” in XI International
Symposium on Telecommunications BIHTEL, Sarajevo, 2016.
[9] M. Halas and S. Klucik, “Modelling the probability density function
For more detailed model it is possible to define PC , of IPTV traffic packet delay variation,” Advances in Electrical and
GoS and AWT parameters particularly for each service group. Electronic Engineering, vol. 10, no. 4, pp. 259–263, 2012.
[10] R. Roka and F. Certik, “Simulation and analysis of the signal trans-
V. C ONCLUSION mission in the optical transmission medium,” in 5th International
Conference on Simulation and Modeling Methodologies, Technologies
In our model we proposed two new parameters: downtime and Applications, SIMULTECH 2015, Colmar, 2015, pp. 219–226.
and administrative task duration. These parameters are useful [11] F. Certik and R. Roka, “Possibilities for Advanced Encoding Techniques
at Signal Transmission in the Optical Transmission Medium,” Journal
for better determination of the optimal number of contact of Engineering (United States), vol. 2016.
center agents. The output of the paper is a mathematical model [12] J. Frnda, M. Voznak and L. Sevcik, “Impact of packet loss and delay
of contact center and a six step recommendation process for variation on the quality of real-time video streaming,” Telecommunica-
tion Systems, vol. 62, no. 2, pp. 265–275, 2016.
determination of the optimal number of contact center agents. [13] A. Brezavscek and A. Baggia, “Optimization of a Call Centre Per-
As we have shown, Erlang C formula is an appropriate tool formance Using the Stochastic Queueing Models,” Business Systems
for modeling of important QoS parameters of contact centers. Research, vol. 5, no. 3, pp. 6–18, 2014.
[14] S. Ding, G. Koole and R. D. Mei, “On the Estimation of the True
When the required number of agents in each service group is Demand in Call Centers with Redials and Reconnects,” European
determined, it is possible to verify the accuracy of the result Journal of Operational Research, vol. 246, no. 1, pp. 250–262, 2015.

122
[15] J. Zan, J. Hasenbein and D. Morton, “Asymptotically optimal staffing of
service systems with joint QoS constraints,” Queueing Systems, vol. 78,
no. 4, pp. 359–386, 2014.
[16] W. Chan, G. Koole and P. Ecuyer, “Dynamic Call Center Routing
Policies using Call Waiting and Agent Idle Times,” Manufacturing &
Service Operations Management, vol. 16, no. 4, pp. 544–560, 2014.
[17] B. Legros, O. Jouini and G. Koole, “Optimal scheduling in call centers
with a callback option,” Performance Evaluation, vol. 95, pp. 1–
40, 2016.
[18] G. Koole, B. Nielsen and T. Nielsen, “Optimization of overflow policies
in call centers. Probability in the Engineering and Informational Sci-
ences,” Cambridge University Press, vol. 29, no. 3, pp. 461–471, 2015.
[19] J. F. Hayes, “Modeling and Analysis of Telecommunications Net-
works,” New Jersey: John Wiley & Sons, 2004.
[20] E. Chromy, T. Misuth and M. Kavacky, “Erlang C Formula and Its Use
In the Call Centers,” Advances in Electrical and Electronic Engineering,
vol. 9, no. 1, pp. 7–13, 2011.
[21] E. Jouini, G. Koole and A. Roubos, “Performance Indicators for Call
Centers with Impatience,” IIE Transactions, vol. 45, no. 3, pp. 341–
354, 2013.
[22] J. P. C. Blanc, “Queueing Models: Analytical and Numerical Methods.
Course 35M2C8,” Departement of Econometrics and Operations Re-
search, 226 p., 2011.

123