Table Of ContentModeling and Analyzing IVR Systems, as a
Special Case of Self-services
Research Thesis
SubmittedinPartialFulfillmentofthe
RequirementsfortheDegreeof
MasterofScienceinOperationsResearchandSystemAnalysis
Nitzan Carmeli
SubmittedtotheSenateofthe
Technion-IsraelInstituteofTechnology
Sivan,5775 Haifa June10,2015
ThisResearchThesisWasDoneUndertheSupervisionof
ProfessorHayaKaspiandProfessorAvishaiMandelbaumintheFacultyof
IndustrialEngineeringandManagement.
IwouldliketogratefullythankProfessorHayaKaspiandProfessor
AvishaiMandelbaumfortheirendlessguidanceandsupport,forgivingme
thewonderfulopportunitytoworkwiththemandlearnfromthem. Ithas
beenagreatprivilege,andapleasure.
IwouldalsoliketothankDr. GalitYom-Tovforhervaluableguidance,
ArikSenderovich,YuvalMichaelandAnatBernshteinfortheir
contributiontothiswork,andgreatlythanktheSEELabteam: Ella
Nadjharov,IgorGavakoandDr. ValeryTrofimovforalltheirsupportand
assistance.
TheGenerousFinancialHelpoftheTechnion
isGratefullyAcknowledged.
Contents
1 Introduction 2
2 LiteratureReview 5
2.1 MethodologiesforevaluatingIVRs . . . . . . . . . . . . . . . . . 5
2.2 DesigningandoptimizingIVRs . . . . . . . . . . . . . . . . . . 6
2.3 ModelingaCallCenterwithanIVR . . . . . . . . . . . . . . . . 7
2.4 StochasticSearchinaForest . . . . . . . . . . . . . . . . . . . . 8
2.5 PredictingSearchSuccessinSelf-ServiceSystems . . . . . . . . 9
3 ModelingCustomerFlowasaStochasticSearchonaTree 11
3.1 TheSearchModel . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.1 BuildingBlocks . . . . . . . . . . . . . . . . . . . . . . 12
3.1.2 StateSpaceandSearchProtocol . . . . . . . . . . . . . . 14
3.2 PropertiesofCandidates . . . . . . . . . . . . . . . . . . . . . . 17
3.3 ModelingtheSearchProtocolasaRootedGraph . . . . . . . . . 24
4 AdmissibleTreeCreation 30
4.1 AlgorithmDescription . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 AlgorithmFormulation . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 The Equivalence between Admissible Tree and the Set of Admis-
sibleCandidates . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.4 ProofsofLemmas5–7 . . . . . . . . . . . . . . . . . . . . . . . 41
4.5 ProofsofLemmas8–10 . . . . . . . . . . . . . . . . . . . . . . . 44
5 IndexCalculationsOvertheAdmissibleTree 50
5.1 IndexCalculations . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2 FindingOptimalSearchPolicyAlgorithm . . . . . . . . . . . . . 52
5.2.1 OptimalPolicyAlgorithm . . . . . . . . . . . . . . . . . 54
6 ExploratoryDataAnalysis 56
6.1 DataDescription . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.2 DemandforIVRServices . . . . . . . . . . . . . . . . . . . . . . 59
ii
6.3 CustomerPaths . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.4 SuccessRateofIVRServices. . . . . . . . . . . . . . . . . . . . 65
6.5 FurtherSupporttoanAbandonmentHypothesis . . . . . . . . . . 72
6.5.1 CustomerExperienceEffectonTimeSpentinIVRServices 73
7 ModelImplications 79
7.1 ComparingDifferentIVRDesigns . . . . . . . . . . . . . . . . . 81
7.1.1 EstimatingtheModelParameters . . . . . . . . . . . . . 81
7.1.2 NumericalExample . . . . . . . . . . . . . . . . . . . . 86
7.2 FurtherImplications. . . . . . . . . . . . . . . . . . . . . . . . . 96
8 SummaryandDiscussion 98
8.1 FutureResearch . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
A IVRDataTable 101
B IVRLastServicesDistribution 103
C Measurements-ReadingTimeofMenuOptions 106
iii
List of Figures
1.1 CallcenterwithanIVR . . . . . . . . . . . . . . . . . . . . . . . 3
3.1 Anexampleofarootedtree,representinganIVRsystem . . . . . 15
3.2 An example of a proper candidate represented by a path in GF.
HereGistakenfromFigure3.1. . . . . . . . . . . . . . . . . . . 28
3.3 ApartialexampleofaProperGraphcreatedfromanIVRtreeG. 29
4.1 ExampleofavertexinF4 . . . . . . . . . . . . . . . . . . . . . 33
exc
4.2 ExampleofapartialProperGraphandexcludedpaths . . . . . . 35
4.3 ExampleofapartialAdmissibleTree . . . . . . . . . . . . . . . 37
4.4 ExamplesofProposition3,caseIV . . . . . . . . . . . . . . . . . 39
4.5 Exampleofpathsrepresentingunwantedsequences,resultingfrom
addingverticesinF2 . . . . . . . . . . . . . . . . . . . . . . . 43
exc
4.6 AnexampleofLemma8 . . . . . . . . . . . . . . . . . . . . . . 46
4.7 AnexampleofLemma8 . . . . . . . . . . . . . . . . . . . . . . 47
6.1 ILBankIVRmenu . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.2 ILBankHybridanimation(link) . . . . . . . . . . . . . . . . . . 62
6.3 ILBankNetworkanimation(link) . . . . . . . . . . . . . . . . . 63
6.4 RecentAccountActivityduration,N = 1,983,161 . . . . . . . . 66
6.5 AccountSummaryduration,N = 1,081,992 . . . . . . . . . . . 66
6.6 AccountActivityTodayduration,N = 353,328 . . . . . . . . . . 67
6.7 CreditCardVouchersduration,N = 608,529 . . . . . . . . . . . 67
6.8 IVRservicesduration . . . . . . . . . . . . . . . . . . . . . . . . 68
6.9 Fittingmixtureofdistributions,RecentAccountActivity . . . . . 70
6.10 Fittingmixtureofdistributions,AccountSummary . . . . . . . . 71
6.11 DistributionofthetimeintheIDphase,asafunctionofthenumber
ofpriorcalls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.12 Zoomin-DistributionofthetimeintheIDphase,asafunctionof
thenumberofpriorcalls . . . . . . . . . . . . . . . . . . . . . . 74
6.13 Distributionofthetimein‘RecentAccountActivity’,asafunction
ofthenumberofpriorcalls . . . . . . . . . . . . . . . . . . . . . 75
6.14 Distributionofthetimein‘RecentAccountActivity’,asafunction
ofthenumberofpriorcalls,zoominfrom0to20seconds . . . . 76
iv
6.15 Distributionofthetimein‘RecentAccountActivity’,asafunction
ofthenumberofpriorcalls,zoominfrom45to75seconds . . . . 76
6.16 Distributionofthetimein‘RecentAccountActivity’,asafunction
ofthenumberofpriorvisitstotheservice,acrosscalls . . . . . . 77
6.17 Distributionofthetimein‘RecentAccountActivity’,asafunction
ofthenumberofpriorvisitstotheservice,withinonecall . . . . 78
7.1 Calculatingorganizationalprofit-Simpleexample. . . . . . . . . 80
7.2 OriginalIVRdesign. Boldrepresentsserviceswithpositivereward 89
7.3 ShallowIVRdesign. Boldrepresentsserviceswithpositivereward 90
7.4 DeepIVRdesign. Boldrepresentsserviceswithpositivereward . 91
A.1 IVRdatatable . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
B.1 RecentAccountActivityduration . . . . . . . . . . . . . . . . . 103
B.3 AccountSummaryduration . . . . . . . . . . . . . . . . . . . . . 104
B.5 AccountActivityTodayduration . . . . . . . . . . . . . . . . . . 104
B.7 CreditCardVouchersduration . . . . . . . . . . . . . . . . . . . 105
C.1 TimemeasurementofILBankIVRmenuoptions . . . . . . . . . 108
v
List of Tables
6.1 TotalnumberofIVRcallsbytheiroutcome,May2008toJune2009 57
6.2 Total number of IVR calls, by customer type, May 2008 to June
2009 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.3 IVRservices-relativedemandfrequency,May2008toJune2009 60
6.4 Frequentpaths . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.5 Fittingmixtureofdistributions,RecentAccountActivity . . . . . 70
6.6 Fittingmixtureofdistributions,AccountSummary . . . . . . . . 71
6.7 Statistics of time in the ID phase, as a function of the number of
priorcalls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.1 AveragesojourntimewithinILBankIVR,May2008toJune2009 85
7.2 Average patience, by customer type, based on Khudyakov et al.
[18] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
7.3 Rewardsandcosts,byAksinetal. [2] . . . . . . . . . . . . . . . 87
7.4 Laplacetransformforsuccessfulandunsuccessfulservicedurations 88
7.5 Results,optimalpathsandexpectedutility,Highpriority . . . . . 92
7.6 Results,optimalpathsandexpectedutility,Mediumpriority . . . 93
7.7 Results,optimalpathsandexpectedutility,Lowpriority . . . . . 94
7.8 Numericalexample-averagenumberofrelevantcallsbycustomer
type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.9 Organizationalrevenue . . . . . . . . . . . . . . . . . . . . . . . 95
vi
Glossary
Abbreviation FullForm
IVR InteractiveVoiceResponse
EDA ExploratoryDataAnalysis
Notation1 Explanation
G = (V,E) Arootedtree.
M Thesetoftreeleaves.
s TherootvertexofG = (V,E).
A(i) Thesetofallimmediatesuccessorsofvertexi.
pre(i) Theimmediatepredecessorofvertexi.
dep(i) Depthoftheuniquepathleadingfromtherootstovertexi.
Γ(i) Thesub-treespanningfromvertexi.
M Thesetofleavesinthesub-treespanningfromvertexi.
Γ(i)
c Costperunitoftime.
τ ∼ exp(θ) Customerpatience.
P Successprobabilityofvertexi.
i
r Rewardearnedfromasuccessfulvisitatvertexi.
i
t (i) Timespentinvertexigivenasuccessfulvisit.
ser
t (i) Timespentinvertexigivenanunsuccessfulvisit.
F
T Timespentinvertexi.
i
N(j) The number of times that option (vertex) j was previously ex-
plored.
N(j)
t Theexplorationtimeofedgee = (i,j) ∈ E, giventhatoptionj
i,j
inmenuiwaspreviouslyexploredN(j)times.
Place (j) Theplaceofoptionj inmenui.
i
l Theleftmostsonofvertexi.
i
N Avectorrepresentingthenumberofrepeatedexplorationsofeach
vertex.
R∗(v) Theexpectedutilityofthetreespanningfromvertexv.
R∗,v Theindexofedge(u,v).
u
L (s) TheLaplacetransformofX.
X
1Listedhereareonlynotationswhichappearinmorethanonechapter.Therestareexplainedin
theirrelevantchapter.
vii
Abstract
Callcentersplayaprominentroleintoday’seconomy. Theyserveasthemain
customer contact channel in various enterprises, which makes them highly labor-
intensive operations. Thus, call centers look for means to reduce the number of
agents handling calls, and trying to do so without degrading service level. Inter-
active Voice Response (IVR) systems are presently one of the main self-service
channels employed by call centers. They are used as means to reduce operating
expensesderivedfromagentemploymentcosts.
The goal of our research is to improve and enhance IVR systems, aiming to
create a body of knowledge that will generalize to other self-service systems. We
model customers flow within an IVR system as a stochastic search in a directed
tree. ThesearchgoalistofindtheoptimalpathontheIVRtree,whichwillresult
in maximal expected discounted utility for customers. We show that a calculable
index can be assigned to each feasible option, and the optimal policy is to choose
theoptionwiththehighestindexateachstage.
OurmodelbuildingblockswerecreatedthroughanExploratoryDataAnalysis
(EDA) of real IVR transactions, in a call center of a large Israeli bank. The EDA
revealed interesting phenomena regarding customer abandonments and learning
withintheIVR.
OurworkenablesthecomparisonbetweenalternativeIVRdesigns,bothfrom
the customer and the enterprise point of view. This complements related research
inotherfields,suchasHuman-Factor-EngineeringandTelecommunication.
The model for IVR systems that we developed can easily be implemented to
otherself-servicesystemssuchasInternetwebsiteswhichhavebecomeprevalent.
1
Chapter 1
Introduction
Call centers play a prominent role in today’s economy. Indeed, they serve as
the main customer contact channel in various enterprises [1], public or private,
product-based or service-based. Call centers are also highly labor-intensive; they
employ sometimes hundreds, or even thousands of Customer Service Representa-
tives(CSRsorAgents)tohandleincomingcalls. Typically,60%-70%oftheover-
all operating expenses of call centers are derived from agent employment costs
[10]. Reducing the number of agents handling calls, without degrading service
level, is thus of interest and importance, and enabling customers to self-serve is
one of the basic means for doing so. As customers self-serve, the agent workload
isbeingreduced,andlessagentsarerequiredinordertomaintainacertainservice
level. Interactive Voice Response (IVR) systems, also known as Voice Response
Units (VRU), are one of the main self-service channels [18], along with Internet
websitesanddesignatedsmart-phoneapplications.
IVR systems, if properly designed, can increase customer satisfaction and
loyalty, cut staffing costs and increase revenue by extending business hours and
market reach [3]. Poorly designed IVR systems, on the other hand, will cause the
oppositeeffectandleadtodissatisfiedcustomers,increasedcallvolumeandmight
even increase agent turnover, as agents would serve frustrated customers [7]. A
PurdueUniversitystudyshowedthatmorethan90%ofUSconsumersareforming
their image on a certain company based on their experience with its call center.
Furthermore, more than 60% stopped using the products of a company in which
they had a negative call center experience [7]. Since the IVR system is the front
gateofmostcallcenters,havinganeffective,efficient,andcustomer-friendlyIVR
systemisextremelyimportant.
The goal of our research is to improve and enhance IVR systems, aiming to
generalizetootherself-servicesystems. Todoso,wemodelandanalyzecustomer
flow within an IVR system. The model building blocks were established and in-
spired by an Exploratory Data Analysis (EDA) of real IVR transactions in a call
centerofalargeIsraelibank,basedonmorethanoneyearofdatawhichincludes
millions of calls. The theoretical basis for our model then relies on the work of
2
Description:Thus, call centers look for means to reduce the number of agents handling Our work enables the comparison between alternative IVR designs, both from Interactive Voice Response (IVR) systems, also known as Voice Response ac.il/serveng/References/CCA-Patience.pdf. 66(3):215–237.