Intelligent Procedures for IntraDay Updating of Call Center Agent Schedules University of Montreal C

1
Intelligent Procedures for Intra-Day Updating of
Call Center Agent Schedules University of
Montreal Call Center Workshop, May 2006
Vijay Mehrotra and Ozgur OzlukDepartment of
Decision Sciences, College of BusinessSan
Francisco State University
2
Presentation Roadmap

• Who is this Guy?
• Customer Conversations / Embedded Problems
• Intra-Day Re-Scheduling Framework
• Literature
• Components
• Numerical Experiment and Results
• Questions and Extensions

3
About Vijay

• PhD in OR, Stanford University, 1992
• 1993 1994 Consultant, DFI
• 1994 - 2002 Co-Founder and CEO, Onward Inc.
• 2002 - 2004 Vice President of Solutions, Blue
  Pumpkin Software

4
More Than 1200 Companies Depend on Blue
Pumpkin/Witness For Workforce Management Software
Insurance Lending
Banking Brokerage
Outsourcers
Retail Catalog
Communications
Technology
Travel Transportation
Healthcare
Consumer Goods
Manufacturing
5
About Vijay

• Fall 2003 Radical Portfolio Adjustment
• Return to Academic World
• SFSU Dept of Decision Sciences, College of
  Business
• Full-Time Tenure Track Position
• Teach Courses in Statistics, Operations, Quality,
  and Project Management to Undergraduates and MBAs
• Still in Real World
• Regular Stream of Consulting Projects
• Focus on Call Center Operations, Enterprise
  Software, and Revenue Management
• Thrust into Brave New World Spring 2004
• Became First-Time Father
• Moved to East Bay from SF

6
Presentation Roadmap

• Who is this Guy?
• Customer Conversations / Embedded Problems
• Intra-Day Re-Scheduling Framework
• Literature
• Components
• Numerical Experiment and Results
• Questions and Extensions

7
Call Center WFM The Right Number of Agents
Working at the Right Times to Deliver the Right
Queues Not So Hard, Right?

• Several Hundred Papers in the Academic Literature
  on Different Aspects of the Call Center WFM
  problem
• Gans, Koole, and Mandelbaum (MSOM 2003) is an
  excellent literature survey
• But We Still Have Many Managers and Executives
  with Real, Unsolved Call Center WFM Problems

8
So Many Improvements to ConsiderThe Exploding
Head of the CC Manager

• MORE Routing Complexity
• Skill-Based Routing
• Multiple Customer Channels
• Inbound/Outbound Blended
• MultiSite / Outsourcing
• MORE Demand Uncertainty
• New Policies/Processes for Existing Businesses
• New Businesses/Services
• MA Activity
• New Operating Hours
• Increased Service Level Goals
• Cross-Channel Dynamics
• MORE Pressure / Urgency
• Tighter Budgets
• Solve the Problem Now

9
Vijays Grand Theory of EvORything
Optimization/ Performance Model
10
The Focus of This Paper Short-Term Decision
Making Based on Newly Available Information
Strategic Cycle
Hiring and Training Plan ? Available Staff
Tactical Cycle
Schedules New Call FCs ? Plan for Future
Week(s)
Real Time Cycle
Adjustments to Schedules Adjustments to Forecasts
Historical Data
11
Conversation 1 CustomerVP of Operations for
Huge Division of Massive FinSvcs Conglomerate

• Vijay So where else do you guys need help?
• Customer (upbeat) We do our forecasts and
  schedules about a month ahead of time.
• C But things are changing all the time, so we
  are monitoring and updating our forecasts all
  the time, every single day.
• C Then, we react by trying to commit and
  de-commit resources as best we can ratchet
  outsourcers up and down, offer our employees OT
  or VTO, offer more hours to our PT staff
• C Last year, we estimate that we saved about
  8mm doing this.
• V (nervously) So whats the problem with
  that?
• C First of all, we have no idea if were doing
  well or not, and we think we might be leaving a
  lot of money on the table.
• C Secondly, its all one big email nightmare,
  and it drives our ops staff nuts trying to keep
  all of it straight.
• V Hmmm.Thanks

12
Conversation 2 CustomerVP of Finance for Big
Division of Large Financial Svcs Conglomerate

• Customer (abruptly) How does your system
  quantify the risk?
• Vijay What do you mean by risk?
• C From what youve said, you take my forecasts
  and my service goals and come up with a number of
  agents for each 15-minute interval. Then, your
  scheduling algorithm tries to match that target.
• V (excited customers never get this!) Thats
  right! Youve got it!!
• C So what percentage of the time will we
  actually meet our goals with that staffing
  plan?
• V Well, what youd need to do is to do a Monte
  Carlo simulation on your forecasts and do a
  bunch of replicationsAnd the answer depends on
  how you respond to different levels of demand,
  and on how accurate your forecasts are
• C Your product doesnt do that for us?
• V Uh, no. But Ill put it on the list

13
Presentation Roadmap

• Who is this Guy?
• Customer Conversations / Embedded Problems
• Intra-Day Re-Scheduling Framework
• Literature
• Components
• Numerical Experiment and Results
• Questions and Extensions

14
Framework for Intra-Day Schedule Updating
15
Key Relevant Literature Workload FC and Update

• Identifying and Modeling Arrival Rates Per Period
  as Random Variables
• Thompson (1999), Chen Henderson (2001)
• Ross (2001), Jongblooed Koole (2001)
• Whitt (2004)
• Which Are Correlated with One Another
• Brown et al (2002)
• Avramidis et al (2004)
• Steckley et al (2004)

16
Key Relevant Literature RT Schedule Adjustments

• Models for Real Time Schedule Adjustments for
  Service Systems
• Thompson (1999)
• Hur, Mabert, and Bretthauer (2004)
• Easton and Goodale (2005)
• Surprisingly Small List
• Absent from the Literature RT Schedule Updating
  Papers in the Context of Call Centers

17
Framework for Intra-Day Schedule Updating
18
Step 0 Operating Parameters Initial Schedules
Initial Schedule Assignments (Typically 1-4
Weeks Prior)
19
Step 1 Update Workload Forecast and Demand
for Agents
Actual Call Volumes (1,2,..u-1)
20
Step 1 Update Workload Forecast
As in (Whitt 99) and (Avramidis et al 2005), we
model arrival process as NHPP with Random Arrival
Rate L(s) H (?(s) s gt0), where ? is
piecewise constant on intervals 1,2,T and H is
a (scalar) Random Variable with EH 1 ? EL
?
21
Step 2 Update Demand for Agents

• Use Standard Queueing Equation for Translation
  (minimum s to satisfy SL goals for M/M/s queue)
  based on updated forecasts to determine
  incremental agent needs dt for tu,u1, T

22
Step 3 Update Agent Schedules for periods uT
Initial Schedule Assignments
Individual Agents Availability Information
Updated Agent Schedules for u, u1,T
Incremental Agent Reqs (u, u1, T)
23
Step 3 Update Agent Schedules for periods uT
24
Step 3 Update Agent Schedules for periods uT
Dimensionality of IP is Quite Small TxN Integer
Variables, Tx(TN) Constraints
25
Step 3 Update Agent Schedules for periods uT

• When Arrival Rate Variability Dominates
  Attendance
• Special Cases
• Strictly overstaffed
• Hlt1 ? dt lt0 for all tu,u1, T
• Address with Voluntary Time Off and/or Release of
  Contracted Agents
• Strictly understaffed
• Hgt1 ? dt gt0 for all tu,u1, T
• Address with Holdover OT and Call-In OT

26
Presentation Roadmap

• Who is this Guy?
• Customer Conversations / Embedded Problems
• Intra-Day Re-Scheduling Framework
• Literature
• Components
• Numerical Experiment and Results
• Questions and Extensions

27
Experimental Framework

• Goal Test Methodology on Real Call Center Data
  to Understand Dynamics
• Model From Saltzman 2005
• Sales-and-Service Call Center in Travel Industry
• Relatively Small Call Center
• Roughly 360 agent hours/day
• Mixture of FT and PT Agents

28
Experimental Framework
Build initial schedules Based on l values
(expected arrival rates)
Choose a value for H, And simulate arrivals for
Periods 1,2,..u-1
29
Results for Overstaffed Cases (0.5 lt H lt 1)
Lesson After recognizing that original FCs are
too high, our Update Methodology delivers
desired SLs with less staff/lower cost
30
Results for Understaffed Cases (1 lt H lt 1.5)

• Key Lessons
• Not responding to new information is very
  damaging to service quality
• When H is large, Schedule Updating cannot fully
  make up for initial poor performance during first
  four hours

31
Results for Understaffed Cases (1 lt H lt
1.5) The Rest of the Story

• When staffing based on expected l, cannot meet
  goals easily without Call-in OT
• ? Given plans to update, where should initial
  staffing be set?
• ? What structure for contingent resource
  contracts makes sense given different arrival
  rate uncertainties?

32
Presentation Roadmap

• Who is this Guy?
• Customer Conversations / Embedded Problems
• Intra-Day Re-Scheduling Framework
• Literature
• Components
• Numerical Experiment and Results
• Questions and Extensions

33
Questions and Ideas? Please Call or Email!
Vijay Mehrotra Ozgur Ozluk Department of
Decision Sciences San Francisco State
University vjm_at_sfsu.edu / 650-465-8443 ozgur_at_sfsu.
edu / 415-338-1007
34
Extension to this Research Currently in Progress

• Almost all Call Center Research to date assumes
  that arrivals in a period is Poisson
  distributed
• Data often strongly refutes this
• e.g., mean 2000, std dev 500 or more
• Model arrival process as B (?(t) t 0), (Whitt
  99), where ? is piecewise constant

35
Random Arrival Rates A Graphical View
?(t)
t
36
Extension to this Research Currently in Progress

• Where to set initial staffing?
• Hypothesis Performance (and Cost-Effectiveness)
  can be improved by accounting for Arrival Rate
  Variability in setting initial staffing levels
• Method Use Analytic Approximations from
  (Steckley, Henderson, and Mehrotra 2005) to
  determine of agents needed to achieve
  particular SL when creating initial weekly
  schedule

Initial Call And AHT Forecasts
M/M/s Queueing Equations
Initial Agent Requirements Per Period