Service Engineering: Data-Based Science & Teaching in support of Service Management Avishai Mandelbaum Technion, Haifa, Israel http://ie.technion.ac.il/serveng Based on joint work with Sergey Zeltyn, . . . Technion SEE Center / Lab: Paul Feigin, Valery Trofimov, RA’s, . . . 1 Main Messages 1. Simple Useful Models at the Service of Complex Realities. Note: Useful must be Simple; Simple often rooted in Deep analysis. 2. Data-Based Research & Teaching is a Must & Fun. Supported by DataMOCCA = Data MOdels for Call Center Analysis. Initiated with Wharton, developed at Technion, available for adoption. 3. Back to the Basic-Research Paradigm (Physics, Biology, . . .): Measure, Model, Experiment, Validate, Refine, etc. 4. Ancestors & Practitioners often knew/apply the “right answer": simply did/do not have our tools/desire/need to prove it so. Supported by Erlang (1915), Palm (1945),..., thoughtful managers. 5. Scientifically-based design principles and tools (software), that support the balance of service quality, process efficiency and business profitability, from the (often-conflicting) views of customers, servers, managers: Service Engineering . Background Material (Downloadable) I Technion’s ‘‘Service-Engineering" Course (≥ 1995): http://ie.technion.ac.il/serveng I Gans (U.S.A.), Koole (Europe), and M. (Israel): “Telephone Call Centers: Tutorial, Review and Research Prospects." MSOM, 2003. I Brown, Gans, M., Sakov, Shen, Zeltyn, Zhao: “Statistical Analysis of a Telephone Call Center: A Queueing-Science Perspective." JASA, 2005. I Trofimov, Feigin, M., Ishay, Nadjharov: "DataMOCCA: Models for Call/Contact Center Analysis." Technion Report, 2004-2006. I M. “Call Centers: Research Bibliography with Abstracts." Version 7, December 2006. 3 The First Prerequisite: Data & Measurements Empirical “Axiom": The data one needs is never there for one to use – always problems with historical data (eg. lacking, contaminated, averaged, . . .) Averages Prevalent. But I need data at the level of the Individual Transaction: For each service transaction, its operational history – time-stamps of events. (Towards integrating with marketing / financial history.) Sources: “Service-floor" (vs. Industry-level, Surveys, . . .) I I I I Administrative (Court, via “paper analysis") Face-to-Face (Bank, via bar-code readers) Telephone (Call Centers, via ACD) Future: Hospitals (via RFID) Measurements: Face-to-Face Services 23 Bar-Code Readers at a Bank Branch Bank – 2nd Floor Measurements 5 Telephone Service:Call-by-Call Call-by-Call Data Measurements: Telephone Data (Log-File) vru+line call_id customer_id priority type date vru_entry vru_exit vru_time q_start q_exit q_time outcome ser_start ser_exit ser_time server AA0101 44749 27644400 2 PS 990901 11:45:33 11:45:39 6 11:45:39 11:46:58 79 AGENT 11:46:57 11:51:00 243 DORIT AA0101 44750 12887816 AA0101 44967 58660291 1 2 PS 990905 14:49:00 14:49:06 6 PS 990905 14:58:42 14:58:48 6 14:49:06 14:53:00 234 14:58:48 15:02:31 223 AGENT 14:52:59 14:54:29 90 AGENT 15:02:31 15:04:10 99 ROTH ROTH AA0101 44968 0 0 NW 990905 15:10:17 15:10:26 9 15:10:26 15:13:19 173 HANG 00:00:00 00:00:00 0 NO_SERVER AA0101 44969 63193346 2 PS 990905 15:22:07 15:22:13 6 15:22:13 15:23:21 68 AGENT 15:23:20 15:25:25 125 STEREN AA0101 44970 0 0 NW 990905 15:31:33 15:31:47 14 00:00:00 00:00:00 0 AGENT 15:31:45 15:34:16 151 STEREN AA0101 44971 41630443 2 PS 990905 15:37:29 15:37:34 5 15:37:34 15:38:20 46 AGENT 15:38:18 15:40:56 158 TOVA AA0101 44972 64185333 2 PS 990905 15:44:32 15:44:37 5 15:44:37 15:47:57 200 AGENT 15:47:56 15:49:02 66 TOVA AA0101 44973 3.06E+08 1 PS 990905 15:53:05 15:53:11 6 15:53:11 15:56:39 208 AGENT 15:56:38 15:56:47 9 MORIAH AA0101 44974 74780917 2 NE 990905 15:59:34 15:59:40 6 15:59:40 16:02:33 173 AGENT 16:02:33 16:26:04 1411 ELI AA0101 44975 55920755 2 PS 990905 16:07:46 16:07:51 5 16:07:51 16:08:01 10 HANG 00:00:00 00:00:00 0 NO_SERVER AA0101 44976 0 0 NW 990905 16:11:38 16:11:48 10 16:11:48 16:11:50 2 HANG 00:00:00 00:00:00 0 NO_SERVER AA0101 44977 33689787 2 PS 990905 16:14:27 16:14:33 6 16:14:33 16:14:54 21 HANG 00:00:00 00:00:00 0 NO_SERVER AA0101 44978 23817067 2 PS 990905 16:19:11 16:19:17 6 16:19:17 16:19:39 22 AGENT 16:19:38 16:21:57 139 TOVA AA0101 44764 0 0 PS 990901 15:03:26 15:03:36 10 00:00:00 00:00:00 0 AGENT 15:03:35 15:06:36 181 ZOHARI AA0101 44765 25219700 2 PS 990901 15:14:46 15:14:51 5 15:14:51 15:15:10 19 AGENT 15:15:09 15:17:00 111 SHARON AA0101 44766 0 AA0101 44767 58859752 0 2 PS 990901 15:25:48 15:26:00 12 PS 990901 15:34:57 15:35:03 6 00:00:00 00:00:00 0 15:35:03 15:35:14 11 AGENT 15:25:59 15:28:15 136 AGENT 15:35:13 15:35:15 2 ANAT MORIAH AA0101 44768 0 0 PS 990901 15:46:30 15:46:39 9 00:00:00 00:00:00 0 AGENT 15:46:38 15:51:51 313 ANAT AA0101 44769 78191137 2 PS 990901 15:56:03 15:56:09 6 15:56:09 15:56:28 19 AGENT 15:56:28 15:59:02 154 MORIAH AA0101 44770 0 0 PS 990901 16:14:31 16:14:46 15 00:00:00 00:00:00 0 AGENT 16:14:44 16:16:02 78 BENSION AA0101 44771 0 0 PS 990901 16:38:59 16:39:12 13 00:00:00 00:00:00 0 AGENT 16:39:11 16:43:35 264 VICKY AA0101 44772 0 0 PS 990901 16:51:40 16:51:50 10 00:00:00 00:00:00 0 AGENT 16:51:49 16:53:52 123 ANAT AA0101 44773 0 0 PS 990901 17:02:19 17:02:28 9 00:00:00 00:00:00 0 AGENT 17:02:28 17:07:42 314 VICKY AA0101 44774 32387482 1 PS 990901 17:18:18 17:18:24 6 17:18:24 17:19:01 37 AGENT 17:19:00 17:19:35 35 VICKY AA0101 44775 0 0 PS 990901 17:38:53 17:39:05 12 00:00:00 00:00:00 0 AGENT 17:39:04 17:40:43 99 TOVA AA0101 44776 0 0 PS 990901 17:52:59 17:53:09 10 00:00:00 00:00:00 0 AGENT 17:53:08 17:53:09 1 NO_SERVER AA0101 44777 37635950 2 PS 990901 18:15:47 18:15:52 5 18:15:52 18:16:57 65 AGENT 18:16:56 18:18:48 112 ANAT AA0101 44778 0 0 NE 990901 18:30:43 18:30:52 9 00:00:00 00:00:00 0 AGENT 18:30:51 18:30:54 3 MORIAH AA0101 44779 0 0 PS 990901 18:51:47 18:52:02 15 00:00:00 00:00:00 0 AGENT 18:52:02 18:55:30 208 TOVA AA0101 44780 0 0 PS 990901 19:19:04 19:19:17 13 00:00:00 00:00:00 0 AGENT 19:19:15 19:20:20 65 MEIR AA0101 44781 0 AA0101 44782 0 0 0 PS 990901 19:39:19 19:39:30 11 NW 990901 20:08:13 20:08:25 12 00:00:00 00:00:00 0 00:00:00 00:00:00 0 AGENT 19:39:29 19:41:42 133 AGENT 20:08:28 20:08:41 13 BENSION NO_SERVER AA0101 44783 0 0 PS 990901 20:23:51 20:24:05 14 00:00:00 00:00:00 0 AGENT 20:24:04 20:24:33 29 BENSION AA0101 44784 0 0 NW 990901 20:36:54 20:37:14 20 00:00:00 00:00:00 0 AGENT 20:37:13 20:38:07 54 BENSION AA0101 44785 0 0 PS 990901 20:50:07 20:50:16 9 00:00:00 00:00:00 0 AGENT 20:50:15 20:51:32 77 BENSION AA0101 44786 0 0 PS 990901 21:04:41 21:04:51 10 00:00:00 00:00:00 0 AGENT 21:04:50 21:05:59 69 TOVA AA0101 44787 0 0 PS 990901 21:25:00 21:25:13 13 00:00:00 00:00:00 0 AGENT 21:25:13 21:28:03 170 AVI AA0101 44788 0 0 PS 990901 21:50:40 21:50:54 14 00:00:00 00:00:00 0 AGENT 21:50:54 21:51:55 61 AVI AA0101 44789 9103060 2 NE 990901 22:05:40 22:05:46 6 22:05:46 22:09:52 246 AGENT 22:09:51 22:13:41 230 AVI AA0101 44790 14558621 2 PS 990901 22:24:11 22:24:17 6 22:24:17 22:26:16 119 AGENT 22:26:15 22:27:28 73 VICKY AA0101 44791 0 0 PS 990901 22:46:27 22:46:37 10 00:00:00 00:00:00 0 AGENT 22:46:36 22:47:03 27 AVI AA0101 44792 67158097 2 PS 990901 23:05:07 23:05:13 6 23:05:13 23:05:30 17 AGENT 23:05:29 23:06:49 80 VICKY AA0101 44793 15317126 2 PS 990901 23:28:52 23:28:58 6 23:28:58 23:30:08 70 AGENT 23:30:07 23:35:03 296 DARMON AA0101 44794 0 0 PS 990902 00:10:47 00:12:05 78 00:00:00 00:00:00 0 HANG 00:00:00 00:00:00 0 NO_SERVER AA0101 44795 0 0 PS 990902 07:16:52 07:17:01 9 00:00:00 00:00:00 0 AGENT 07:17:01 07:17:44 43 ANAT AA0101 44796 0 0 PS 990902 07:50:05 07:50:16 11 00:00:00 00:00:00 0 AGENT 07:50:16 07:53:03 167 STEREN 6 Averages Prevalent ACD Report: Health Insurance Time Total 8:00 8:30 9:00 9:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30 16:00 16:30 17:00 17:30 18:00 Calls 20,577 332 653 866 1,152 1,330 1,364 1,380 1,272 1,179 1,174 1,018 1,061 1,173 1,212 1,137 1,169 1,107 914 615 420 49 Answered 19,860 308 615 796 1,138 1,286 1,338 1,280 1,247 1,177 1,160 999 961 1,082 1,179 1,122 1,137 1,059 892 615 420 49 Abandoned% 3.5% 7.2% 5.8% 8.1% 1.2% 3.3% 1.9% 7.2% 2.0% 0.2% 1.2% 1.9% 9.4% 7.8% 2.7% 1.3% 2.7% 4.3% 2.4% 0.0% 0.0% 0.0% 7 ASA 30 27 58 63 28 22 33 34 44 1 10 9 67 78 23 15 17 46 22 2 0 14 AHT 307 302 293 308 303 307 296 306 298 306 302 314 306 313 304 320 311 315 307 328 328 180 Occ% 95.1% 87.1% 96.1% 97.1% 90.8% 98.4% 99.0% 98.2% 94.6% 91.6% 95.5% 95.4% 100.0% 99.5% 96.6% 96.9% 97.1% 99.2% 95.2% 83.0% 73.8% 84.2% # of agents 59.3 104.1 140.4 211.1 223.1 222.5 222.0 218.0 218.3 203.8 182.9 163.4 188.9 206.1 205.8 202.2 187.1 160.0 135.0 103.5 5.8 quantiles of waiting times to those of the exponential (the straight line at the right plot). The t is reasonable up to about 700 seconds. (The p-value for the Kolmogorov-Smirnov test for Exponentiality is however 0 { not that surprising in view of the sample size of 263,007). Beyond Averages: Waiting Times in a Call Center 9: Distribution of waiting time (1999) Large U.S. Bank Small IsraeliFigure Bank Chart1 29.1 % 20 18 Relative frequencies, % 600 Mean = 98 SD = 105 13.4 % 8.8 % 200 6.9 % 400 Exp quantiles 20 % 5.4 % 3.9 % 3.1 % 2.3 % 16 14 12 10 8 6 4 2 1.7 % 0 0 2 5 8 11 14 17 20 23 26 29 Time 0 30 60 90 120 150 180 210 240 270 300 0 200 Time 400 Waiting time given agent 600 Page 1 Medium Israeli Bank waitwait 0.9 Relative frequencies, % 0.8 exponentials: Interestingly, the means and standard deviations in Table Remark on mixtures of independent 19 are rather close, both annually0.7and across all months. This suggests also an exponential distribution for each month separately, as was0.6indeed veried, and which is apparently inconsistent with the observerd annual exponentiality. The phenomenon recurs later as well, hence an explanation is in order. We shall be 0.5 satised with demonstrating that 0.4 a true mixture W of independent random varibles Wi , all of which have coeÆcients of variation C (Wi ) = 1,0.3can also have C (W ) 1. To this end, let Wi denote the waiting time in month i, and suppose it is exponentially distributed with mean mi . Assume that the months are independent 0.2 and let pi be the fraction of calls performed in month i (out of the yearly total). If W denotes the mixture 0.1 of these exponentials (W = Wi with probability pi , that is W has a hyper-exponential distribution), then 0.0 20 40 60 2 80 100 120 140 160 1802200 220 240 260 280 300 320 340 360 380 C (W ) = 1Time + 2(Resolution C (M ); 1 sec.) where M stands for a ctitious random variable, dened to be equal mi with probability pi . One concludes that if the mi 's do not vary much relative to their mean (C (M )8 << 1), which is the case here, then C (W ) 1, Page 1 32 35 The Second Prerequisite: Models Through Examples Only. Each example starts with a Complex Reality and ends with a useful insight due to a Simple Model. ‘‘Theorem": A useful model must be simple (yet not too simple). Models in decreasing simplicity-levels: I Conceptual: Service Networks = Queueing Networks I Descriptive: Averages, Histograms I Explanatory: Comparative, Regression I Analytical/Mathematical: Little’s Law, Fluid Models, Queueing Models, Diffusion Refinements. “Corollary": To be useful, a simple model sometimes requires deep analysis. 9 Conceptual Model: Face-to-Face Services Bank Branch = Queueing Network / / Manager Entrance Xerox Teller Tourism / Bottleneck! 10 23 Descriptive Model: Transition Probabilities (Averages) Bank: A Queuing Network Transition Frequencies Between Units in The Private and Business Sections: Private Banking To Unit Bankers From Unit Bankers Private Authorized Personal Banking Compensations Authorized Compens - Business Tellers Tellers Overdrafts Authorized Personal - ations 1% 1% 4% 4% 5% 4% 6% 18% 12% 7% 4% Tellers 6% 0% 1% Tellers 0% 0% 0% 90% 0% 0% 0% 73% 6% 0% 0% 1% 64% 1% 0% 0% 0% 90% 1% 0% 2% 94% 5% 8% 64% 11% 69% 1% 0% 0% 0% 2% 0% 1% 1% 19% Authorized Personal 2% 1% 0% 1% 11% 5% Full Service 1% 0% 0% 0% 8% 1% 2% Entrance 13% 0% 3% 10% 58% 2% 0% 0%-5% 5%-10% 10% -15% >15% Dominant Paths - Business: Unit Parameter Station 1 Tourism Station 2 Teller Total Dominant Path Service Time 12.7 4.8 17.5 Waiting Time Total Time 8.2 20.9 6.9 11.7 15.1 32.6 Service Index 0.61 0.41 0.53 11 Exit Service Services Overdrafts Legend: Full Personal 88% 14% 0% Conceptual Model: Hospital (ED) Network (Marmur, Sinreich) proportion of patients process requires bed 01 Else Lab Imaging 02 Nurse reception Physician 04 triage 03 vital signs E.C.G 06 05 07 handling patient&family initial examination 09 10 08 11 Labs 100% consultation bloodwork 14 13 25,26 32,33 34 labs decision 35 18 consultation 21 29 17 16 treatment 19 treatment decision ultrasound 28 15 imaging 27 23 24 imaging 20 Xray 36 22 labs 37 38 decision 39 40 CT 30 treatment 31 observation 43 45 46 follow up every 15 minutes 48 hospitalization/ discharge 49 52 50 awaiting evacuation 51 discharge 53 instructions prior discharge 55 discharge / hospitalization else 54 56 estimated max time 60 probability of events 10% decision point for alternative processes reference point Figure 2. The Unified Patient Process Chart 12 42 treatment follow up 47 every 15 minutes awaiting discharge 41 44 else 12 consultation imaging /consultation / treatment Hours Descriptive Model: Service Times (Averages) or, Even “Doctors" Can Manage Operations Time - Morning (by Hour) vs. Afternoon (by Case): 6 AM 5 Hours 4 PM 3 2 1 0 EEG Orthopedics Surgery Blood Surgery Plastic Surgery Department Afternoon, by Case Morning, by Hour 13 Heart/Chest Surgery Neuro-Surgery Eyes E.I. Surgery Conceptual Model: The “Production of Justice" “Production” Of Justice The Labor-Court Process in Haifa, Israel Open File Allocate Prepare Activity Mile Stone Proceedings Queue Phase Phase Transition Closure Avg. sojourn time ≈ in months / years Processing time ≈ in mins / hours / days Appeal 14 Analytical Model: Little’s Law in Court (still Averages) Judges: Judges: Operational Performance Performance Analysis – Base Case Judges: Performance byAnalysis Case-Type Judges: Performance Operational Performance: Rate λ & Time W Case Type 0 Case Type 01 Case Type 3 10 Average Number of Months - W 9 Judge1 Judge2 Judge3 Judge4 Judge5 0 0 01 8 7 . 45 (6.2, 7.4) . 100 (13.5, 7.4) 3 3 01 3 6 5 (7.2, 4.6) 01 4 . 3 33 0 59 (12, 4.9) . 0 0 118 (26.3, 4.5) 3 . 01 01 3 2 1 0 0 5 10 15 20 Average Number of Cases / Month - λ 15 25 30 Expertise “invite" Skills-Base-Routing (SBR) Operational Performance: Judges’ Heterogeneity Judges: Judges: Operational Performance Performance Analysis – Base Case(Best / Worse) Judges: Performance byAnalysis Case-Type Judges: Performance Case Type 0 Case Type 01 Case Type 3 10 Average Number of Months - W 9 Judge1 Judge2 Judge3 Judge4 Judge5 0 0 01 8 7 . 45 (6.2, 7.4) . 100 (13.5, 7.4) 3 3 01 3 6 5 (7.2, 4.6) 01 4 . 3 33 0 59 (12, 4.9) . 0 0 118 (26.3, 4.5) 3 . 01 01 3 2 1 0 0 5 10 15 20 Average Number of Cases / Month - λ 16 25 30 Analytical Model: Little’s Law in Court (still Averages) Judges: Performance Analysis Judges’ Profiles (Operational) Case Type 0 Case Type 01 Case Type 3 10 9 0 0 8 Avg. Months - W Judge1 Judge2 Judge3 Judge4 Judge5 01 7 . (6.2, 7.4) . (13.5, 7.4) 3 3 01 3 6 5 (7.2, 4.6) . 3 01 4 . (12, 4.9) 0 0 3 (26.3, 4.5) . 0 01 01 3 2 1 0 0 5 10 15 Avg. Cases / Month - λ 17 20 25 30 Analytical Model: Little’s Law in Court (still Averages) Judges: Performance Analysis Judges: The Best/Worst (Operational) Performer Case Type 0 Case Type 01 Case Type 3 10 9 0 0 8 Avg. Months - W Judge1 Judge2 Judge3 Judge4 Judge5 01 7 . (6.2, 45 7.4) 3 3 . (13.5, 100 7.4) 01 3 6 5 (7.2, 4.6) 01 4 . 3 33 0 59 (12, 4.9) . 0 0 118 (26.3, 4.5) 3 . 01 01 3 2 1 0 0 5 10 15 Avg. Cases / Month - λ 18 20 25 30 Call-Center Network: Gallery of Models Index Service Engineering: Multi-Disciplinary Process View (75% in Banks) Information Design Marketing, Operations Research Lost Calls ( Waiting Time Return Time) Organization Design: Parallel (Flat) Sequential (Hierarchical) Sociology/Psychology, Operations Research Agents Queue Redial Function Scientific Discipline Multi-Disciplinary Call Center Design Service Completion Experts (Consultants) (Invisible) (Retrial) Computer-Telephony Integration - CTI MIS/CS Busy (Rare) Good or Arrivals (Business Frontier Bad Job Enrichment Training, Incentives Human Resource Management of the 21th Century) VRU/ IVR Forecasting Statistics Internet Chat Email Fax Customers Interface Design Human Factors Engineering New Services Design (R&D) Operations, Marketing Agents (CSRs) To Avoid Starvation Skill Based Routing (SBR) Design Marketing, Human Resources, To Avoid Operations Research, MIS Delay Customers Segmentation CRM Marketing Tele-Stress Psychology (Turnover up to 200% per Year) (Sweat Shops of the 21th Century) Psychological Process Archive Expect 3 min Willing 8 min Perceive 15 min Back-Office VIP VIP Queue (Training) Service Process Design Abandonment Psychology, Logistics Statistics Lost Calls Positive: Repeat Business Negative: New Complaint 19 Operations/ Business Process Archive Database Design Data Mining: MIS, Statistics, Operations Research, Marketing Service Completion (If Required 15 min, then Waited 8 min) (If Required 6 min, then Waited 8 min) Psychology, Operations Research, Marketing The “Phases of Waiting" for Service Common Experience: I Expected to wait 5 minutes, Required to 10 I Felt like 20, Actually waited 10 (hence Willing ≥ 10) An attempt at “Modeling the Experience": 1. Time that a customer expects to wait 2. willing to wait 3. required to wait 4. actually waits 5. perceives waiting. Experienced customers “Rational" customers ⇒ ⇒ ((Im)Patience: τ ) (Offered Wait:V ) (Wq = min(τ, V )) Expected = Required Perceived = Actual. Then left with (τ, V ). 20 Call Center Data: Hazard Rates (Un-Censored) Required/Offered Wait V (Im)Patience Time τ −3 5 x 10 4.5 4 Israel hazard rate 3.5 3 2.5 2 1.5 1 0.5 0 0 50 100 150 200 time, sec actuarial estimate spline smoother 16 0.3 14 0.25 12 hazard rate U.S. hazard rate 0.35 0.2 0.15 10 8 6 0.1 4 0.05 0 0 2 10 20 30 time, sec 40 50 0 0 60 21 10 20 30 time, sec 40 50 60 A Patience Index Quantifying (Im)Patience: “Willing to wait 15 min" = Patient? Willing to wait E[τ ] = , Expected to wait E[V ] “assuming" Experienced; further “assuming" that τ and V are Exponential, the M-L estimate of Index is the easily-measurable: ∆ Theoretical Patience-Index = ∆ Empirical Patience-Index = % Served % Abandoning Index Validation: Theoretical vs. Empirical 10 9 Theoretical Index 8 7 6 5 4 3 2 1 0 2 3 4 5 6 Empirical Index 7 8 9 Predicting Performance Model Primitives: I I I I Arrivals to service (random process) (Im)Patience while waiting τ (r.v.) Service times (r.v.) # Servers / Agents (parameter / r.v.) Model Output: Offered-Wait V (r.v.) Operational Performance Measure calculable in terms of (τ, V ). I I eg. eg. Average Wait = E[min{τ, V }] % Abandonment = P{τ < V } Application: Staffing – How Many Agents? (When? Who?) 23 The Basic Staffing Model: Erlang-A (M/M/n +M) agents 1 arrivals queue 2 λ … n µ abandonment θ Erlang-A Parameters: I I I I λ – Arrival rate (Poisson) µ – Service rate (Exponential) θ – Impatience rate (Exponential) n – Number of Service-Agents. 24 Erlang-A: Fitting a Simple Model to a Complex Reality Hourly Performance vs. Erlang-A Predictions (1 year) % Abandon E[Wait] %{Wait > 0} 1 250 0.9 0.5 0.8 0.4 0.3 0.2 Probability of wait (data) Waiting time (data), sec Probability to abandon (data) 200 150 100 0.7 0.6 0.5 0.4 0.3 0.2 50 0.1 0.1 0 0 0.1 0.2 0.3 0.4 0.5 Probability to abandon (Erlang−A) I I 0.6 0 0 50 100 150 200 Waiting time (Erlang−A), sec 250 0 0 0.2 0.4 0.6 Empirically-Based & Theoretically-Supported Estimation of (Im)Patience: θ̂ = P{Ab}/E[Wq ]) Small Israeli Bank (more examples in progress) 25 0.8 Probability of wait (Erlang−A) 1 Testing the Erlang-A Primitives I I I Arrivals: Poisson? Service-durations: Exponential? (Im)Patience: Exponential? Validation: Support? Refute? 26 May 1959! Arrivals to Service: only Poisson-Relatives Time 24 hrs Arrival Rate to Three Call Centers (Lee A.M., Applied Q-Th) Dec. 1995 (U.S. 700 Helpdesks) May 1959 (England) Q-Science Arrival Process, in 1999 Arrival Rate % Arrivals Yearly Monthly Dec 1995! May 1959! Time 24 hrs Time 24 hrs (Help Desk Institute) November 1999 (Israel) 28 (Lee A.M., Applied Q-Th) Daily Hourly % Arrivals Dec 1995! Observation: Peak Loads at 10:00 & 15:00 27 Time 24 hrs Service Durations: LogNormal Prevalent Israeli Bank Log-Histogram Survival-Functions Service Time by Service-Class Survival curve, by Types 900 800 700 NW (New) = 11 PS (Regular) = 1 600 500 Survival Frequency Means (In Sec Average = 2.24 St.dev. = 0.42 400 300 NE (Stocks) = 2 IN (Internet) = 200 100 0 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 Log(service time) frequency Time normal curve 3 I I New Customers: 2 min (NW); Regulars: 3 min (PS); I I Stock: 4.5 min (NE); Tech-Support: 6.5 min (IN). Observation: VIP require longer service times. 28 (Im)Patience while Waiting (Palm 1943-53) Irritation ∝ Hazard Rate of (Im)Patience Distribution Regular over VIP Customers – Israeli Bank 16 I 14 I Peaks of abandonment at times of Announcements Call-by-Call Data (DataMOCCA) required (& Un-Censoring). Observation: VIP are more patient (Needy) 29 A “Service-Time" Puzzle at a Small Israeli Bank ' $ Inter-related Primitives Figure 12: Mean Service Time (Regular) vs. Time-of-day (95% CI) (n = 42613) 180 160 100 120 140 Mean Service Time 200 220 240 Average Service Time over the Day – Israeli Bank 7 & 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 % Time of Day 11 30 Prevalent: Longest services at peak-loads (10:00, 15:00). Why? Explanations: I Common: Service protocol different (longer) during peak times. I Operational: The needy abandon less during peak times; hence the VIP remain on line, with their long service times. 30 Erlang-A: Simple, but Not Too Simple Experience: I Arrival process not pure Poisson (time-varying, σ 2 too large) I Service times not exponential (typically close to lognormal) I Patience times not exponential (various patterns observed). I Customers and Servers not homogeneous (classes, skills) Questions naturally arise: 1. Why does Erlang-A practically work? justify robustness. 2. When does it fail? chart boundaries. 3. Generalize: time-variation, SBR, networks, uncertainty , . . . Answers via Asymptotic Analysis, as load- and staffing-levels ↑ , which captures model-essentials: I I Efficiency-Driven (ED) regime: Fluid models (deterministic). Quality- and Efficiency-Driven (QED) regime: Diffusion refinements (eg. revealing that the patience-density at the origin is “all" that is needed). 31 DataMOCCA = Data MOdels for Call Center Analysis I I I Technion: P. Feigin, V. Trofimov, Statistics / SEE Laboratory. Wharton: L. Brown, N. Gans, H. Shen (UNC). industry: I I U.S. Bank: 2.5 years, 220M calls, 40M by 1000 agents. Israeli Cellular: 2.5 years, 110M calls, 25M calls by 750 agents; ongoing. Project Goal: Designing and Implementing a (universal) data-base/data-repository and interface for storing, retrieving, analyzing and displaying Call-by-Call-based Data / Information. System Components: I Clean Databases: operational-data of individual calls / agents. I Graphical Online Interface: easily generates graphs and tables, at varying resolutions (seconds, minutes, hours, days, months). Free for academic adoption: Mini version available on a DVD; working version 7GB tables, or 20GB raw zipped, for each call center – ask for my mini-HD. 32 Arrivals to Service: Predictable vs. Random Arrival Rates on Tuesdays in a September – U.S. Bank 3500 3000 Arrival rate 2500 2000 1500 1000 500 0 0:00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 Time (Resolution 30 min.) 04.09.2001 I I I 11.09.2001 18.09.2001 25.09.2001 Tuesday, September 4th: Heavy, following Labor Day. Tuesdays, September 18 & 25: Normal. Tuesday, September 11th, 2001. 33 A “Waiting-Times" Puzzle at a Medium Israeli Bank waitwait 0.9 Relative frequencies, % 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 Time (Resolution 1 sec.) Page 1 Peaks Every 60 Seconds. Why? I Human: Voice-announcement every 60 seconds. I System: Priority-upgrade (unrevealed) every 60 seconds. Served Customers Abandoning Customers waithandled waitab 0.6 0.9 0.5 Relative frequencies, % Relative frequencies, % 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.4 0.3 0.2 0.1 0.1 0.0 0.0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 20 Time (Resolution 1 sec.) 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Time (Resolution 1 sec.) Page 1 Page 1 34 Priorities, Economies-of-Scale, SBR Regular vs. VIP Customers: Cellular – March 23, 2004 Average Wait Staffing Level 50 60 50 40 Number of agents Average waiting time, sec 45 35 30 25 20 15 10 40 30 20 10 5 0 7:00 9:00 11:00 13:00 15:00 17:00 19:00 21:00 0 7:00 23:00 9:00 11:00 13:00 Time Private 15:00 17:00 19:00 21:00 Time Private Platinum Private Private Platinum I Design: VIP-dedicated agents, Regular-dedicated Agents. VIP’s are not served better than Regular’s I Solutions: Add VIP agents (costly), or Re-Design. I 35 23:00 Priorities and Routing Protocols I Regular vs. VIP Customers: Cellular – October 2004 Average Wait 40 90 35 80 Average Wait, sec Delay Probability, % Delay Probability 100 70 60 50 40 30 20 25 20 15 10 5 10 0 7:00 30 9:00 0 7:00 11:00 13:00 15:00 17:00 19:00 21:00 23:00 Private 9:00 11:00 13:00 15:00 17:00 19:00 21:00 23:00 Time (Resolution 30 min.) Time (Resolution 30 min.) Private Platinum Private Private Platinum More VIPs delayed than Regulars, yet their average wait is shorter. What changed since last March? 36 Priorities and Routing Protocols II Waiting-Time Histograms: Cellular – October 2004 Regular Customers VIP (Platinum) Customers private_hist platinum_hist 3.0 18.0 Relative frequencies, % Relative frequencies, % 16.0 2.5 2.0 1.5 1.0 0.5 14.0 12.0 10.0 8.0 6.0 4.0 2.0 0.0 0.0 2 14 26 38 50 62 74 86 98 2 Time (Resolution 1 sec.) 14 26 38 50 62 74 Time (Resolution 1 sec.) Page 1 Page 1 After 25 seconds of wait, VIP customers are routed with high priority to Regular agents. Hence, almost no long waiting times for VIP’s. 37 86 Distributed via Call Center (U.S. Bank) at a U.S. Bank Network Balancing Inter-Queues 10 AM – 11 AM (03/19/01): Interflow Chart Among the 4 Call C t f Fl t B k Internal arrivals: 224 • External arrivals:2092 2063(98.6%Served)+29(1. 4%Aban) • • Not Interqueued:1209(57.8%) • Served: 1184(97.9/56.6) • Aban: 25(2.1/1.2) Interqueued :883(42.2) • Served here:174(19.7/8.3 ) • Served at 2: 438(49.6/20.9) S d t3 • Internal arrivals: 643 • Served at 1: 67 (29.9) Served at 2: 41 (18.3) Served at 3: 87 (38.8) Served at 4: Served at 1: 157 (24.4) Served at 2: 195 (30.3) Served at 3: 282 (43.9) Served at 4: 4 (0.6) Aban at 1: 3 • • • • 2 NY 1 179 +5 619 +3 19 +1 • Not Interqueued: 1665(98.3) Served: 1659 (99.6/97.9) Aban: 6 (0.4/04) • Interqueued:28+1 (1.7) • Served here: 17(58.6/1) • Served at 1: 3(10.3/0.2) • • • • Served here: 110 (41.2/6.2) Served at 1:58 (21 7/3 3) External arrivals: 122 112(91.8 Served)+10(8.2 Aban) M A4 Internal arrivals: 613 • Served: 1497 (99.6/84.6) Aban: 6 (0.4/0.3) Interqueued:258+9 (15.1) • 11 +1 101+ 2 PA 2 • Not Interqueued: 1503(84.9) 74 +7 508 +2 External arrivals: 1694 1687(99.6% Served)+7( 0.4% Aban) • RI 3 8+ 1 20 External arrivals: 1770 1755(99.2 Served)+15(0.8 Aban) Internal arrivals: 81 Served at 1: 41(6.7) Served at 2: 513(83.7) Served at 3: 55(9.0) Aban at 1: 2(0.3) • • • 38 Served at 1: 17(21) Served at 3: 42(51.9) Served at 4: 15(18 5) Not Interqueued: 93 (76.2) • • • • Served: 85 (91.4/69.7) Aban: 8 (8.6/6.6) Interqueued:27+2 (23.8) Served here: 14(48.3/11.5) Served at 1: 6 Balancing Protocols and Performance Level U.S. Bank: Histograms of Waiting Times Retail Business Chart1 Sheet3 Chart 1 12 20 10 16 Relative frequencies, % Relative frequencies, % 18 14 12 10 8 6 4 8 6 4 2 2 0 0 2 5 8 11 14 17 20 23 26 29 32 35 2 Time 8 14 20 26 32 38 44 50 56 62 68 74 80 86 Time Page 1 Page 1 Peak for Retail service at 10 seconds – Why? After 10 seconds of wait, Retail customers sent into the inter-queue. Business customers – peak at 5 seconds, for the same reason. Second peak – unclear, maybe priority-upgrade. 39 92 Data-Based Service-Research (with DataMOCCA, even before tenure) I I I Contrast with EmpOM: Industry / Company / Survey Data (Social Sciences) Converge to: Measure, Model, Validate, Experiment, Refine (Physics, Biology, . . .). Prerequisites: I I I OR, OM, IE, (Mktg.) - for Design CS, IS, Stat. - for Implementation. Outcomes: Relevance, Credibility; Interest, Fun; Call Centers as a Pilot (eg. for Healthcare). Moreover, I I I Teaching: Class, Homework (Experimental Data Analysis); Cases. Research: Validate Existing (Queueing) Theory/Laws and Suggest New Models/Research. Practice: OM Tools (Scenario Analysis), Mktg. (Trends, Benchmarking). 40 Live Demonstration of DataMOCCA 5-7 minutes, to emphasize “online" capabilities. U.S. Bank I I I Daily Reports: October 2003, weekdays; typically takes 10-20 sec till a first output, but this is because of PowerPoint/Windows. Then do few additional Daily Reports, say Monday, Tuesday,... (starting with STATCCA, as opposed to by minimizing the powerpoint screnn) - this will be now happening very fast. Time-Series: Number of agents, for ALL classes, all months, weekdays. (Including total). Shows scale, trends. Then do Service Durations, indicating that 1 second of 1000 agents could cost $500M per year. Could also do Unhandled (lower middle entry in list), for only Retail and Premium - Premium is worse, and deteriorating, Daily Summaries: I I I Tuesdays in September 2001; September 11th; shown during lecture under the heading “Predictable or Random"; 30 sec scale - stoch. variability, 1 hour scale - the “right" scale; % to mean, to show very similar shape over 3 Tuesdays. Suggests the model λ(t) = λ0 (t) · Z , for 0 ≤ t ≤ T . 41