Revenue Management Through Dynamic Cross-Selling in Call Centers

1. Revenue Management Through Dynamic Cross-Selling in Call Centers O. Zeynep Akşin Koç University Joint work with E. Lerzan Örmeci, Koç University

2. Call Centers in Banking • The Tower Group estimates that nearly 39 billion retail banking transactions were conducted in the US during 1999, growing to 44 billion in 2003. Call centers processed 18% of transactions in 1999 and are projected to represent almost 25% of transactions by 2003

3. Shifting from Cost to Profit CentersSource: Merchants International 2000

4. Growth in the Financial Services Industry • Mature markets: Increase “share-of-wallet” • Existing customers are better sales prospects (Felvey 1982) • Dislike for telemarketing • Sales increasingly in the form of cross-selling (Kresbach, 2002; Walker, 2003) • Cross-selling as an important CRM initiative • Call centers as a key customer contact point

5. “This new loan option is exactly what I need!” + $$$$ “I don’t want another sales pitch, just transfer the money! ” Lost time, increased waiting! Cross-selling basics Data-mining Training & incentives Customer reaction Try to Cross-sell? • Paas and Kuijlen, 2001 • Kamakura et al. 1991, 2003 • Insurance Advocate, 2003 • American Banker, 2003 • Retention: Marple and Zimmerman, 1999 • Churn: Kamakura et al. 2003 • Switch: Kamakura et al. 2003

6. The Basic Trade-off Service Quality Revenue enhancement • Current load in system • Talk times • Call volumes • Who is likely to buy? • Real-time automation • Historical data Key questions: When to cross-sell? To whom to cross-sell?

7. Cross-selling and Operations • Akşin and Harker (1999): first model that considers cross-selling impact on service delivery performance and design • Güneş and Akşin (2004):Incentive design for cross-selling; illustrates link to market segmentation • Byers and So (2003): performance of cross-selling policies that consider queue state information and/or customer profile information. • Netessine, Savin, and Xiao (2004): cross-selling in e-retailing.

8. Dynamic admission control • Admission control with random revenues (revenue realization) • Ghoneim and Stidham (1985): Optimal threshold policies for a system in series • Ormeci, Burnetas, Emmons (2002): Optimal threshold policies for a loss system, existence of preferred jobs • Admission control with fixed revenues (expected revenues) • Koole (1998), and Altman, Jimenez, Koole (2001): Optimal threshold policies for a loss system • Ormeci, Burnetas, van der Wal (2001), and Savin, Cohen, Gans, Katalan (2003): Existence of preferred classes for a loss system

9. Two models of cross-selling • The model with revenue realization • Real-time automation • Server knows revenue potential of a customer at the beginning of a call • The model with fixed revenues • Analysis of historical segment based data • Server makes decision based on expected revenue potential at the beginning of a call • At the end of the call, a random revenue is realized

10. Market segmentation • Aggregation of customers into homogeneous groups according to their cross-sell revenue generation potentials. • Discrete segmentation: Most widely used • In cross-selling, this typically leads to cross-sell to only class-H calls type policy • Overlapping segmentation: • a more realistic and theoretically accurate segmentation scheme. Lilien and Rangaswamy (1998) • In cross-selling, this typically leads to cross-sell to all calls type policy L H L H

11. Model with random revenues: A 2-class c-server no-wait-room system S1 1 lH 2 S2 Poisson arrivals lL 1 : xsell rate m1 =m-k 2: service rate m Sc • service revenue: r • xsell revenue: • rs~Fs, s=L,H exponential service times

12. Market segmentation based on cross-sell revenues Upper bound on random revenues: Lower bound on random revenues: Segments in marketing discrete Possible scenarios: overlapping irrelevant

13. Markov decision model of the system • State: x=(x1, x2) & (x, rs, s)= (x1, x2, rs, s) • x1: number of cross-sell customers • x2: number of service customers • rs : random revenue observed upon the arrival of the call • s: class/segment of the incoming call • Actions (only upon arrivals): • Objective: maximize total expected b-discounted revenue over an ∞-horizon • Maximal total expected b-discounted revenue: u(x1, x2), v(x1, x2, rs, s)

14. Illustration of the Model u(x1, x2+1)+r Service only Class-s arrival v(x1, x2, rs, s) u(x1, x2) Cross-sell Service completion u(x1+1, x2)+r+rs Cross-sell completion u(x1, x2-1) u(x1-1, x2)

15. Definition • D(21)(x): Expected loss in future rewards because of the increased load due to the slower service of a cross-sell compared to a pure service call in state x a threshold on revenues in state x D(21) = max { D(21)(x) : 0 ≤ x1+ x2 ≤ c }

16. Preferred calls & classes • Preferred calls are those that always generate a cross-sell attempt. If D(21) < rs, call (rs, s) is preferred. • Class s is preferred if all class-s calls always generate a cross-sell attempt. If D(21) < rs, class s is preferred. • Use of upper bounds on D(21) to find sufficient conditions to have preferred calls/classes.

17. Cross-sell dynamically (Policy 0) not interesting for us

18. Cross-sell attempt to chosen calls of segment H only (Policy 1)

19. Cross-sell attempt to segment H only (Policy 2) not possible for Scenario 2

20. Cross-sell attempt to segment H and chosen calls of segment L (Policy 3)

21. Cross-sell attempt to chosen calls of segment H and L (Policy 4) not possible for Scenario 1

22. Cross-sell attempt to everyone (Policy 5)

23. Proposition for Scenario 1:

24. Proposition for Scenario 2:

25. Back to cross-selling in call centers • Attempt + forward (f): Inbound service representatives attempts a cross-sell and forwards to sales personnel to close and book • Attempt + close (c): Inbound service representative has the required capabilities to close and book the sale • Different talk-time implications

26. Set-up for numerical experiments • Case 1 (constructed from real data of a retail banking call center) • Regular service time = 2.7 minutes • Regular service revenue = 1 unit • 27% increase in talk time for forward xsell • 220% increase in talk time for close xsell • Case 2 (constructed from Case 1 to reflect the environment of insurance/investment) • Regular service time = 5.5 minutes • Regular service revenue = 20 units • 27% increase in talk time for forward xsell • 220% increase in talk time for close xsell

27. Experimental design • Load (l/cm): 0.75, 0.90, 1.05, 1.20 • Number of servers (c): 100, 150, 200 • : 75, 125, 175 • : • : • : 0 • Total number of experiments : 1296

28. Performance of expected vs random realization

29. Size load Size load 100 100 150 150 200 200 120 120 105 105 90 90 75 75 When are policies more dynamic? Attempt and forward Attempt and close

30. When are policies more dynamic? • When • There is a premium H segment that is difficult to distinguish from remaining calls (narrower H segment revenues, wider L segment revenues or more overlap) • Talk times and cross-sell durations are long

31. Performance of our sufficient conditions in numerical experiments • The structure of the optimal policy is predicted correctly in 74% of the experiments. • Use conditions within a heuristic  Good static policies

32. Performance of static policies: H-only • The performance of this policy is quite bad in some of our examples

33. Performance of static policies: All

34. D21 bound as a threshold vs. best of static

35. Conclusions • Current marketing practice: Optimal dynamic policies do not overlap with revenue based segment policies • However a smart choice of static policies can lead to good performance • Theoretical results characterizing • Preferred class/call structure • Monotonicity of optimal policy: concavity? • The use of these structural results to generate a static threshold based heuristic leads to consistently good performance: further refinement?

36. Conclusions contd. • Resorting to dynamic optimization rather than simple static policies for cross-selling in call centers will be more beneficial for call centers with • narrow range for revenues of segment H, and wider or more overlapping revenues of segment L • long additional talk time for x-sell • Large call centers operating in a quality efficiency or quality regime • The value of real-time-automation can be high, so it is important to understand environments where this is the case