Analytics in the call center

One recent Monday morning, I received an unexpected “cold call” at my office. The caller identified himself as a consultant with an IEOR background from a reputable graduate program, and he wanted to talk to me about call centers. Having recently done some operational improvement projects in a few call centers, he was very surprised to discover powerful call forecasting and agent scheduling software packages (known in the industry as “workforce management” systems, or WFM for short) whose capabilities were highly underutilized.

He had been even more stunned to find that the analysts responsible for operational planning and analysis were often former call center agents with little or no background with math, statistics or optimization, treating these powerful software tools as very delicate and somewhat scary black boxes. Multi-site operations, multiple call types, multi-skilled agents, forecast errors and real-time schedule changes all added operational complexity to the planning and analysis process.

All of this sounded quite familiar to me [1].

Like many other enterprise software systems (CRM and BI come to mind immediately), there has been a sustained wave of excitement about the promise of call center WFM. As far back as 1997, I wrote, “As computers grow faster and cheaper, more and more information is captured by call center databases. It is my strong belief that the potential buried in these databases will continue to increase, and that we have just begun to help the call center industry understand how to tap into this potential” [2].

Back then, only about 10 percent of call centers had purchased WFM solutions, while a 2010 survey put the penetration rate at more than 75 percent, with over 84 percent of respondents stating that such systems were “mission critical” [3]. But this same study reported that customer satisfaction with these systems was “strikingly low,” just as it was back in my day. While the usual blame for this customer dissatisfaction is placed on the technology and its vendors, my sense has always been that the primary source of the unhappiness is rooted in the selection, training and empowering of the people tasked with using the software to magically capture some promised ROI. And the frustration that I heard in that consultant’s voice over the phone was a rueful reminder of this.

A couple of days after receiving the cold call, I got a call from my old friend Rick Casares. Rick works for Verint, which had long ago acquired the company that had acquired Blue Pumpkin Software, a WFM software vendor and a former client and employer of mine. “I’m coming out there to learn about voice biometrics and fraud detection. You know we acquired a company out in Menlo Park [4] a while back, right? It’s a really interesting ‘big data’ application, and some of the big players in the credit card industry have been early adopters of this stuff…”

We spent the next 45 minutes talking about how these kinds of 21st century digital detectives do their work. Though unable to share any proprietary information, Rick still managed to enlighten me quite a bit. Everything in the world of credit cards is about speed and volume, he pointed out, so the key concepts here are “real-time analysis” and “risk economics.”

First off, when a call comes into a call center that has utilized Verint’s technology, the caller’s voice is captured and tested (against his/her own voice if a validated sample is available or against voice records of known fraudsters) in a matter of just a few seconds to determine whether the caller is actually who he/she claims to be.

Meanwhile, even as this authentication is taking place, all the relevant data about this call – including the date and time of the call, the card number about which the customer is inquiring, and information about the phone number from which the call was placed – is captured, also in real time. This data is fed back to a sophisticated predictive model that utilizes this information along with a large number of historical data elements (such as the number of recent calls, the content of those previous calls and any recent changes to the account) to estimate the probability that the call is fraudulent.

From here, a team of analysts (employed by Verint and located in a secure area to provide specialized professional services for the credit card issuer) pores over all records that are estimated as having a probability of fraud above some threshold, looking at each individual record before sending alerts directly over to the bank. From a statistical perspective, these analysts are key players in managing the trade-off between Type I errors (that is, a false positive determination of fraud) and Type II errors (letting fraudulent transactions pass by undetected).

The final determination of what course of action to take is ultimately made by the bank that issued the card, which has its own team of specialized analysts. This extra layer of scrutiny is primarily because of the economic impact of slowing down the lending and spending machine. Specifically, a credit card issuer is loath to put a freeze on a card unnecessarily because this puts its coveted “front of wallet” position with that customer at risk (Americans have nearly five cards each on average), because losing it means that the customer begins using another card more frequently, with that competitor then capturing the (steady) margin on the transactions and the (lucrative) interest on the outstanding credit balances.

While Rick couldn’t give me any specific numbers about the true probability of fraud, he did allude to the fact that a huge proportion of fraudulent activity was conducted by an extremely small part of the population. This suggests that the challenge of intelligently managing false positives is still a big issue. But the Verint models continue to get better, in part because of faster computers and smarter algorithms and in part because of the constant model tuning based on updated data.

Though early in its product lifecycle, these voice biometric and fraud prediction technologies are already having a serious impact. According to a recent Associated Press survey of software vendors, more than 65 million voiceprints have been captured and stored already, and Verint V.P. Mark Lazar has been quoted as claiming that these kinds of tools have reduced fraudster phone calls by 90 percent in some cases [5].

Given the explosive growth in credit card fraud over the past decade – a recent article [6] in The Economist reports that losses exceeded $11 billion in 2012 from less than $4 billion in 2003 – one can safely anticipate that the demand for and sophistication of such solutions will continue to grow. Though if WFM is any guide, the long-term success of such applications will depend not only on the intelligent algorithms embedded within them but also on having enough technically skilled and business savvy analysts in place to make the value chain work. We will see.

REFERENCES

1. Full disclosure: much of my research over the past several years has been in these areas, as a quick look at my Google Scholar page (https://scholar.google.com/citations?user=MmO4WSwAAAAJ&hl=en) quickly reveals.
2. Mehrotra, Vijay, 1997, “Ringing Up Big Business, OR/MS Today, August issue; see:http://www.lionhrtpub.com/orms/orms-8-97/CallCenter.html.
3. These numbers are from the 2010 edition of the annual DMG Consulting study as cited inhttp://www.customerzone360.com/workforce-management/articles/86777-survey-reveals-that-84-percent-call-center-managers.htm.
4. For more on this acquisition, see http://www.darkreading.com/attacks-breaches/verint-acquires-voice-biometric-company-victrio/d/d-id/1140788?
5. These numbers are from an October 2014 article by Raphael Sutter of the Associated Press. Seehttp://m.startribune.com/business/279079751.html or http://www.businessinsider.com/banks-use-voiceprint-on-calls-to-detect-fraud-2014-10 for various versions.
6. See http://www.economist.com/news/finance-and-economics/21596547-why-america-has-such-high-rate-payment-card-fraud-skimming-top.