Using data science for Contact Center optimization

Transcript

Note: This transcript was created using speech recognition software. It may contain errors.

Siddharth Gard:

Hi, my name is Siddharth and I welcome you on this presentation about data center optimization using machine learning. And so something about me. I’m currently working as a senior analyst in TransUnion. I’m responsible for building algorithms that works behind the scene to provide most optimized products. Previously, I have worked with Networks and CGI Group on market leading products. So what will be the agenda for you today? so we’ll, we’ll have some in production about the contact center. we’ll see the numbers how data science and machine learning can be used to improve contact center experience. then we’ll have some description about cloud call analytics solution types ML modeling strategy the strategy use cases and then we’ll use time series forecasting for inbound call centers. so the general flow of the talk will be we have inbound and outbound call centers which use cloud call analytics to perform the same function.

 we have in-house data strategy that is more relevant to the outbound call center. we’ll see the strategy use case because of time limitation. We won’t be going in details about time series forecasting. It’s it’s own topic and it’s time consuming effort. So we’ll be most significantly talking about outbound call centers and in-house data product strategies. So, before going to that, like I know everybody seems like, okay, I have no contact center. What is contact center? And this is what you imagine as soon as somebody says, okay, this is contact center. But, you know, generally contact center supports customer interaction over wide range of channels, including phone call, email, web chats, and, you know, still in the era of various automations. people prefer automat prefer, prefer real people over automated systems. So, you know, whenever your system goes down or you have any issue, you want to straight away talk to somebody or get hold of somebody instead of talking to a automated system because, you know the people give you the solutions most of the time. so, so it matters a lot. like if you have a really good contact center strategy you have a very good marketing team. And there will be a very good impression about the organization

Itself. will be limiting the conversation to the finance industry mortgage loan servicing, which is about 26% globally. So it’s the biggest chunk. So we’ll be discussing that. So some of the issues faced by contact centers bad consumer experience, highway times, then you have bad agent management, lot of dead here and call hanging agent compliance issues, agent attrition. So there, there are lots and lots of issues. you all the whenever you call to a call center, you have long wait times. They’ll keep on saying, okay, be online and we’ll get hold of somebody. and then there’s one limitation. That company always said that we need to adhere to the target tech service levels, and we need to keep the staffing calls down. So it’s a huge balancing act where you have cost limitations, but at the same time, you have to manage your service level agreements really well.

So contact centers in numbers before jumping to any of the remedies we have for all the ailments we discussed previously. so we see, like the contact center is a market is expected to reach by nearly 500 billion by 2027. US companies lose over 62 billion in and with revenue, revenue due to poor customer service. So it’s a huge chunk, even if you are able to use some of the techniques, you can leverage nearly 10% that comes around 6 billion, and, you know, it’s a huge chunk of money. So it really matters. 88% of customers prefer voice calls with live agent rather than other means. 87% of customers find it frustrating to repeat themselves in multiple channels, and 73% question doing business with that brand as a result. So if your questions are not being answered properly, you are having issues with the product next time, you will be having some issues dealing with that brand again. So it really comes down to customer engagement in a very necessary way. customers are likely to spend one 40% more after positive experience rather than a negative one. So you see it’s a huge impact. You can see a very huge value boost if you have a good customer contact center. And 77% of customers view our business more positively if they are proactive with customer service. So we see it, it has a huge impact.

So ML machine learning can help to find many pain points in the process based on previous data itself. So you can avoid most of the problems with customers if you have a very good machine learning strategy, because it’ll help you to avoid most, most of the calls in the first place by implementing some of the very good strategies to save costs, inbound and outbound call centers we can use various for forecasting techniques and machine learning techniques. okay, you know, this is very time consuming and capital consuming. So any contact center in-house data strategies, time consuming, and there’s lot of effort that is needed that can also be a risky one because, you know, sometimes it gives good results. Sometimes it if it’s not that the business is not taken into account, it can result into cost overrun combination of off-the-shelf product and in-house data products help in optimizing operations by providing insights on the call data and strategic data respectively.

So you have two things. You have in-house products and you have off-the-shelf products. So off-the-shelf products are basically cloud products or analytic products, which are available by various vendors. in-house data products are, which you develop for the strategic versions because you have to implement a strategy for any loan servicing. So like, so some of the benefits of your call cloud call analytics solutions which is you know, of the shelf data products that is you get best better customer insights improve execution decision making by extracting detailed conversion and insights. boot boost agent productivity by analyzing call sentiment, script compliance, and speech characteristics. operational efficiency analyze up to a hundred percent calls to identify product improvement and data security. you know, you need to remove some of the call transcription details.

You don’t need to provide ssn personal details. So all these can be done with the help of a cloud call analytics solution that is off the shelf. And these are like some of the various leading, so solution providers like, you know, a w s, Google Cloud, Microsoft Digital contact center. So we have discussed something about off the shelf data products, but our focus for the today is in-house data products. So in-house data products will help you with the strategy for the contact center and you know, of the, you know, cloud contact centers provide really good call details and call volumes about inbound and outbound call. But we need a strategy which is done by, in in-house data products in-house data products. today for the time limitations, there are really huge number of use cases which can be done.

But this is strategy, which we will be discussing today, can be applied for the two, two use cases that is phase calling and right time to call. We’ll discuss that in upcoming slides. But yeah, these are the two cases, but there the strategy we are going to discuss can be applied to really huge number of use cases. So what is the risk? what is the methodology for this strategy? we’ll be using machine learning to examine the loan default. we’ll be using 70% of the data for rolling five years to fit the model. for validation, we’ll be using 30% data. we’ll be evaluating model performance, and then we’ll apply that to the current month. So we’ll be discussing what is that? So basically we are using logistic regression to find out risk about default from various variables we have in our hand.

So we have seasonality we have months in servicing autopay. These all attributes these is are logistic regression but you can use various data forest random forest random decision tree, and other algorithms for the implementation. So some of the ML modules model attributes which will, will be we are looking at are so these are classified in two, three categories. First is time based, second is payment based, and third are general attributes. So time based attributes are which, which uses time as a measure that is seasonality or monthly servicing or due dates, term remaining nonsense modification. So these use time as one of the methods to forecast the lift risk for the loans. we’ll be using payment based that is changing payment, interest rates ever delinquent status autopay, unpaid balance, credit scores.

 and there are some more attributes which are like bankruptcy status property region modification status. there’s location segmentation. so segmentation data looks something like you can use various tapestry data that can be implemented and to improve the model risk by to a model improvement parameters by 20%. so model summary. in the summary we can look like what if the customer is on autopay, then there is 46% less likely. if the customer doesn’t have autopay, then it’s 46% more likely as customer who do for default, the customer who defaulted in the past three months are two times more likely to default on the mortgage. A person with high credit scores, 20% less likely to default as compared to a person having credit score between 500 and 700. So these are some of the attributes which we can come up from the model summary.

So we see that as the credit score decreases, the risk, cost default increases. as the credit, as the autopay increases, the risk of default is really really decreases. So the four criteria which we have used for finding out the model evaluation. first is A I c b i c G E, and a u C. So A I C and b I C should be less genie and a u C should increase with the more genie and E U C you can find, you can see that the model is performing better. So, and A I C and B I C should be less. So these are four criteria in which we use, but you can use your own criteria. There’s no, no limitation how you compare the model. You can use any number of criteria in which you find better to use.

So basically we have divided the loans into 10 deciles. so we can see that as the auto pay is increasing we see that the probability of default is getting lower and lower, but as the average score is decreasing, we say the risk of loan default is getting higher and higher. And you can use like any number of parameters to find if your model is performing better or worse. So first calling is something where we find after we find out the loan decile, we can see where whether the for which loan we should follow first, rather than the loan, which is not defaulting, we can drop them off and not use our outbound call sent to, to follow those loans. So basically, the diagram stays the same. Using this methods, we have decreased the cost by 30% using this phase calling strategy.

So we can see that low, low risk loans are followed late at the month, or we can just drop off following them. But the high risk loans can be followed at the start of the month. the right time to call is another use case where we use the same parameters to find with if they’re the if this is the right time to call, and we reach the right person, the right person contact to contact the lo about the loan. So we have seen that there was a one 46% relative increase in right person contact. So the, this was a huge incentive to implement this strategy. the time span of the forecast, we can use this strategy for planning decisions and daily and real time control. the strategic decisions, which are taken for two to five years is for strategy, and we won’t be using these strategies for this, but we are using for planning decisions and real time control that varies from monthly to weekly and half hourly to daily. same strategies can be applied to various loan servicing. that is auto loan, personal loan credit card loans, student loans, payday loans pawn shop loan, small business loan vacation loans. So all these the same strategies can be applied for these all loans. for inbound center, we can use various time series forecasting. but for the time constant we won’t be discussing because it, it is a huge topic in itself. we won’t be going into the details for time series forecasting. so but these are some of the various methods which can be used for the inbound call center. first is triple exponential smoothing. one of this is one of the oldest method, which has been used to smoothen various anomalies in the time series forecasting. And then we can apply the forecasting. so yeah thanks.