Intelligent customer service dispatching system

Background

When it comes to scheduling, you may think of scheduling Alibaba Cloud's massive machine resources, but for Alibaba Group's customer experience business group (CCO), what we need to schedule is not machines, but our customer service resources.

Why does customer service need to be dispatched? CCO currently undertakes the customer service business of Ali Group and the ecosystem. Our customers seek to solve various problems through various channels. The daily incoming lines are huge, and they are often accompanied by sudden incoming lines, such as Tmall vouchers. If there is a problem, thousands of hotlines or online consultations can be generated in a few minutes. In the face of various, massive, and sudden customer problems, our service capabilities are often difficult to meet, often causing users to queue up or even give up. Naturally, we have a need for scheduling.

What is the core problem of customer service scheduling? Improve the utilization rate and service level of customer service resources, and obtain better user experience with fewer customer service resources. If we recruit a large number of customer service staff, we can also give users a better experience, but it is easy to cause waste of manpower. More manpower means more training costs, management costs and labor costs.

Compared with machine scheduling, customer service scheduling has its complex points:

1) A new physical machine is added to the computer room, which can be used quickly after the machine is virtualized, but recruiting a new customer service requires long-term training to enable him to have online service capabilities;

2) There are large differences among customer service personnel, and the business skills of different customer service personnel are different. It is difficult to directly let the customer service personnel of skill group B handle the tasks of skill group A. The difference is not big, many businesses can use the same type of machine;

3) Customer service is human, and he has the right to choose to go to work or take a short break. His work efficiency and quality will fluctuate with his mood, experience, service members, and working hours. When scheduling, they need to consider their feelings, and scheduling machines without any hesitation;

4) There are many unexpected scenarios, business problems, system failures, etc. are all irregular outbreaks, and the fluctuations are particularly large. It is difficult to accurately arrange manpower for one day in advance.

Can on-site administrators cope with such complex customer service scheduling? the answer is negative. Before there was no scheduling system, on-site administrators basically relied on manual scheduling. As the volume increased, the defects were gradually exposed:

1) Slow response: For example, when queuing up online on weekends, the on-site administrator may receive phone feedback, and then turn on the computer to manually schedule a temporary shift, etc. It is normal for more than ten minutes from the queuing to the scheduling to take effect;

2) Inaccurate: Lack of data guidance and weak overall planning and optimization capabilities. For example, when skill group A is queuing up, the on-site administrator wants to cut some of the traffic of skill group A to B, how much to cut, and who to distribute to. It may be a brainstorming decision, and the decision result cannot be settled;

3) Lack of means: There are very few means available, nothing more than manual scheduling and shift shifts, manual cutting of streams, control of small breaks, and posting announcements, etc., and the ability and potential of customer service have not been fully tapped.

After clarifying the core issues of customer service scheduling, knowing the difficulties, and seeing the current situation, we decided to create an automatic and intelligent customer service scheduling system - XSigma.

XSigma Big Picture

The XSigma scheduling system can be divided into hand, brain and eye according to the functional modules. Hands are means that can improve customer service resource utilization, customer service level, and customer satisfaction, such as overflow diversion, appointment callback, on-site control, incentives, shift scheduling, emergency shift release, training, etc. With so many methods, how to choose in different business and different scenarios is a difficult point. Here, the brain, that is, the dispatching center, is needed to make decisions. How can the complex scheduling logic resulting from the decision be better understood by field administrators, business people, and developers? We transform the complex scheduling logic into an understandable real-time graphical interface through visualization technology, which is the eye of the scheduling system - the large scheduling screen. After the hand, brain, and eye functions are complete, how to make them work together better and better? We temper through the simulation exercise system.

2. Prepare in advance: Schedule your shifts

If demand can be predicted and supply ready, then customer service scheduling is half the battle. In our business, different types of customer service scheduling models are different. The cloud customer service adopts the self-selecting class mode. The administrator only needs to set the number of people who are selected for each time period, and let the cloud customer service select classes according to their own time. The SP (partner) adopts the scheduling mode, which requires the administrator to arrange each customer service according to the traffic volume of each time period. It must not only ensure the maximum connection rate of each time period, but also be able to coordinate The rest and working hours of customer service personnel should be ensured to ensure that the total working hours of each customer service personnel are roughly equal. This is a test of the administrator's overall planning ability. When the number of customer service personnel increases, manual scheduling brings great challenges to the administrator.

Regardless of the model, it is necessary to predict the required service volume in the next two weeks in advance (the business schedule is based on the granularity of 1 to 2 weeks). This is actually a standard time series forecasting problem. Combined with historical data, we can predict the service volume in the next two weeks according to the granularity of the department-skill group. Of course, this offline prediction is only an approximation, and it is difficult to predict accurately.

For the customer service scheduling of partner companies, it can be abstracted as an optimization problem under multiple constraints. In actual scenarios, we use a combinatorial optimization algorithm.

3. Horizontal expansion: predictive emergency shift

It is difficult to accurately estimate the service volume by scheduling in advance. It is impossible for us to know in advance that there will be a voucher problem at 13:25 next Monday, which will cause a large number of users to come in for consultation.

For this kind of sudden traffic or traffic that is larger than the service volume at work, can we quickly and horizontally expand a group of customer service personnel to work like dispatching machines. For socialized cloud customer service, we can do it. For example, when the number of queues exceeds a certain value, the emergency shift of cloud customer service is automatically triggered. Through practice, it is found that cloud customer service usually takes more than ten minutes from shift selection to work. How to further save this golden processing time of more than ten minutes? Upgrade emergency shift to predictive emergency shift! Predict the upcoming heavy traffic a few minutes in advance, and leave work early.

There are two models involved here, one is the real-time forecasting model of service volume, which can comprehensively predict the future of a skill group based on real-time data such as member's operation behavior, member's behavior, fault scenarios, and combined with historical incoming lines 30 minutes of incoming line volume per minute.

With the input of service volume forecast data, the emergency shift model can combine the current service member situation, customer service scheduling status in the next 30 minutes, member consumption speed, overflow relationship and other comprehensive indicators to infer whether to trigger emergency shift and release Class service volume. Once the emergency off-duty is triggered, the offline notification module will notify the appropriate customer service staff to come to work through phone calls, text messages and other means.

Different from scheduling machines, we need to always consider customer service experience. In order to avoid disturbing customer service who are not willing to go to work, we allow customer service to independently set whether to receive notifications.

4. Load balancing: overflow, diversion

Although the predictive emergency shift effect is good, it is only effective for cloud customer service at present. What should I do for non-selected customer service such as SP? We found that when queuing online, there are often a large number of queuing scenarios in certain skill groups. For example, the merchant line is bursting, and the customer service of the consumer line may be idle. How to solve this uneven busy-leisure problem? An intuitive extreme idea is to turn all groups into a large pool subgroup, and make every customer service member busy through load balancing distribution, so as to maximize efficiency. In fact, not all skill groups can take on each other. Here, we must balance the business and offline training to equip the customer service with multiple skills.

XSigma provides the configuration function of skills group diversion and overflow. As long as the trigger conditions are met, the overflow can be diverted in real time, which solves the pain of manually changing customer service skill groups by on-site administrators in the past.

For some scenarios, the overflow granularity between skill groups is a bit rough. For example, if the queue of skill group A is set to overflow to skill group B, not every customer service member of skill group B can undertake the business of A. Only those who have been trained Only customer service can take over, and XSigma also provides the function of labeling skills for customer service.

5. Vertical expansion: flexibility +1

Some businesses are more complicated, and it is difficult to find other skill groups to overflow, so we turn our attention to the customer service who is at work. Online customer service can serve multiple members at the same time. If the maximum service capacity of a customer service is 3, then he can serve up to 3 members at the same time. This value is set by the administrator according to the historical service level of the customer service.

We found that although the maximum concurrent capacity of many mistresses is the same, when they serve members at full capacity, their service levels are very different, and their busyness is also very different. Why?

There are differences in the level of the second grader

As shown in the figure below, the maximum service capacity of the customer service of a skill group is 3. In the last month, the average response time distribution of the customer service of this skill group in the scenario of serving 3 members at the same time (the average response time is proportional to the response speed of the customer service), can be It can be seen that the data shows a roughly normal distribution, indicating that there are differences in the service level of the secondary school.

The scene is different

For example, the maximum service capacity of both customer service A and B is 5, and they are both dealing with 5 members, but the 5 members of A are almost at the end of the conversation, and the 5 members of B have just started. In the example, the current busyness of customer service staff A and B is obviously different.

Since there are differences in the service levels of secondary students, and the actual scenarios vary greatly, can those secondary students who have spare capacity break through the maximum service limit when the skill group is queuing up?

XSigma provides two strategies to allow Xiao Er to break through the service limit.

1) Active +1 mode

When the skill group reaches the trigger condition, XSigma will actively light up the +1 button on the customer service workbench (as shown in the red box in the figure below), and the customer service can click to actively add a member's incoming line. This method is equivalent to handing over the expansion rights to Customer service, because only the customer service knows whether they are busy or not.

2) Mandatory +1 mode

If some skill groups are of strong control type, you can choose to enable the mandatory +1 mode, and XSigma will automatically select some suitable customer service based on the data to break through the upper limit of service capacity. For example, his previous maximum service capacity was 5, and we will let him serve 6 at the same time members.

6. Cut peaks and fill valleys: call back by appointment

For the hotline, it is impossible for the second person to answer several calls at the same time, and there are few offline customer service that can be undertaken in business. What should I do if there is a large queue at this time?

Through data analysis, it is found that the busyness of many skill groups fluctuates within a day, with peaks and low peaks. The figure below shows the remaining number of services of a skill group. It can be seen that there are two busy time periods, 10~13 From 17 to 21 o'clock, the number of idle services in these two time periods is often 0, while other time periods are relatively idle. If the incoming lines in these busy time periods can be moved to non-busy time periods, then It can greatly improve the utilization rate of customer service personnel, and can also avoid the trouble of customers queuing.

How to do it? Turn today's service into tomorrow's service by scheduling a callback. As shown in the figure below, there are mainly two modules.

1) Appointment trigger. After the user calls in, the appointment trigger will determine whether to trigger the appointment according to the busy situation of the skill group;

2) Callback trigger. Using the system's active outbound call mode, once the busyness of the skill group is found to be at a low peak, a callback will be triggered. As long as the user's phone is connected, it will enter the allocation link with high priority, so that the customer service personnel can work within the effective working hours Inside are actually talking to customers.

7. Optimal Allocation

The goal of dispatching is: "improve the utilization rate and service level of customer service resources, and obtain better user experience with fewer customer service resources". The previous strategies focus more on improving the utilization of customer service resources. Are there any strategies that can improve user satisfaction? We start with the allocation part.

In essence, what we want to solve is the problem of "membership (task)-customer service matching optimization". In the traditional mode, allocation is to find a member who has been waiting for the longest time from the queuing queue of a certain skill group, and then find the most idle customer service under the skill group to complete the match. This fair distribution method considers a single dimension, and fails to grasp various information related to membership, customer service, and problems related to scheduling and distribution at the global level.

The matching optimization problem is actually a bipartite graph matching problem. As shown in the figure, at a certain moment, we can get unassigned customers (tasks) under a skill group and customer service personnel with remaining service capabilities. If we can know the relationship between each task and The matching probability between each customer service, then the best match can be found through the stable marriage algorithm.

How to find the matching probability between tasks and customer service? Abstracted as a classification regression problem, the core is to construct a large number of samples (x1,x2,x3,…,xn)(y). For one-pass historical session tasks, y is the customer rating or session duration (target is optional), and x includes customer service features such as offline indicators such as satisfaction in the past 30 days, average response time, etc., as well as the number of service members in the current session of the customer service , the maximum number of members and other real-time indicators, and also includes the characteristics of the task, such as question type, waiting time, order number, number of repeated consultations, etc. After the samples are available, the following is to choose a classification algorithm for training, and finally we use CNN.

During the iterative process, it was found that the model will allocate more traffic to good customer service, while the traffic of customer service with relatively poor indicators will be reduced. introduced into the model.

8. Intelligent Training: Rhubarb Robot

An important reason for improving satisfaction through optimal allocation is to allocate more traffic to highly capable and high-level customer service, but the proportion of this part of customer service is not high. Why? In order to cope with the high traffic in the two special months of November and December, the business team has to recruit and train a large number of cloud customer service personnel. The influx of these novices will inevitably have an impact on satisfaction. In other words, if you want to further improve the satisfaction index, you must improve the service level of novice customer service.

For novices, the only way to improve their level before taking up the job is training. The traditional training is to let cloud customer service members watch videos and other learning materials offline, and then take a written test. If you are not familiar with the tools and solutions of the platform, you will serve members directly, and the members will feel poor.

Comparing the car training scene, we found that the car training has different processes such as subject 1, subject 2, and subject 3. Subject 1 learns theory, subject 2 and subject 3 actual combat simulation. If we introduce this kind of actual combat simulation, the service of new customer service can be greatly improved level.

We innovatively proposed a new customer service training model (patent applied for) using robots (rhubarb) to train customer service. In the training tenant, the new customer service will generate a very real simulated session by clicking on the rhubarb avatar. By chatting with members, they will continue to learn the use of platform tools and continuously improve their ability to solve customer problems. Once the conversation is over, the rhubarb robot will evaluate the conversation and tell the user that a specific solution should be used to answer the user's question.

For new customer service personnel, currently they must complete 80 conversations with Rhubarb before they can start working. Tens of thousands of customer service personnel have been trained throughout the fiscal year, and the number of service sessions has reached several million rounds. The abtest shows that the customer service personnel who have passed the rhubarb trial post have significantly improved in various indicators such as satisfaction, dissatisfaction, average response time, and average service duration.

9. Unified dispatch center

From the above, we can see that our customer service scheduling strategies are many and complex, and each strategy has played a certain role in improving the utilization rate of customer service resources and service level. Now the question is, how to choose so many strategies in different scenarios? For example, now skill group A suddenly queues up 100 members. At this time, should it directly overflow to other skill groups, or trigger active +1 or trigger emergency dismissal? A brain is needed to make decisions here.

How to make this brain applicable to various complex business scenarios is difficult. Our platform currently has dozens of tenants. The Taoxi tenant alone has dozens of customer service departments. Each department is subdivided into a series of skill groups. Different departments have different business scenarios. In the case of a serious lack of historical data accumulation, it is difficult to directly adapt to a variety of businesses by training a decision-making model. So our idea is to directly use the expert knowledge of on-site administrators to let them precipitate the decision-making logic into rules.

At present, tens of thousands of rules have been configured on the platform, and thousands of rules take effect every day. The precipitation of these data allows us to realize a real intelligent scheduling decision-making brain through intelligent optimization technology.

10. Scheduling and monitoring large screen

There are many customer service scheduling strategies and complex logic, and the scheduling results will actually affect the feelings of the participants in the whole process. Therefore, we have built a large XSigma scheduling screen to facilitate everyone's understanding. In practice, it is found that the dispatching large screen can establish the user's trust in the dispatching system, and reduce the cost of developers and administrators to discover, locate and solve system problems. For example, the administrator sets some rules on the XSigma platform. For example, the number of queuing numbers of skill group A >= 1 triggers overflow to skill group B. Let the development students confirm whether it has taken effect again, and now there is a large visual scheduling screen, which can not only observe the real-time monitoring data such as the service volume and remaining service volume of each skill group, but also see the process of real-time scheduling various strategies taking effect , as well as the real-time summary detailed data scheduled every day.

11. Simulation exercise

In the scheduling optimization scenario, how to evaluate the quality of the scheduling system is very important. Is there a way to evaluate whether XSigma can adapt to various scenarios? Can you prove in advance that it can be smoothly scheduled during the double 11 promotion period? Can problems in the scheduling process be discovered in time? This is not only what we but also business students urgently need to know.

After careful consideration, we found that the problem to be solved is very similar to that of the technical full-link stress test. What we need to do is actually a full-link stress test for business, so we built a simulation exercise system for customer service scheduling.

Based on the rhubarb robot, we have already been able to simulate member access. Through customization and transformation, the robot can create various types of topics, such as double 11 type scenarios. On this basis, combined with the estimated amount of business students, the incoming line amount of each skill group can be set.

Before Double Eleven, business students used this drill system to conduct large-scale drills twice. Because the drills were based on real service volumes, rather than the previous word-of-mouth method, it put pressure on every participating student in the upstream and downstream of the scheduling feel. After some problems found during the drill were improved, our confidence in dealing with sudden traffic during the big promotion was greatly improved.