AI for Customer Service: Actual Projects Insights

AI and Machine Learning in Customer Service – the Question is already HOW, not WHETHER

Artificial Intelligence and Machine Learning have been among the most discussed topics lately.

Even though the degree of enthusiasm differs, most experts agree that ML and AI are going to influence modern business landscape and trigger business transformations in almost every industry.

• 55% of established companies either have started making investments in the potential of artificial intelligence or are planning to do so by 2020. – Gartner.
• AI technologies are projected to increase business productivity by average of 38% by 2035. – Accenture.

However, for some business areas ‘those trendy things’ are already a part of daily routine. Customer service was among pioneers that acknowledged ML and AI potential and started utilizing it. Recent researches report that:

• Already today 21% of consumers see chatbots as the easiest way to contact a business – Ubisend.
• Eight out of 10 businesses have already implemented or are planning to adopt an AI-enabled customer service solution by 2020 – Oracle.
• AI bots will power 85% of customer service interactions by 2020 – Gartner.

Artificial Intelligence for Help Desk

When it goes to AI and automated customer service, most of our customers would ask for some intelligent chatbot. That’s probably the most widespread AI application for the support area. However, we prefer using a holistic approach to AI-powered customer service rather than limiting wide AI and ML capabilities to chatbots only.

In most cases when we talk about call center automation and increasing customer support efficiency (without being bound to the AI concept at all), there is a number of areas for improvement. And guess what – being applied properly, AI and ML can contribute to better KPIs in all of those.

Table: Vulnerabilities and Possible AI Improvements in Customer Service

Best Customer Service Chatbots to Relieve 20-30% of the Load

As the hottest topic in 2017, AI support bots were expected to be a game changer. And even though the ebullience faded a bit when it became clear that even the best customer service chatbots still have a lot to learn, a properly designed and trained chatbot can substantially relieve the load on 1-st line support.

As statistical data show, 87% of chatbot users worldwide consider them efficient or very efficient in resolving issues.

As of today, there is a number of proven and efficient 3-rd party AI customer service chats, namely:

• Watson Assistant
• Dialogflow
• Wit.ai
• LUIS

Those can come in handy when no sensitive data is going to be handled by customer service bots, e.g. for pizza or theatre ticket ordering, common information bot, etc.

However, in some cases, chatbots for customer support handle sensitive information (e.g. internal corporate know-how, pricing info, etc.). In those cases, an in-house chatbot looks like a better option. E.g., in one of our projects we have developed a virtual assistant for 1-st line support engineers. What it does is simply:

• Answering questions (with AI-enabled search in the knowledge base)
• Executing commands (e.g. check ticket status)
• Returning documents (by document name)

Introducing this simple virtual assistant resulted in 1-st line support engineers being able to process 12-18% more tickets per day. While for more sophisticated assistants and chatbots the figures are even more optimistic: on average, an AI chatbot for customer service can take over 20-30% of the 1-st line support load.

Efficiency of implementing an in-house chatbot depends on a wealth of factors. Thus, feasibility of doing this should be evaluated depending on a business case. Going for a small PoC project before moving to a fully-fledged chatbot development can also be a good idea (please refer to a Case Study on such a PoC).

ML for Smart Tickets Handling

Accuracy in classifying a ticket is one of the key tasks for customer service automation influencing resolution time, quality of ticket handling, and customer satisfaction level. In our artificial intelligence customer support projects, we have so far focused on three major areas for improvement.

1. Classifying tickets by problem area and severity

Thanks to Natural Language Processing, AI-enabled models extract data from a ticket, define its problem area and severity, and trigger corresponding workflows automatically. Or return suggestions on problem routing to an operator in simpler solutions.

In one of our projects, we trained a neural network model to classify between low and high severity. Around 700,000 tickets were split to train, test and validation sets with test and validation sets not involved in training. Model accuracy on a validation set was 79% which is a very good result for a relatively simple model.

2. Suggest similar tickets to get ideas about how the task can be solved

Once a ticket is registered in the system, the probability that there is a similar case resolved in the past is quite high. Prompting similar tickets shortens resolution time as support engineers will have info on steps and actions taken in similar cases.

3. Find the most relevant engineer to solve the task

With an AI-enabled solution defining similar tickets in the system, it can also find engineers who have already successfully resolved those cases or the ones who worked in related areas. Those engineers are more likely to resolve the new ticket as well and are prompted by the system as suggested engineers. In one of the projects we have delivered, the system provided a list of 5 engineers who would be best-fit specialists to take over the task allowing call center specialists to manually select from the list based on the info about engineers’ availability. If integrated with a resource management system containing workload information, the selection can also be done automatically.

The model developed was trained on 580,000 tickets with historical data and tested on 145,000 tickets. Testing results boasted 78% accuracy of the model. This means that in 78% of cases the engineer who was assigned to the ticket and successfully resolved it in the past was in the list of 5 specialists suggested by the model as best-fit.

Predictions and Forecasting with ML tools

Forewarned is forearmed, and one would hardly doubt that accurate forecasting is essential to any business. ML forecasting in customer support includes but is not limited to:

1. Forecasting peak loads for proper resource management so that you are prepared and have some aces up in your sleeve when the peak comes.
2. Predicting time needed to solve a request for customer expectations management and efficient resource planning.
3. Predicting possible issues and preventing support requests by taking preventive actions instead of reacting to the issue when it’s already there.

Organizing Handy Knowledge Base

A comprehensive and handy knowledge base is fundamental for support operations. All documentation, regulations and policies, FAQs, instructions, and historical data on previous issues form the basis for fast and high-quality service.

In ideal world your knowledge base would be a priceless asset. But in real life it quite often turns into a massive of partially structured information which is hardly searchable.

For example, using cosines similarity or pre-trained RNN comparison instead of traditional context search substantially increases search accuracy.

Conclusion: AI/ML is a Game Changer in Customer Support

As you see AI and ML benefits for customer service are already obvious. AI and ML are definitely changing the rules of the game in the customer support industry and the use of AI/ML technology in this industry will only grow.

However, AI and ML are rather additional tools, which can amplify but not replace traditional customer service automation software and savvy approach to automation. No matter how intelligent your systems are – they are just the reflection of your organization and your processes, which should never be neglected.

• #AI/ML
• #Information Technology
• #Support & Maintenance