Image result for customer service machine learning

Source: PCMag

Customer Service is likely one of the most complex and frustrating parts of your business, but it doesn’t have to be. Machine Learning is making strides in automating and improving parts of the Customer Service (CS) stack quickly, like auto-routing tickets to the right agent or improving your knowledge base. Our Vertical Spotlight on Customer Service will give you all the information you need to get started.

All of our vertical spotlights use our Machine Learning Vertical Framework: we analyze unique use cases, leadership, domain specific problems, and model tradeoffs.

Unique Use Cases

There are more than a dozen relevant use cases for Machine Learning in Customer Service, so we’ll just cover a few of them. Use cases can generally be split along two axes: improving the customer experience and saving the big bucks.

Improving Your CX

Detecting Customer Intent

Inefficiencies in Customer Service are often driven by information asymmetry: your company doesn’t know what your user needs help with in advance. Machine Learning can help predict why a customer is contacting you based on previous contact patterns. That allows your team to prioritize in advance and run more efficient workflows.

Right Channel, Right Time

In today’s technology landscape there are a number of channels to reach customers on: website, SMS, Facebook Messenger, and even more in developing economies (think WeChat, WhatsApp, LINE). To make your customers happiest with their resolutions, you want to reach them how they want to be reached and when they want to be reached. Machine Learning can help predict what those variables are based on past outcomes and similar customer profiles.

Intelligent Agent Routing

As your operation grows, you’ll have different agents tackling different cases, priorities, and personalities: accordingly, your customer experience could be meaningfully improved by pairing the right agent with the right case. Machine Learning can use your historical Customer Service data to suggest those pairings quickly and automatically.

Saving on Customer Service Costs

Similar Cases Classification

If your agents have an easier time deciding how to resolve cases, they’ll save time and address more tickets. Machine Learning (clustering) can find similar cases to the current one that your team has dealt with before, and recommend a resolution based on what has worked in the past.

Automatic Data Capture

After resolving an issue, your agents most likely spend time manually recording the details of their customer interaction. That’s time that could be spent on core ops: helping customers. Machine Learning (NLP) can help transcribe conversations and pull relevant data from a conversation so your agents don’t have to.


Perhaps the most ubiquitous and maligned element of Machine Learning in Customer Service, chatbots help your operations by answering customer queries automatically. Now I know you’re thinking “what customers want to talk to a robot and not a person?” – but these bots can have utility when used correctly. Many customer tickets are simple and can be answered automatically: the question is just which ones.

Source: InTheChat


As far as Machine Learning verticals go, Customer Service is relatively mature. There are a number of companies implementing it, startups selling it, and Machine Learning practitioners who are experts in it. One of the leaders in the space is Uber and their COTA system.

COTA stands for Customer Obsession Ticket Assistant (catchy, I know), and it’s a system Uber has designed to aid their agents in satisfying customer tickets. It’s important to note that COTA exists to assist Uber’s Customer Service teams (hence the “Assistant” moniker) – not to replace or automate them out of their jobs.

Source: Uber Engineering

The way COTA works is actually pretty simple: any time a new customer request hits Uber’s Customer Service backend, it gets pushed to the Machine Learning models they developed using their Michelangelo platform. The model scores the customer request and sends it back to the frontend, which eventually presents the agent with 3 suggestions for how to resolve the ticket.

Early results for COTA are promising: Uber ran controlled experiments that showed that the system was able to reduce average ticket handling time by 10%. For a company managing hundreds of thousands of requests per day, that’s a pretty significant boost. They were able to achieve an even higher accuracy bar when they added Deep Learning into the mix, too.

Uber isn’t the only company building internal software to handle Customer Service, though. Bank of America released their digital assistant Erica to much fanfare in March, a service that was built internally. Capital One’s version is called Eno, and it works through SMS. As users become more comfortable interacting with digital personalities, expect this portfolio to expand.

Domain Problems

Customer Service is an unadulterated, direct line to your users: that means any modeling you do needs to be carefully tested and monitored so as not to destroy your user experience. This thread informs Machine Learning in Customer Service’s domain-specific problems.

User Experience: Direct

If users have one bad experience with your product or service (and certainly if they have many), they’re much more likely to churn over time. Any measures you take in automating or improving your Customer Service with Machine Learning need to be mapped directly to how they’ll impact the user experience practically (e.g. directing a user to a chatbot who can’t answer their questions).

Source: Boston Interactive

User Experience: Indirect

Even if users don’t feel negative about a particular interaction, Customer Service is a holistic operation. If things begin to take longer, routing pairs users with the wrong agent, or intent is misinterpreted, these have subconscious effects on how you’ll be perceived. Avoiding a negative customer experience isn’t only about not leaving them on hold: it’s about delight and ease.

Data Volume

Models need data, and that data isn’t always easy to come by. Small or medium businesses haven’t dealt with that many customer tickets, which means they may not have enough data to train models properly. Even large enterprises like banks, who do have years (or decades!) of Customer Service experience and records to draw on, face data format, source, and engineering issues.

Model Tradeoffs

To address the sensitivity around impacting the user experience negatively, model and system design in Machine Learning for Customer Service will often distance a model’s predictions from any interactions with users. Uber’s COTA system is a good example: it’s a system fully integrated with Machine Learning, but model outputs aren’t served as actions to users: they’re presented as options to agents.

As a Data Scientist or Architect building one of these systems, it makes sense to distance your predictions from the user as much as possible while still maintaining utility. Even once you’re confident in your predictions, any number of factors can spontaneously impact model performance, and the user experience is too sacred to hamper.

Another tradeoff to think about is model type and size. In Uber’s post about COTA, they showed how their first system utilized more traditional Machine Learning methods, and over time they were able to show meaningful improvements using Deep Learning. Practitioners’ first instinct might be to jump to Deep Learning, but recall that DL models generally require orders of magnitude more data than their simpler counterparts. If you’re a business just getting off the ground or one with data integrity and cleanliness problems, you might want to start simple and build up over time.


Machine learning implementation strategy for a customer service center (365 Blog) – “Many customer service centers are already thinking about adopting machine learning for their day to day operations and these techniques will soon be a part of industry standard best practices. It is therefore imperative for all call centers to adopt new age technology for improving their performance metrics and to stay competitive. This post describes how traditional call centers can create a strategy for adopting machine learning capabilities by evaluating the technical capabilities offered against their KPIs.”

COTA: Improving Uber Customer Care with NLP & Machine Learning (Uber Engineering) – ”Leveraging our Michelangelo machine learning-as-a-service platform on top of our customer support platform, COTA enables quick and efficient issue resolution for more than 90 percent of our inbound support tickets. In this article, we discuss our motivations behind creating COTA, outline its backend architecture, and showcase how the powerful tool has led to increased customer satisfaction.

Use Cases of AI for Customer Service: What’s Working Now (techemergence) – “While the technology is not yet able to perform all the tasks a human customer service representative could, many consumer requests are very simple ask that sometimes be handled by current AI technologies without human input. In this article we’ll shed light on the current trends and use-cases that business leaders should be considering today.

Automatic Insights: How AI and Machine Learning Improve Customer Service (Entrepreneur) – “Recently, Air Canada wanted to learn more about its mobile app users to identify new opportunities for customer experience improvements. That’s why the company entered into a partnership with technology provider Glassbox, that company’s CEO, Yaron Morgenstern, told me in an interview. Air Canada needed to know more about the customers using its mobile app — including what uses they preferred and which devices they were using. Glassbox’s Digital Behavioral and Digital Experience Performance Analytics solutions was the tool used to glean this information.