Chatterbox NLP Framework - Use Case - Customer Churn

Find dissatisfied customers in real-time using Statistical Machine Learning

The Problem

Losing customers to a competitor is an expensive business; better to identify dissatisfaction early and act to stop them leaving. The ‘churn’ of customers is inevitable; however in the telecoms sector rates of annual churn averages vary between 10-67%. The cost of losing a single customer and then acquiring a replacement customer is expensive; running into hundreds of dollars. ‘Churn’ is a major and expensive issue for telecoms companies.

Customers lose patience and get upset with your service – whether that is a billing error, slow network, service is down, an inability to contact customer services or sales. More and more they express their frustration online via social networks or via text message to friends and family. It is an impossible task for a customer services team or helpdesk to be able to track and manage the scale and frequency of messages related to performance at any given time. The solution for a partner’s client was to build a machine learning solution using our NLP Framework.

The Solution

Our partner used our NLP Framework to create highly customised classifiers around particular services provided by their client. Utilising historical data linked to customer dissatisfaction the classifiers were trained and iterated to provide a high level of accuracy. Our classifiers understand natural language and a mixture of slang and emoticons as used in short form messaging; they are smart and learn as we utilise statistical rather than rules based NLP machine learning techniques and processes. This is custom classification based on the client’s own data; specifically trained using our statistical machine learning methods.

The new product with trained classifiers was deployed in a single day within a CRM environment connected to a live feed of short form social data. It immediately started to capture and assign different levels of customer dissatisfaction in real-time for the client’s customer service team to prioritise and proactively engage and manage. The deployed solution continues to improve its accuracy as it keeps learning; this is the advantage of a statistical machine learning approach rather than a rules based one.

Example Messages

“signal is a joke! Can't watch more than 2 minutes without it cutting out!”

“I was assured my call barring facility would be set up today. It isn't. Can you resolve this for me?”

“overcharged for an upgrade. Phones put down twice, live chat unwilling to transfer me.” “after spending 53 mins on the phone , my sim still hasn't be swapped. Your customer service is the worst” “No service all day - paying through the roof for a phone I cant use”

“@companyA..your service is rubbish – I’m moving to “companyB”


Finding the messages in real-time allowed the issues to be highlighted and sent to a customer services team member to take action. They were fully empowered to resolve each customer issue dependent on the severity of service issue; or how close the customer was to leaving the provider. The critical issue was to react in a timely fashion; ‘in’ moment and to give the customer a voice.

Early Results

The results to date have seen a decrease in the overall % of customer churn; on a like for like basis a 15% improvement. The customer services team are made aware of issues in real-time and a specialist team now handles the issues highlighted by the product.

The deployed product is viewed by the customer as a critical piece in the overall strategy of reducing churn and improving customer experience and overall profitability.

Future Plans

The customer plans to broaden the remit of our statistical machine learning approach to encompass a range of other data sources within the organisation. They have data from a wide range of sources including web usage, credit information, call history etc and there is scope for using machine learning to predict the likelihood of a given customer leaving and moving to a competitor.

For more information please visit our web site or contact Andrew Watson, VP Strategic Alliances,