Case Study: Advanced natural language processing use cases for AAA AUTO
Recordings of calls in a call centre contain a large quantity of information about the customers and their needs, the operators, and the communication practices in general. More and more companies realise that this data has a critical potential for advancing the operations and impact of call centres. Natural language processing uses machine learning to process and interpret text data, that can be used for customer grading (call prioritisation), operator oversight (call scripts matching, detection of challenging calls), topic detection and much more.
DataSentics helped us implement NLP (Natural Language Processing) into our call centre – the key lead management tool for both selling and buying. Before the NLP application, our analysis depended on hand-written data submitted to a CRM application by our operators. Thanks to the great work of DataSentics, we can now analyse the conversations between our operators and customers, measure sentiment, and predict business potential. All of this helps increase our efficiency and provide better service to our customers. None of this would be possible without the expertise of DataSentics in the area of applied machine learning.
Viktor Cirbus | Head of Data Science & Group Digital Officer at AAA AUTO
AAA AUTO is one of the biggest used car dealerships in Central Europe, with more than 29 years of experience with used cars and over 2000 professionally trained employees. They have a very well-established call centre where all calls are recorded. Storing such an enormous amount of valuable data gives them a great opportunity to improve their customer service with modern AI/ML-based solutions.
Although the client has access to tons of useful data, it was not utilised to the maximum. The calls were recorded but not transcribed to allow for a full-text search or further analysis. Operators wrote notes manually after each call, which was extremely time-consuming, and there was a lot of room for mistakes or loss of relevant information. They have been using customer grading, but only for customers who sent their requests via an online form.
In order to start transcribing calls and use them to their full potential, we introduced a finetune speech-to-text technology. We wanted to improve the tools for call centre managers and use the transcripts to prioritise customers in the lost zone (=customers who were supposed to perform an action but did not, e.g., those who did not visit the branch to see a car after scheduling a meeting). We were able to do this using propensity grading based on these transcripts. Giving the salespeople at a branch office extracts of the customers’ conversations with the call centre on a relevant topic would lead to a better understanding and business continuity with the customers.
DataSentics delivered a complex natural language processing (NLP) solution based on Azure cloud infrastructure and Databricks. The first step was to finetune the Azure Speech-to-Text service to perform best on the customer’s domain. Next, we used this service to transcribe a significant amount of their calls.
The transcripts were used as a source for NLP features predicting the customer’s tendency to buy a particular car. We developed and trained a propensity model that uses these language processing features. As it turned out, it significantly outperformed the original static grading that the client had in place up until then. This will allow our client to focus the call centre resources on the customers with the highest potential profit.
Finally, we developed and implemented an algorithm focused on detecting and searching for topics in the transcripts. This summary is meant to serve mainly the salespeople in the branches, where each customer finalises their journey during their visit, look at and potentially buy a car.
If displayed in a mobile app, the topic summary (e.g., who will be driving the car, its primary usage or whether the customer talked to competition) allows the salespeople to see the key points of the customers’ previous conversations with the call centre. They can use it to provide a personalised experience to the customer and read and listen to the relevant parts of the customer’s conversation on the relevant topic.
- Searchable transcripts enable better monitoring of call quality and feedback targeting to individual call centre agents.
- Improved propensity model allows allocating operators effectively on high potential customers, Profit per Customer in the top 20 % of customer base was increased by 29 %.
- Automated topic detection provides more context to the salespeople while giving the customer more continuity, both of which can improve the salesmen’s chances of closing the deal.
In the future, we plan to implement a sentiment analysis, which will give the client an insight into customer satisfaction and detect problematic calls. We also want a rollout to the Polish and Slovak branches of the business.