Six Figure Savings: Transforming Customer Service With AI

Posted 2020-07-27 Posted by Tom O'Connell

Pivigo worked with, a furniture retailer making high-end lifestyle design accessible. MADE has a direct-to-business model that cuts out the middle-men, working directly with designers and customers. Customer service is crucial and meeting customer needs and responding to queries is at the forefront of what they do. MADE is also an online-only retail platform which is an ideal environment to accumulate data that can be analysed to improve elements of the business.

A group of aspiring data scientists from Pivigo’s Science to Data Science bootcamp worked intensively over five weeks to automate customer services, as MADE faced huge volumes of queries and found it difficult to keep up with demand.

Customer service is crucial to MADE, as they cut out the middle-men, working directly with designers. However, customer service agents face huge volumes of queries and MADE were looking to automate part of this service to keep up with demand.

MADE were dealing with 2300 queries a day, 800 submitted by email, with an average response time of two days. The S2DS team divided these queries into ones which were easy to respond to, i.e. a one-line reply would resolve the query and then those which were judged to be more complex. Of those more complex queries, a categorisation model built by the S2DS team found a relevant automated response from more than 100 category options.

The team analysed more than 17 million data points, both structured, orders and customer information, and unstructured data, correspondences with customer service.

To understand which response could be sent by the model, it required NLP (Natural Language Processing) to decipher the content of emails; as although they would be easy for a human to understand, without this, an AI would struggle.

This process can be adapted to fit many business problems; Pivigo removed the need for a human to check 100,000 job descriptions for error for a large recruitment firm, by creating an NLP model which cut down unique job titles by 70% and eliminated 10% of job descriptions with inaccurate job skills required.

The MADE team’s first classifier used a technique called Logistic Regression to decide whether a query could be automated end-to-end. Once that decision was made, it needed a second model to produce the most appropriate response.

Again, Logistic Regression had the greatest precision when choosing a category for the query; for instance, the sentence, “I write you to know if you also deliver a sofa in Italy,” could be placed into an ‘Overseas Delivery’ category and the correct response sent by AI.

Pivigo’s work to automate responses to customer queries achieved substantial reductions (40%) in response times with an estimated £100,000 in savings to MADE, delivered in just five weeks. To hear more about the outcomes of this project and how it was achieved, watch the webinar in full here.

Get priority access to Pivigo news, features, events and networking opportunities