Feb 17, 2020
Using ML and NLP to Transform Customer Support
artificial intelligenceB2B supportbuild vs. buyITSMmachine learning
Welcome everyone! I am the VP of Engineering at SupportLogic and I’ll be contributing here to our blogs to discuss a wide variety of topics related to the machine learning (ML) techniques that we are leveraging at SupportLogic. Additionally, we will discuss the challenges around operationalizing these ML techniques, and building a fully integrated solution that is capable of affecting change across an organization and delivering measurable value.
It’s an exciting time in the artificial intelligence (AI) space, with an abundance of open-source ML software to experiment with and build from. For instance, Google open-sourced BERT, an example of transfer learning, that is capable of accurately detecting extremely nuanced sentiment. However, hardly a day goes by without another article being published discussing the general disillusionment with AI in the corporate world and urging everyone to temper their expectations.
I’ve been at SupportLogic for two years. Previously, I’ve focused on building enterprise data science platforms, and had the privilege to work with amazing data scientists, who successfully analyzed terabytes of data using sophisticated machine learning techniques. However, in every case, we struggled with the ‘last mile’ problem — how do we integrate these predictions into people’s everyday workflows, such that the insights actually become a catalyst for change. In reality, the outcome of these data science projects is an inspiring powerpoint presentation and little more.
Operationalized ML solutions are rapidly shifting the way companies think about leveraging ML, and Support and ITSM are especially ripe for transformation via ML for a few key reasons:
Support is frequently extremely manual:
- Metadata: Look at a case in any CRM and you’re confronted by dozens of metadata fields that must be manually entered, updated, corrected and validated by both the agent and the customer. Sadly, in many instances these fields are null or erroneous, impacting support quality.
- Case Prioritization: Customers are asked to rank and self-report the criticality of their support tickets, frequently resulting in a disconnect between the customer and the company with respect to the severity of the problem.
- Case Assignment: In many companies the assignment of agents to tickets is fairly ad hoc causing a sub-optimal experience, delays in resolution time and costly ticket reassignment.
Support is critical in today’s subscription economy: While in the past support may have been viewed as a cost center, now support is critical to customer loyalty and revenue generation from upsell / cross-sell activity.
Support has a wealth of data waiting to be unlocked: Support conversations with a company’s customers provide invaluable insights into how customers are using the company’s products, their pain points, views on the product (strengths, weaknesses, shortcomings, and missing features), potential competitive threats and upsell / cross-sell opportunities. If this information can be mined from the tickets, it can prove to be invaluable for further revenue generating activities (product improvement, upselling, churn prevention, future marketing messages and engineering bug prioritization).
Effectively training the ML is only half the battle in any data science project! Operationalizing the ML is just as critical, and frequently, almost as hard to do. Significant effort is required to build an operational deployment; UI dashboards, connectors to pull data and disseminate the findings, real-time alerting infrastructure, glue logic, automated model monitoring, retraining and uptime guarantees are required for a production-grade solution. And, for an internal company project, it’s very difficult to marshal the necessary engineering and data science resources and keep them in place for the long term to ensure that the solution remains operationalized and doesn’t immediately become stale.
Happily, as the ML space matures and ML algorithms continue to commoditize, it becomes even easier to buy these solutions rather than be forced to build them internally. In the last couple of years, there has been an explosion in the number of companies offering complete pre-packaged solutions for a wide variety of problem spaces. Companies are no longer required to spend months (or even years), and significant monetary investment, attempting to build their own solution in house, but can instead buy and rapidly deploy a third-party solution to solve their problem.
Ultimately, it’s why I decided to join SupportLogic – a startup focussed on using ML & NLP to transform the Support and ITSM spaces.