Beyond Chatbots: Transforming Customer Service with Personalized Precision

When I started my career, my first job was to set up a call center for an Asian Telecom operator. It was a bustling call center, with more than 700,000 calls landing daily. The average talk time per call was approximately 90 seconds, which is nearly two years of talk time in one single day. We had 3,000 full-time human agents handling those calls. As a customer, do you recognize this prompt: "This call will be recorded for quality and training purposes"? The call center manager always asked me to help him select which call to listen to for "quality and training" purposes. We had no choice but to choose something like a three-sigma rule and look for call lengths outside the norms.

We did a cost analysis for the calls, and we found out that if the call was self-serviced via the IVR, then the cost of the call to business was 1.5 Rupees; if the caller pressed '0' to talk to the human agent, the cost of the call jumped up to 44 Rupees. This is a staggering 30X increase in the cost to serve. You multiply that cost by 700,000 calls, and you are talking about nearly One Million $ per day, and across the whole year, the operating cost jumps up to a staggering 365 Million $. Of course, this is the worst-case scenario, and not all calls land at the human agent. This is the scale of the problem many customers in the customer services industry face today.

TM Forum Catalyst

TM Forum , a non-profit telecom industry body, plays a crucial role in fostering collaboration between digital service providers, technology suppliers, consultancies, and systems integrators. Its primary aim is to address the most pressing challenges in the industry. One of its key initiatives is the incubation of proof of concept projects called catalysts. These projects, championed by telecom service providers, partner with various technology solution providers to develop innovative solutions.

du (Telecom service provider in the UAE) and TurkNet (Broadband provider in Turkey) came up with the problem statement of modernizing the customer service industry. The catalyst was called "Beyond chatbots – Revolutionizing telecom with advanced generative AI". Snowflake , Amdocs , EY , and solvatio AG came together to work on this problem.

It is an "AND" story, not an "OR" story

When people think about AI, some doomsday predictors talk about it taking over the world. They believe it is either human or AI, but when we think of most tasks today, there is an augmentation story where humans and AI can work together. In the book, Human + Machine by Paul Daugherty and H. James Wilson discuss how AI can augment human capabilities.

They discuss how the bulk of customer service representatives' work could be broken down into 13 existing tasks. They then analyze how the introduction of generative AI might affect each task. Four of the tasks remain unchanged and can be performed entirely by humans. Four tasks can be fully automated. Five tasks can be augmented to help humans work more effectively.

The Solution Prototype

To build the solution, we used Snowflake Cortex. Snowflake Cortex is a suite of AI features that use large language models (LLMs) to understand unstructured data, answer freeform questions, and provide intelligent assistance. The solution followed the following blueprint for Call Center Analytics with Snowflake Cortex.

This above solution architecture shows how to build a customer insights analyzer app using Snowflake Cortex, Snowpark Container Services and Streamlit .

• The solution uses OpenAI Whisper that transcribes audio files to text for analysis
• Using Snowflake Cortex LLM functions, customer details, sentiment analysis, call summary and resolution are extracted
• A RAG based AI app is built to ensure contextual responses for user queries.

With Snowflake Cortex, you bring the applications to the data rather than sending the data to the applications. Through the Snowflake governed environment

If you want to build this solution in your own environment, you can follow this quickstart by Phani Raj and Karunakaran T Nadadur .

The Menu of One

Most call center agents have dozens of applications open on their screens, and they constantly switch between different screens. Using this solution, the agents don't have to switch between multiple screens, and the solution can predict why the customer has called because we have captured all of the customer interactions with the telecom provider. Hence, we refer to the solution as a "Menu of One." Based on our deployment, we have seen that in more than 70% of the scenarios, the solution correctly predicted the reason for the customer's call before the customer stated the problem to the agent.

Imagine your customer hearing: "Are you calling about your last bill?" Instead of the usual: "How can I assist you today?"

Say goodbye to long phone menus. Embrace the menu of one, where we’ve set a new customer-service benchmark by using Generative AI to predict customer intent, enabling first-contact resolution through IVR, call centers, and chatbots. This solution provides a quick, personalized, and efficient customer experience while reducing call-center costs.

How it works?

Predict Customer Intent and Seamlessly Route Customers to a Gen-AI-Powered Virtual Agent. Our solution harnesses the power of Generative AI and ML to revolutionize customer service. 

Innovative Analysis:

1. We used GenAI in a non-standard way, not just to chat with the customer, but by combination with ML, to create outstanding models, enabling high accuracy in anticipating customer needs.
2. This allows us to label call reasons and key discussion topics accurately.
3. These labels form the foundation for our ML Intent-Score models.

The use of GenAI:

• We utilized large language models to discover the customers' intentions from past interactions, label the call reasons, and analyze them.
• Implement use cases like Bill Shock, International Roaming, AI-Augmented Customer Operations, and Proactive Network Maintenance.
• Each use case targets specific customer needs, enhancing the customer experience and operational efficiency.

In our solution, if we can predict with high confidence, then the call will be answered by IVR automation; if the score is medium confidence, then we will ask additional questions, and in case of low confidence, we will route them to typical IVR experience. Once the system learns about the customers, overtime we will see most customers being served through automation with high confidence.

Outcomes

Based on the solution concept, we were able to showcase the following:

1. 90% First Call Resolution
2. 40% reduction in call duration reaching human agents
3. 50% reduction in operating cost,
4. 32% reduction in average handling time.

The solution was selected as a finalist in three categories at the DTW - Ignite conference in Copenhagen:

• Outstanding Catalyst - Innovative and Futuristic
• Outstanding Catalyst - Rising Star
• Outstanding Catalyst - Application of AI & Automation

Conclusion

If we were to map the catalysts on the TM Forum Autonomous Model by Dean Ramsay then this solution would fall somewhere between levels 4 and 5; with an option for human override.

This is not just another GenAI chatbot solution. It represents a shift in customer interaction, anticipating customer needs with high accuracy in prediction and seamless integration, setting a new standard for efficiency and customer satisfaction. We hope that the solution can bring benefits to other customers looking to optimize their call center operations as well.