For a D2C Fashion brand, Robylon automatically resolved 85% chat queries and 60% tickets

Being in D2C business, the customer query volume is high for our client. On a typical day, they receive 2-2.5k queries on chat and 500-600 on tickets. 70% of these were found to be repetitive and were regarding tracking orders, placing returns & checking refund status. 

Before Robylon, they were using a deterministic flow chatbot to address the queries. Hence, only 50% chat queries were automated while ticketing remained completely manual. 

To solve for this, they were looking for an AI agent that could fulfill the following criteria

1. Understand free flowing language: The AI must be able to understand the intent even if the customers use vague expressions. For example “order where” should be understood as “Where is my order” and responded to accordingly)
2. Generate personalized responses: The responses should be contextual and personalized based on users’ past interactions. For instance, AI must be able to fetch data about user’s past purchases and interactions on chats, tickets & agents to generate personalized messages
3. Seamless escalations to human agents: Certain queries which the AI is unable to solve, are to be escalated to human agents. The AI must be able to determine the type of query and route it to the correct department (e.g. payment, tracking order, etc.) 

Provide instant replies: Users want quick responses and even 10-15 seconds of wait can be frustrating. So the AI must not only be accurate, but also fast.

We studied the client’s specific requirements and devised our plans accordingly. We deployed 100+ pre-built ecommerce workflows, but also modified them as per the current context. 

Here’s a step-by-step breakdown of how the AI automates customer support:  

‍

Step 1: Intent Identification

We made an intent identification engine. It was trained on 5k+ chats and 25k+ tickets. This exercise mapped a user’s intent with a list of keywords/ phrases they actually used. 

Each new chat message/ ticket is passed through this AI engine to identify the user's intent. The intent could be easily identified even in free flowing or vague languages. 

Example: If a user writes “order not received” in the chat, the AI would identify its intent as “Where is my order?”. Then, it would check the user’s order list to figure out which order is pending to be delivered. 

‍

Step 2: Automated query resolution

Each intent identification is given a confidence score. For high confidence scores, the AI responds based on the respective SOP associated with the agent. If any additional data is needed, relevant APIs are triggered automatically and responses are accordingly generated.

Example:  For the above example, AI would trigger APIs to fetch data from the delivery partners and update the user with “Sorry for the delay, your order would be delivered by 6pm today. You can track the order via this tracking link: XXX”

‍

Step 3: Escalation to human agents

For intents with low confidence scores, or where AI is unable to generate satisfactory responses, the chat/ ticket is seamlessly handed over to a human agent. All the context is readily summarized and provided to the human agent to prevent further delays. 

Example:  In case the AI is unsure of why the particular order is still not delivered, the chat would be escalated to a human agent. “Apologies for the inconvenience. I am connecting you to an agent.”

The average per reply time is 3-6 seconds. 

‍