Virtual Agent Intent Prediction and Confidence-Based Deflection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing customer support systems face inefficiencies due to automated conversational agents (VAs) being unable to accurately interpret customer queries, leading to increased time spent by human agents in handling conversations and a lack of timely assistance for customers.
Innovation Solution
A method and apparatus that predict customer intents and compute confidence scores for VA responses, deflecting conversations to human agents when confidence is low and back to VAs when confidence is high, ensuring timely and effective assistance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If the VA is configured to request the customer to ask the query in a different form when the customer query is not clear, then the VA can maintain its automated operation, but the query disambiguation takes longer and the customer does not get the best experience in a timely manner
Solution Approach 1:
A human agent serves as an intermediary to review ambiguous customer queries that the VA cannot confidently interpret. The human agent provides disambiguation guidance to the VA, enabling the VA to learn from these interactions and improve its interpretation capabilities over time, reducing the need for repeated disambiguation requests.
Solution Approach 2:
The system implements a feedback mechanism where human agents review VA performance on ambiguous queries and provide corrections. This feedback loop allows the VA to learn from human expertise, improving its intent recognition accuracy and reducing future disambiguation needs while maintaining automated operation.
2Reliability
If the VA deflects the conversation to a human agent when it cannot interpret the customer query correctly, then the customer receives accurate assistance, but a large number of conversations are handed over to human agents increasing their time spent substantially
Solution Approach 1:
The patent replaces the traditional mechanical handoff mechanism with an intelligent learning system. Instead of directly deflecting ambiguous queries to human agents, the system uses machine learning models to predict customer intent and only deflects when confidence thresholds are not met. This substitution reduces unnecessary human agent involvement while maintaining accurate customer assistance.
Solution Approach 2:
The system dynamically adjusts the confidence threshold parameter for query interpretation. By optimizing this parameter, the system balances between maintaining high customer assistance accuracy and reducing the volume of conversations requiring human agent intervention, thereby reducing human agent time spent.
3Adaptability or versatility
If the VA uses NLP algorithms and special grammar to interpret customer's natural language inputs, then the VA can respond to various queries, but the VA is limited in its ability to provide assistance due to the sheer variety of requests
Solution Approach 1:
The system performs preliminary learning by having human agents review and correct VA interpretations of ambiguous queries before the VA encounters similar queries in production. This preliminary action allows the VA to pre-load knowledge from human expertise, improving its ability to handle diverse requests reliably without sacrificing adaptability.
Data Source
AI summary
In a method and apparatus for facilitating agent interactions with customers of an enterprise, one or more intents corresponding to an input provided by a customer during a conversation with a Virtual Agent (VA) are predicted. A confidence score corresponding to each intent is computed that is indicative of an ability of the VA to provide an effective response to the input. The confidence score corresponding to each intent is compared with a predefined threshold score. If the confidence score is less than the predefined threshold score, the conversation is deflected from the VA to a human agent to respond to the input of the customer. The conversation is deflected from the human agent to the VA for a subsequent input if a respective confidence score of at least one intent predicted for the subsequent input is greater than or equal to the predefined threshold score.