AI Contact Center Summarization With LoRA for Faster Agent Wrap-Up
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing contact center systems require agents to manually prepare post-interaction data, which can lead to information loss or inefficiency, and existing AI models are computationally inefficient and resource-intensive.
Innovation Solution
A computing system uses an AI model with low-rank adaptation to automatically generate post-interaction data from interaction data, incorporating historical data and agent feedback, and trains the model with transformer neural networks to enhance efficiency and reduce resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If agents manually prepare post-interaction data, then information accuracy is maintained through human review, but productivity decreases due to time-consuming manual entry
Solution Approach 1:
The system enables self-service by having the AI model automatically generate post-interaction data from interaction transcripts without requiring agent intervention. The model processes interactions autonomously, extracting key information and creating structured data that agents can review and confirm with minimal effort.
Solution Approach 2:
The patent replaces the mechanical system of manual data entry with an AI-based automated system. The transformer model processes interaction data and generates post-interaction records, substituting human manual work with intelligent automation that maintains accuracy while dramatically improving productivity.
2Productivity
If existing AI models are used to automate post-interaction data generation, then productivity increases, but computational resource consumption increases excessively
Solution Approach 1:
The system changes key parameters of the AI model by using low-rank adaptation (LoRA) technique. Instead of training full-sized models, the patent applies parameter-efficient fine-tuning that modifies only a small subset of parameters (rank decomposition matrices), reducing computational resource requirements while maintaining generation quality.
Solution Approach 2:
The patent applies local quality by focusing computational resources only where needed - using low-rank adaptation to modify specific parameter matrices rather than entire model weights. This localized approach to model adaptation reduces overall computational burden while preserving the model's ability to generate accurate post-interaction data.
3Measurement precision
If full AI model training is performed, then model accuracy is maximized, but storage requirements and computational costs increase
Solution Approach 1:
The patent extracts only the essential components needed for accurate post-interaction data generation. By using low-rank adaptation, the system separates the pre-trained model weights (which remain frozen) from the trainable rank decomposition matrices, storing and training only the smaller adapter parameters rather than entire model weights.
Solution Approach 2:
The model parameters are segmented into frozen pre-trained weights and trainable low-rank adaptation matrices. This segmentation allows the system to maintain high accuracy through the pre-trained knowledge while requiring minimal storage and computational resources for the adaptive components that are actually trained and stored.
Data Source
AI summary
A method for enabling efficient performance of post interaction operations in a contact center may include obtaining, by a computing system, interaction data indicative of an interaction between a client and an agent of a contact center. The method may further include producing, by the computing system, with an artificial intelligence model trained with low-rank adaptation and as a function of the obtained interaction data, post interaction data to enable the agent to efficiently proceed to a subsequent interaction associated with the contact center. The post interaction data may be indicative of a summarization of the interaction. The method may also include providing, by the computing system, the post interaction data to the agent for confirmation. Further, the method may include storing, by the computing system and in response to receipt of agent feedback, the post interaction data in a data set of the contact center for analytics.