In the initial phase of the engagement, CloudCookies dedicated their efforts to enhancing the accuracy of customer intent recognition and routing. Their goal was to achieve a minimum accuracy rate of 75% for routing customer intents and an overarching engagement goal of exceeding 90% accuracy when comparing expected outcomes to actual results across 18 specific intents. These included tasks like updating account information, checking order status, refunds, unauthorized charges, general customer service requests, and resolving issues with orders.
To optimize efficiency, CloudCookies committed to an agile approach, targeting a rapid turnaround time for each phase. Their objective was to complete every phase within just two weeks, ensuring a swift deployment of improvements and solutions to enhance DoorDash’s customer support operations.
Phase 2 of CloudCookies’ solution for DoorDash, the focus shifted towards conversational design, particularly in identifying and enhancing personas for customers, Dashers, and merchants. This phase involved several key components, including theme and pattern detection using existing transcription data, word cloud generation to analyze contact requests, triaging self-service opportunities, and creating a self-service benefit heatmap from transcription data. CloudCookies also aimed to discover potential API integration points for comprehensive self-service solutions. They developed conversational design storyboards, tailored interactions, and an architecture that leveraged existing transcription and interaction data. Furthermore, they trained and fine-tuned natural language bots based on DoorDash’s existing transcriptions and interaction data. The solution also encompassed touch-tone DTMF fallback experience storyboards for each use case and robust error handling practices to enhance overall customer support capabilities.
Lastly, the primary focus in Phase 3 was on enhancing self-service capabilities for seven distinct use cases. This was accomplished through a comprehensive conversational design process that included voice input, natural language understanding, and yes/no confirmations. The solution incorporated architectural diagrams and involved the integration of AWS Lambda functions with existing APIs to facilitate self-service interactions. Performance tuning was a key aspect, with a focus on key performance indicators such as containment, intent accuracy, and word accuracy/error rates. Dynamic prompts were introduced to suggest self-service options based on user data and input.