How we use AI for Conversions

Overview

EngineRoom's Phone Call Intelligence feature is designed to extract valuable insights from phone call transcripts for your business. This documentation outlines the technical process behind the transcription and analysis pipeline, detailing each stage of data handling, from audio capture to actionable insights.

Transcribing Audio to Text

Transcribing audio files into text is a complex task due to several challenges:

• Overlapping Conversations

• Unclear Pronunciation

• Background Noise

• Hold Messages (Interactive Voice Response)

We are currently utilising a third-party platform, Assembly AI, an industry leader in speech transcription, to handle the audio-to-text transcription.

Transcript to Insights

We leverage a variety of Large Language Models (LLMs) from leading institutions such as OpenAI, along with our own proprietary models, to analyse the cleaned transcripts. The steps involved are:

• Transcript Refinement:

  • We implement specialised algorithms and natural language processing (NLP) techniques to identify and exclude IVR messages and irrelevant text to differentiate between multiple speakers and focus on the conversation between the customer and the sales representative

• Custom Prompt Creation:

  • A critical part of our process involves generating a curated prompt that includes your business's specific goals and key focus areas.

  • This prompt ensures that the AI model is aligned with your business context, enabling more accurate and relevant insights. It includes:

    • Role Specification, defining AI's role as a summarisation and categorisation assistant

    • Company Information, to provide context on company's brand and goals

    • Pre-defined output format and categories, allowing the output to be saved in a database and displayed in the platform

• Model Inference: The prompt is then fed into an AI model to extract the following insights:

  • Key Topics Asked About: Gain a clear understanding of what your customers are inquiring about most, enabling you to create better content and address any gaps

  • Issues Mentioned: Proactively identify common issues your customers face, allowing you to address them and track over time

  • Category: Calls will now be classified into general types to help your team quickly understand the nature of each lead:

    • Existing Customer Enquiry

    • Genuine Lead

    • Issue Needs Fixing

    • Callback Requested

  • Segment: Personalisation is key. We will segment calls based on the specific categories relevant to your business.

  • Sentiment: Assess customer sentiment during calls to better understand their experiences

Example

    "Key Topics Asked About": ["Why was the free installation denied?", "What steps can be taken to resolve this issue?"],
    "Issues Mentioned": ["Incorrect installation promotion information, Difficulty booking installation"],
    "Category": "Issue Needs Fixing",
    "Segment": "Filtered Water Taps",
    "Sentiment": "Neutral"

AI Limitations

It is essential to acknowledge that AI isn't perfect - there may be instances where the AI misinterprets information or overlooks certain nuances in conversations. Context or subtleties may not always be fully captured, but we are continuously working to improve our models and welcome your feedback to enhance accuracy.