Definition
Utterance is part of Salesforce's AI capabilities that bring intelligent automation and insights into CRM workflows. It applies advanced algorithms to organizational data to generate predictions, recommendations, or autonomous actions.
Real-World Example
a solutions architect at DeepSight Analytics recently implemented Utterance to enhance decision-making with AI-driven insights embedded directly in the CRM workflow. Utterance processes thousands of records and delivers actionable recommendations that help the team prioritize their efforts and improve outcomes measurably.
Why Utterance Matters
An Utterance in the context of Salesforce AI refers to any text or spoken input that a user provides to a conversational AI system, such as an Einstein Bot or a natural language processing model. Utterances are the raw material that AI uses to determine user intent - the system analyzes the utterance, matches it against trained intents, extracts relevant entities (like dates, product names, or account numbers), and routes the conversation accordingly. Defining a diverse and representative set of training utterances for each intent is what makes the difference between a bot that understands users and one that constantly responds with fallback messages.
As conversational AI deployments scale from simple FAQ bots to complex multi-turn assistants handling thousands of interactions daily, utterance management becomes a discipline of its own. Poor utterance training leads to misclassified intents, frustrated customers, and high escalation rates to human agents, defeating the purpose of AI automation. Organizations need to continuously mine conversation logs for new utterance patterns, handle regional language variations and slang, and maintain clear boundaries between similar intents to prevent overlap. Teams that invest in regular utterance refinement see measurable improvements in bot containment rates and customer satisfaction scores over time.
How Organizations Use Utterance
- SwiftBank Digital — SwiftBank trained their Einstein Bot with 150 utterance variations for the 'check balance' intent, including phrases like 'how much money do I have,' 'show my balance,' and 'what are my funds.' This diversity improved intent recognition accuracy from 72% to 94%, reducing the number of customers being incorrectly routed to a live agent for simple balance inquiries.
- TravelEase Vacations — TravelEase discovered through conversation log analysis that customers were using 30 different ways to ask about trip cancellation policies. They added these naturally occurring utterances to their bot's training data, which reduced the cancellation policy intent's fallback rate from 35% to 8% and decreased average handle time for cancellation inquiries by 40%.
- MedConnect Health — MedConnect's bot team identified that patients frequently used medical abbreviations and informal language when describing symptoms. They expanded their utterance training to include variations like 'BP check,' 'blood pressure reading,' and 'my pressure numbers,' ensuring the bot correctly routed these to the vitals inquiry intent instead of the generic help fallback.