Agentforce Voice

Agentforce Voice enables Agentforce Service agents to speak and understand voice conversations with customers. It is designed to understand customer intent from spoken input and take actions to resolve customer queries.

Required Editions

Available in: Lightning Experience

Available in: Enterprise, Unlimited, and Developer Editions with Foundations or Agentforce 1 Editions, and Salesforce Voice add-ons.

Note Agentforce Voice is supported only in English.

Agentforce Voice uses multiple models to handle voice interactions.

The user’s spoken words are first converted to text using a speech-to-text model, such as Deepgram.
The transcribed text is then processed by a Flash Planner, which functions like the Salesforce Atlas reasoning engine. Using an OpenAI GPT large language model (LLM), the most relevant subagent and actions are identified to use to respond to the user’s query, resulting in a text response.
Finally, the generated text is converted back to audio using a text-to-speech model and delivered to the user using a text-to-speech model, such as ElevenLabs.

We have built-in instructions and guardrails to help make sure agent responses are clear, easily understood, and follow content safety and security policies. These instructions are integrated into the Flash Planner to manage interactions.

These are a few examples of what the built-in instructions and guardrails accomplish.

Optimize spoken delivery: Instructions on how agents format spoken replies, such as strategic pauses for clarity when conveying information like addresses. For example, separating numbers from street names.
Provide concise responses: Guiding agents by word limits to prevent lengthy responses.
Help ensure content safety: Guardrails to help prevent toxic or offensive language in agent responses and safeguard proprietary information.

Zero Data Retention Policies

Your data isn’t retained by third-party LLMs. We partner with all our external model providers, such as OpenAI, to enforce the zero-data retention policy.

Large Language Models: We partner with all our external large language model (LLM) providers to enforce the zero-data retention policy. For more details on other trust related issues, see Trust and Agentforce.
Deepgram: We use a dedicated opt-out mechanism that Deepgram provides to make sure data is only retained for the duration necessary to process the request and not utilized for model improvement.
ElevenLabs: We use ElevenLabs zero retention mode, an enhanced privacy feature to make sure the data from our requests and responses is immediately deleted upon completing the request. The specialized mode restricts the logging of sensitive data points, ensuring they are never sent to a database for long-term storage. ElevenLabs also has agreements with third-party LLM providers that prevent them from training their models on customer data.