Salesforce Trusted AI Principles

This document sets out what Salesforce Einstein promises about how it handles your data and AI outputs, and what it expects from you in return. Your data belongs to you, not Salesforce — prompts and AI responses are never stored or used to train underlying AI models, and sensitive information is masked before it reaches those models. In exchange, you are expected to use Salesforce Einstein responsibly and in accordance with its Acceptable Use Policy.

The Salesforce Trusted AI Principles establish Salesforce Einstein's substantive commitments regarding data ownership, AI transparency, content safety, and acceptable use. Customers retain exclusive ownership and control of their data at all times, with a zero data retention policy prohibiting storage or use of prompts and responses to train third-party large language models. Salesforce Einstein applies data masking at the prompt stage to de-identify sensitive PII and proprietary data before it reaches the underlying LLM, and operates real-time toxicity detection and automated content filtering for harmful outputs. The document places affirmative obligations on customers to use AI responsibly and in compliance with Salesforce's Acceptable Use Policy, while Salesforce Einstein commits to model explainability at the point of prediction, public model cards, and bias-mitigation tools for employees, customers, and partners.

As an individual user, your data and prompts remain your property and are never stored or used to train the underlying large language models. Sensitive personal information in prompts is automatically replaced with non-identifiable tokens before reaching the model, and any harmful content identified in AI outputs is filtered or blocked before you see it. Salesforce Einstein also provides an explanation of how each AI prediction or recommendation was reached at the time it is delivered, and publishes public model cards describing each model's design, intended and unintended uses, known ethical implications, and performance — resources you can consult to understand the AI systems you are using.