A Deep Dive into Salesforce's Einstein Trust Layer

Balancing generative AI benefits with data security is a challenge. The Einstein Trust Layer, built into Salesforce, ensures safety with robust agreements, security measures, and privacy controls, empowering secure AI adoption.

1. Secure Data Retrieval: The prompt journey begins with secure data retrieval from the org, adhering to Salesforce permissions that govern access to objects, fields, and other data.

2. Dynamic Grounding: To generate personalized responses, the LLM needs additional context from your CRM data, a process called grounding. Grounding enriches prompts with CRM data using merge fields, such as record fields, flows, Apex, Data Cloud DMOs, and related lists. This ensures responses are more relevant and tailored to your needs.

3. Data Masking: The prompt may contain sensitive data like customer names and addresses, which are not ready to send to the customer. The Trust Layer adds protection through data masking, tokenizing each value and replacing it with a placeholder. This enables the LLM to maintain context and generate relevant responses while ensuring the privacy and security of customer data.

4. LLM Response Generation: With the prompt populated with relevant data, and protective measures in place, it’s ready to leave the Salesforce Trust Boundary by passing through the Secure LLM Gateway to connected LLMs. External LLM uses this prompt to generate a relevant, high-quality response for Jessica to use in her conversation with her customer.

5. Zero retention policy: Salesforce's zero data retention policy ensures no customer data, prompts, or responses are stored outside Salesforce, ensuring privacy.

6. Toxicity Detection: The response passes back into the Salesforce Trust Boundary from the LLM, where it's checked for toxicity. The Salesforce assessment tool, using deep learning models, scans for harmful content. Toxic language detection protects Jessica and her customers from toxic, hateful, violent, sexual, identifiable, physical, and profane responses. If the score is high, the reply is discarded.

7. Data Demasking: The placeholders we created for masking the data during the prompt journey are now replaced with the actual data. The relationship between the original entities and their respective placeholders is used to rehydrate the response so the response is useful and meaningful when sent back.

8. Audit Trail: All interactions, including the prompt, unfiltered response, toxic language scores, and feedback, are timestamped and collected into an audit trail. The Einstein Trust Layer ensures accountability, providing a clear record of data handling to guarantee customer data protection. These audit trail records are available for 30 days in Salesforce.

9. Feedback Framework: The response is reviewed for quality and alignment with the case-handling style. The user can accept, edit, or ignore the response before sending it. They can also provide qualitative feedback with a thumbs up or down, and if the response wasn’t helpful, specify a reason. This feedback is securely collected to improve future prompt quality.

The entire process happens in a blink of an eye. Within seconds, the conversation moves from a prompt initiated by a chat, through the security of the Trust Layer, to a relevant, professional response ready to be shared with the customer.