What the Trust Layer actually does
Every generative call from a Salesforce feature routes through a proxy that performs, in order:
1. Secure Retrieval — merge fields + Data Cloud grounding
2. Data Masking — PII replaced with tokens (name → [NAME_1])
3. Prompt Defense — jailbreak / injection detection
4. Zero-Data-Retention — provider contractually prohibited from training on prompts
5. LLM call — OpenAI, Anthropic, Bedrock, private endpoint
6. Toxicity Detection — response scanned
7. De-Masking — tokens restored for the end user
8. Audit Trail — full transcript stored in Data Cloud (`GenAiInteraction`)
Nothing in this chain is optional; it's how Salesforce ships "enterprise-safe" AI.
Prompt Template anatomy
Prompt Templates are metadata. Types:
- Sales Email — bound to Contact + Recipient
- Field Generation — populates one field on save
- Record Summary — surfaces on a record page
- Flex — you define inputs
Merge with Handlebars-style syntax on Salesforce/Data Cloud objects:
You are a rep at {!$User.Company.Name}.
Draft a renewal outreach for {!Opportunity.Account.Name}.
Context:
- Product: {!Opportunity.Product__c}
- Term ends: {!Opportunity.CloseDate}
- Last 30d support cases: {!Related.Account.RecentCases}
Rules:
- Under 150 words.
- Reference at least one recent case.
- Sign off as {!$User.FirstName}.
Grounding with Data Cloud vector search
Prompt Template
{!Retriever:KnowledgeArticles(query=input.question, k=5)}
At runtime, the retriever runs a vector search on the specified DMO/CIO and injects the top-K chunks with citations.
Bring Your Own Model (BYOM)
Model Builder connects an existing endpoint:
{
"provider": "AzureOpenAI",
"endpoint": "https://my-openai.openai.azure.com/",
"deployment": "gpt-4o-mini",
"auth": { "type": "namedCredential", "name": "AzureOpenAI_NC" },
"context_window": 128000
}
Trust Layer wraps it automatically — masking, defense, audit all apply.
Testing prompts before shipping
Use the Prompt Builder → Preview with test rows from Data Cloud. For CI:
curl -X POST "https://.../services/data/v62.0/einstein/prompt-templates/Draft_Followup_Email/generations" \
-H "Authorization: Bearer $TOKEN" \
-d '{"inputs":[{"variables":{"Opportunity":"006xx..."}}]}'
Assert on presence/absence of PII, length, forbidden terms — not exact wording.
Cost & throughput
- Every generation costs "Einstein Requests" — meter per feature
- Cache repeated summaries via Data Cloud (
AccountAiSummary__c) - Batch offline summaries in Automation Studio / Jobs; keep interactive latency low
Governance checklist
- All prompts in Git as metadata
- Every template reviewed by security before publish
- Sensitive fields tagged for masking policy
- Audit dashboard: which reps used which templates, acceptance rate
- Kill-switch: disable a template in production without a deploy
Prompt BuilderEinsteinAI Trust