AI Engine sends requests from your Salesforce org straight to your AI provider, through the Named Credential you create. These calls do not go through the Salesforce Einstein Trust Layer, so Salesforce's own agreements with LLM providers do not cover them: the terms of your provider account do. AI Engine requires a provider account with zero data retention (ZDR) enabled. With ZDR the provider does not keep your prompts or answers after it responds, does not use them for training, and nobody at the provider reviews them.
Which calls go to the provider#
- Every AI feature (chat, actions, Meet Prep, AI Scoring, Magic Snap, PDF templates) calls the provider selected in the Org Profile: the OpenAI API, the Anthropic API or your own Azure OpenAI resource.
- Web-search features call the same provider. On Azure OpenAI they call the OpenAI API through a Named Credential named
OpenAI, because Azure OpenAI has no web search tool; see Choosing a provider for compliance. - Cloud Connected runs no server or proxy in this path and receives none of the content.
How to enable zero data retention#
| Provider | Training on API data | Without ZDR | How to enable ZDR |
|---|---|---|---|
| OpenAI | Not by default | Abuse-monitoring logs kept for up to 30 days | Request Zero Data Retention or Modified Abuse Monitoring from OpenAI; after approval, set it for the organization or project whose API key AI Engine uses |
| Anthropic | Not without your express permission | Retention under Anthropic's commercial data retention policy | Contact Anthropic sales; ZDR is enabled per organization, so use an API key from that organization |
| Azure OpenAI | No | Prompts and answers can be stored for abuse monitoring and reviewed by people | Apply to Microsoft for modified abuse monitoring for the subscription that holds your Azure OpenAI resource |
Provider terms change. Check the current terms and keep the provider's written confirmation of ZDR for your audits.
What AI Engine does on its side#
- No stored responses at OpenAI. Every OpenAI Responses request sends
store: false. - No stateful provider features. AI Engine uses no file storage, batch jobs, assistants or threads, or fine-tuning. Each request carries the context it needs and refers to no earlier one.
- Personal data is masked by default. Names, e-mail addresses and phone numbers are replaced with tokens before a request is sent, unless an administrator allows personal data; see PII masking.
- No prompts or answers stored in Salesforce. See How AI Engine handles your data.
Exceptions the providers keep#
Even with ZDR, providers keep the right to retain data that the law requires them to keep, and content their automated safety systems flag as misuse. Anthropic, for example, may keep flagged inputs and outputs for up to two years. Some provider features fall outside ZDR; AI Engine does not use them.
Checklist before go-live#
- Choose the provider; see Choosing a provider for compliance.
- Get ZDR (or modified abuse monitoring) enabled on that provider account.
- Create the Named Credential with an API key from that account; see API keys and Named Credentials.
- On Azure OpenAI: if you create the
OpenAINamed Credential for web search, enable ZDR on that OpenAI account too, or leave the credential out. - File the provider's ZDR confirmation with your compliance records.