Zero Data Retention Verification: an Agentforce Architecture Guide

What is Zero Data Retention Verification in Agentforce architecture?

In Agentforce architecture, verify that every LLM endpoint beyond your org boundary retains nothing, with contractual or configuration evidence on file. This guide explains the design decisions, controls, and implementation checks needed to apply the pattern in production.

Agentforce prompts and completions leave your Salesforce org and travel to external LLM endpoints. The Einstein Trust Layer enforces zero data retention on these calls, meaning the provider does not store the prompt or completion and does not use them for training. This is what keeps customer data from persisting outside your control.

Policy claims are not enough for a production agent. Architects need to confirm zero retention for every endpoint in the request path, including traffic routed through external gateways, and keep evidence that stands up in a security review or audit.

Treat every endpoint as a retention risk

• Prompts and completions must not be retained by the model provider or used for training once they leave the org boundary.
• The check covers every endpoint in the path, not just the primary model: fallbacks, gateways, and regional endpoints all count.
• Zero retention must be provable through contract terms or provider configuration, not assumed from a marketing page.
• Any new model, region, or integration changes the endpoint list and invalidates stale evidence.

Build and maintain the evidence file

1. Inventory every LLM endpoint the agent can reach, including those behind external gateways or custom integrations.
2. Collect zero-retention evidence per endpoint: data processing agreements, provider configuration exports, or documented Trust Layer settings.
3. Confirm in the provider console that logging and training opt-outs are active, and capture dated proof of the configuration.
4. Re-verify endpoints and refresh evidence on a schedule and whenever models, gateways, or regions change.