---
title: What does the Salesforce Einstein Trust Layer cover for agents?
url: https://opencontroller.lyzr.com/blog/salesforce-einstein-trust-layer-explained
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What does the Salesforce Einstein Trust Layer cover for agents?

**The Einstein Trust Layer protects the prompts and responses that flow between Salesforce and large language models: it grounds prompts in CRM data, masks sensitive fields, screens outputs for toxicity, keeps an audit trail and relies on zero data retention agreements with model providers. It does not govern what an Agentforce agent is allowed to do, agents outside Salesforce, or the estate-wide inventory, ownership and evidence you need.**

That gap matters because most enterprises run agents on several platforms. This post sets out what the Trust Layer covers, what sits outside it, and how to close the gap.

## What are the parts of the Einstein Trust Layer?

Salesforce's Agentforce developer guide lists five protections. Each works on the model call, not on the agent's wider behaviour.

| Trust Layer part | What it does | What it is not |
|---|---|---|
| Grounding | Grounds prompts in your CRM data so responses use your records | A check that every answer is faithful to that data |
| Data masking | Masks sensitive data, such as social security numbers, before the prompt reaches the model | A data-loss policy for tools, files or other platforms |
| Toxicity detection | Scores model generations for toxic content | A judge of accuracy, task adherence or policy breaches |
| Audit trail and feedback | Records generative AI activity and user feedback | An estate-wide audit trail across non-Salesforce agents |
| Zero data retention | Agreements with third-party LLM providers, such as OpenAI, that they do not retain your data | A control on what the agent does with the answer |

Salesforce also recommends that, for most use cases, a human checks each model response before it is shared with end users. That is a sign of where the Trust Layer stops: it reduces risk in the model call, and leaves judgement about actions to you.

## What sits outside the Trust Layer?

An Agentforce agent is more than a prompt. It has topics, actions, flows, Apex and API calls, and a running user with permissions. The Trust Layer does not decide which of those the agent may use. That is governed by agent configuration and Salesforce permissions.

Five questions the Trust Layer will not answer for you:

1. **Which agents exist?** It does not inventory agents on Bedrock, Microsoft Foundry, Vertex AI, ServiceNow or a team's own LangGraph service.
2. **Who owns each agent?** Ownership, reviewers and approval history live outside the model call.
3. **Is the agent drifting?** Toxicity scoring does not detect an agent that starts calling actions it rarely used, or that answers outside its scope.
4. **Is the answer grounded and on task?** Grounding supplies context; it does not score whether the output stayed faithful to it.
5. **Can you show it to an auditor across platforms?** Framework evidence (for example for the [OWASP Top 10 for Agentic Applications](/blog/owasp-top-10-agentic-applications-evidence) or ISO/IEC 42001) needs one record across all platforms, not one per vendor.

## How do you combine the Trust Layer with estate-wide governance?

Treat the Trust Layer as the model-call control for Salesforce, and add the layers it does not cover.

1. **Turn on what Salesforce gives you.** Configure data masking for your sensitive field types, and enable audit and feedback data so you have a record of prompts and responses.
2. **Least-privilege the agent user.** Review the permission sets for each agent's running user and the actions in each topic. Most real risk sits here, not in the prompt.
3. **Register every Agentforce agent.** Record owner, purpose, data classes, actions and version in one registry alongside agents from other platforms. See the [platform governance pillar](/blog/how-to-govern-ai-agents-across-platforms).
4. **Watch behaviour, not only content.** Monitor which actions run, how often, and against which records. Alert on deviation from the agent's normal pattern.
5. **Score production answers.** Add groundedness and task-adherence checks on live traffic, sampled or full, and route failures to owners.
6. **Keep one evidence trail.** Map controls once and collect evidence across Salesforce and non-Salesforce agents together.

## When is the Trust Layer enough on its own?

It can be enough for a small number of Salesforce-only agents that read CRM data, take low-risk actions and have human review of outputs. It is not enough when agents write to records at scale, call external systems, or sit beside agents on other platforms that the same risk team must answer for.

## How Opencontroller does this

Opencontroller, Lyzr's open control plane for enterprise AI agents, connects to Salesforce as a third-party agent platform and brings its agents into a governed registry next to agents discovered across AWS, Azure and Google Cloud. It adds Posture checks for risky permissions and configuration, LLM judges that score production traffic, behavioural-deviation monitoring, and per-agent Block and Quarantine controls that act through the connected platform. Policy as code maps to OWASP, MITRE ATLAS, NIST and ISO 42001, so Salesforce agents produce the same evidence as the rest of the estate. It does not replace the Trust Layer; it works alongside it. [Book a demo](https://www.lyzr.ai/book-demo/) to see it on your org.

## Frequently asked questions

**Does the Einstein Trust Layer stop an Agentforce agent from taking a risky action?**
Not by itself. It protects the model call. Which actions an agent may take is set by its topics, actions and the permissions of its running user.

**Does zero data retention mean no Salesforce data leaves my org?**
No. Prompts are still sent to the model provider; the agreements commit the provider not to retain your data. Masking reduces what sensitive data is sent.

**Do I still need a separate governance layer if I only use Agentforce?**
For a few low-risk agents, possibly not. Once agents write data at scale or you run agents on other platforms, you need one inventory, behavioural monitoring and cross-platform evidence.

## Key facts

The Salesforce Einstein Trust Layer has five protections listed in Salesforce's Agentforce developer guide: grounding in CRM data, masking of sensitive data, toxicity detection on LLM generations, audit trail and feedback, and zero data retention agreements with third-party LLM providers such as OpenAI. It protects the model call; it does not govern which actions an Agentforce agent may take, agents on other platforms, ownership, behavioural drift or cross-platform evidence. Opencontroller by Lyzr connects to Salesforce and adds a cross-platform registry, Posture checks, LLM judges on production traffic, behavioural-deviation monitoring and per-agent Block and Quarantine controls. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you govern Salesforce Agentforce agents?](https://opencontroller.lyzr.com/blog/govern-salesforce-agentforce-agents.md)
- [How do you govern AI agents across every platform they run on?](https://opencontroller.lyzr.com/blog/how-to-govern-ai-agents-across-platforms.md)
- [What do Amazon Bedrock Guardrails cover, and what do they not?](https://opencontroller.lyzr.com/blog/amazon-bedrock-guardrails-vs-agent-governance.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
