Employees continue to paste customer records, support transcripts, and opportunity details into consumer versions of ChatGPT and similar tools. The practice creates immediate exposure because those accounts sit outside any contract, logging, or retention policy your organization controls.
Salesforce addresses the risk inside its own generative features through a set of technical and contractual controls. Those controls do not reach the consumer tools employees open on their own devices. The distinction matters because the two problems require different solutions.
What Salesforce controls inside its own AI paths
The Einstein Trust Layer applies Zero-Data Retention agreements to every prompt sent to third-party large language models. Under those agreements the provider may not store prompts or responses after the reply returns to Salesforce, may not use the data for training, and may not allow human review except in narrow cases of suspected illegal activity. The policy covers generative features and Agentforce actions.
Data masking runs before the prompt leaves Salesforce infrastructure. Pattern-based detection uses regular expressions and context models to find PII and PCI values. Field-based detection reads metadata tags, Shield encryption status, and sensitivity levels you set on objects. Masked values travel as placeholders; the response is demasked on return. Masking works for embedded Einstein features such as Service Replies, yet it is disabled for full Agentforce agents because the added latency and accuracy loss affect planner and action workflows.
Dynamic grounding respects the running user’s existing field-level security and record access. No additional permissions are granted simply because an AI feature is involved. Audit records for prompts, responses, and trust signals flow into Data 360 (formerly Data Cloud) so you can review what was sent and when.
The self-service Customer Data Opt-Out toggle prevents Salesforce from using your org data to improve global predictive models. The toggle sits in Setup under “Opt Out of Customer Data Access.” It applies at the enterprise ID level in many cases and does not affect the Zero-Data Retention behavior of generative calls.
Hyperforce keeps data residency decisions inside your chosen geography. When you deploy in a supported region, both storage and processing stay inside that boundary for the platform services that support it. Some hosted models, such as certain Anthropic Claude instances via Amazon Bedrock, can also run inside the Salesforce trust boundary rather than crossing to an external provider.
The gap that remains for personal AI tools
None of the mechanisms above observe or block an employee who opens a browser tab to chat.openai.com and pastes a customer email thread. The Trust Layer operates only on prompts that travel through Salesforce-approved generative paths. Consumer accounts have no Zero-Data Retention contract with your organization, no masking layer, and no audit trail visible to you.
The technical controls Salesforce ships are therefore necessary but not sufficient on their own.
How to configure the controls that do exist
Begin by confirming the required editions and add-ons: Enterprise, Performance, or Unlimited plus the relevant Einstein licenses. Data 360 (formerly Data Cloud) must be enabled for full masking and audit functionality.
In Setup, locate Einstein Trust Layer settings. Enable pattern-based and field-based masking for the data types your compliance program requires. Tag sensitive fields in Object Manager with the appropriate compliance categories and sensitivity levels so field-based rules can act on them. Test the configuration in Prompt Builder by sending sample prompts that contain known PII and confirming the placeholders appear in the logged request.
Turn on the Customer Data Opt-Out toggle if your governance policy requires it. The change takes effect immediately and removes the org from future model improvement cycles.
For residency, confirm the Hyperforce region selected during your org’s deployment matches the regulatory requirements that apply to your customer data. Review the supported regions list periodically because Salesforce continues to add countries.
Finally, write and publish an acceptable-use policy that explicitly prohibits pasting customer data into any consumer AI account. Pair the policy with short training sessions that show employees the approved paths inside Salesforce and the single-click opt-out they can request when a new tool appears.
Where these controls are the wrong choice
If your organization must prevent every possible exfiltration vector, including browser-based consumer tools, the Salesforce-native stack alone will not achieve that outcome. You will need additional endpoint controls, browser restrictions, or enterprise versions of external models that carry their own contractual protections. The platform features described here were built to secure data that stays inside approved Salesforce flows, not to police every tab an employee opens.
Limitations worth stating plainly
Masking is not perfect. Cross-language or heavily obfuscated values can slip through. Latency-sensitive agent workflows lose the masking step entirely. Third-party AppExchange components and custom integrations fall outside the Trust Layer unless their developers have built equivalent controls. Audit logs appear only after the fact, so prevention still depends on the policy layer.
Frequently asked questions
Does the Einstein Trust Layer stop employees from using ChatGPT?
No. The Trust Layer only governs prompts that route through Salesforce generative features. Consumer accounts remain outside its reach.
What happens to prompts sent to third-party models under Zero-Data Retention?
The provider contract states that prompts and responses are not stored after the reply returns, are not used for training, and receive no human review except in documented cases of suspected illegal activity.
Can I still use Agentforce if I enable data masking?
Masking is currently disabled for full Agentforce agents because the added processing affects planner accuracy and response time. Embedded features such as Einstein Service Replies continue to support masking.
How do I know whether my org data is being used to train global models?
Check the self-service toggle in Setup. When the opt-out is active, Salesforce does not use your data for those purposes. The setting is visible without opening a support case.
What should replace consumer AI accounts inside the organization?
Approved paths include Prompt Builder, Agentforce actions, and any enterprise-licensed external models that carry equivalent contractual protections. The acceptable-use policy should name the sanctioned options explicitly.
Integrations & Data Cloud
Connect your systems. Unify your data. Power smarter AI.
Read Next

What controls do you need to keep Salesforce AI data sovereign?
Learn the essential controls for Salesforce AI data sovereignty, including BYO Key for models, Trust Layer, regional deployment, Shield encryption, and Data 360 governance.

Who Can Help Me Implement Salesforce Data Cloud? BYO Key Explained for Secure AI
Understand how Bring Your Own Key (BYOK) works with Data 360 (formerly Data Cloud) to give you control over encryption for AI workloads while meeting sovereignty requirements. (148 characters)

How Agentforce Routes Conversations Between Agents and Humans
Agentforce uses defined routing patterns to move tasks and context between AI agents or to people while keeping full conversation history intact.

