Section 2 · Prompt Engineering · 20% of exam
2.8 Managing and Preventing Specific Models from Being Accessed
Exam objective
Explain how administrators govern which models can be used by prompt templates and how to prevent access to disallowed models.
Model governance means controlling which language models are available for an organization's generative AI use cases. The exam guide includes managing and preventing specific models from being accessed as a Prompt Engineering objective. The decision is not simply about choosing the model with the most impressive output. Administrators should use approved models that meet the organization's security, compliance, quality, and cost requirements, and then make the chosen model available only through authorized Salesforce configuration.
Apply an approved model policy
First identify the business and data requirements. Consider whether the use case involves sensitive information, the type of response required, the expected latency and quality, and any organizational restrictions on providers or regions. Consult the current Salesforce model settings and provider support for the product feature. Availability may vary by release, edition, org configuration, and feature. Do not infer that every model supported by an underlying provider is automatically available in Prompt Builder or Agentforce. Salesforce documents two controls for embedded generative AI such as Prompt Builder. In Setup, an administrator can disable a model provider through Einstein Setup and Configure Model Providers; that blocks the provider's models across org features that use them. In AI Models, an administrator can hide a language-model configuration so it is no longer selectable in Prompt Builder. These controls have different scopes: disabling a provider can affect agents, templates, APIs, and other generative AI components, while hiding a configuration restricts selection in Prompt Builder. Use least privilege for administrators who change these settings and document which models are approved.
Validate the result
Before disabling a provider, identify every component that references its models and select an approved replacement where needed. Salesforce warns that turning off a provider can stop affected agents, prompt templates, APIs, or other AI components from running. After applying a restriction, test the affected templates and workflows. Check that approved templates still execute, prohibited configurations no longer appear in Prompt Builder, and integrations do not provide an unreviewed alternate path. Review permissions for administrators and users who can modify templates, model settings, flows, or connected applications. A control that hides a configuration in Prompt Builder does not necessarily disable that provider across other features; understand the scope of each setting.
If a template is moved from sandbox to production, verify model availability and policy in the target org. Settings, entitlements, or supported models may differ between environments. Deployment of a prompt template should not be assumed to carry every org-level model governance configuration with it.
Tie governance to use-case risk
A model decision should be reviewed when source data changes, the use case expands, Salesforce adds supported models, or the organization's policies change. Monitor quality, safety, and operational outcomes using available Salesforce capabilities. If a model is disallowed for a particular type of data, enforce that restriction through approved configuration and access controls instead of relying on prompt instructions such as "do not use model X." Prompt text guides generation; it is not a reliable substitute for access governance.
If a scenario asks how to prevent a particular model from being used, look for an administrative model-access control or approved configuration policy, not an instruction written inside the prompt.
Model names and setup options are release-sensitive. Verify the current org settings and Salesforce Help before configuring a production policy.
Extra concept: Governance covers both availability and execution path. Review templates, permissions, integrations, and environment differences so that a restriction is effective wherever the prompt can run. Official references: LLM support and access - https://help.salesforce.com/s/articleView?id=ai.generativ e_ai_large_language_model_support.htm&language=en_US&type=5 | Model provider controls - https:// help.salesforce.com/s/articleView?id=release-notes.rn_einstein_models_provider_access.htm&langua ge=en_US&release=256&type=5
Key takeaways
- Approve models according to security, compliance, quality, and operational requirements.
- Use administrative access controls to restrict model availability.
- Test template, deployment, and integration paths to confirm disallowed models cannot be used.
