Model Builder
Model Builder is Salesforce's tool for managing the large language models that power Einstein generative AI features.
Definition
Model Builder is Salesforce's tool for managing the large language models that power Einstein generative AI features. It is the layer where admins choose which LLMs Prompt Builder and Agentforce use, configure connections to third-party model providers, monitor model usage, and bring custom models into the platform. Model Builder exists because the right LLM for a given use case varies: Salesforce's own models work for many cases, OpenAI or Anthropic models work for others, and enterprise customers sometimes need to bring their own fine-tuned models for specific domains.
The product supports several model sources. Salesforce-hosted models include xGen (the Salesforce-trained foundation models) and other native options. Trust Layer connections expose third-party LLMs (OpenAI GPT-4, Anthropic Claude, Google Gemini, AWS Bedrock-hosted models) through zero-data-retention agreements that protect Salesforce data from being used for model training. Custom Model Connections let enterprises register their own hosted LLMs as endpoints Salesforce can call. Each model has performance, cost, and quality trade-offs that admins can compare and choose based on the use case. Model Builder is the configuration center for these decisions.
How Model Builder governs LLM use across Salesforce
The three model sources: Salesforce, Trust Layer, custom
Salesforce-hosted models are the default. They include xGen (Salesforce''s own foundation models) plus other native options. They are tuned for Salesforce-specific use cases and licensing is included with Einstein subscriptions. Trust Layer models are third-party LLMs (OpenAI, Anthropic, Google, AWS Bedrock) accessed through Salesforce''s zero-data-retention agreements; these models offer broader capability but at different cost and latency. Custom Model Connections let customers register their own hosted LLMs, useful for industries with regulatory or model-specialization requirements.
The Einstein Trust Layer and why it matters
The Trust Layer is the differentiator between using ChatGPT directly and using OpenAI''s models through Salesforce. Trust Layer features: zero-data-retention contracts with model providers (your data is not used to train their models), data masking (PII is anonymized before LLM submission), prompt and response logging (audit trail of every AI interaction), toxicity filtering (harmful responses are blocked), and prompt injection defense. Without the Trust Layer, sending Salesforce data to a generic LLM endpoint would violate most enterprise data agreements.
Bring-your-own model (BYO LLM) connections
Some customers need specific models: regulated industries that require on-premises hosting, organizations with proprietary fine-tuned models, customers using region-locked models for data sovereignty. Model Builder supports BYO LLM connections that point at customer-managed endpoints. The customer hosts the model, Salesforce calls it through the connection, and the same Trust Layer features apply. Setup is more involved than Salesforce-hosted or Trust Layer models but unlocks use cases that off-the-shelf models cannot serve.
Model assignment per prompt template
Each Prompt Template specifies which model to use. Different templates can use different models based on the use case. A sales email template might use a Salesforce-hosted model for cost efficiency; a complex multi-turn agent conversation might use Anthropic Claude for better reasoning; a regulated-industry summary might use a customer-hosted compliant model. Model Builder exposes the model picker that each template uses; switching models is a configuration change, not a code change.
Cost, latency, and quality trade-offs
Different models have different cost and quality profiles. GPT-4 and Claude Opus are higher quality but more expensive per token. Salesforce-hosted models are cheaper but less capable on complex reasoning. Smaller models (GPT-4o-mini, Claude Haiku) trade some quality for significantly lower cost. Latency also varies: bigger models take longer to respond. Model Builder''s monitoring shows the trade-off metrics so admins can pick the right model per use case rather than defaulting to the most powerful (and most expensive) option.
Usage monitoring and credit tracking
Model Builder includes usage dashboards. Per-model token counts, response times, error rates, and credit consumption. Setup > Einstein Generative AI > Usage shows aggregate spend. For consumption-priced models (most third-party LLMs), the credit cost can become significant at scale; monitoring is essential. Build alerts for credit spend approaching the entitlement, particularly for templates that fire frequently in customer-facing surfaces.
Model Builder and the Einstein Studio relationship
Model Builder is part of Einstein Studio, which also includes Prompt Builder, Agent Builder, and Copilot Builder. The Studio is the umbrella for designing generative AI experiences; Model Builder is the foundation layer that the other tools build on. Templates and agents reference models managed in Model Builder; changes to model configuration propagate to consumers. The architecture supports separation of concerns: prompt designers focus on the prompts, model admins focus on the LLMs.
How to configure models in Model Builder
Configuring Model Builder is about three decisions: which Salesforce-hosted models to enable, which Trust Layer third-party models to connect, and whether to set up any bring-your-own LLM connections. Most orgs start with defaults; customers with specific cost, quality, or compliance needs configure more deliberately. Plan the model strategy with your account executive because licensing varies by model source.
- Identify model needs based on use cases
List the use cases the org plans to support. Sales email drafting, service response generation, agent conversations, document summarization. Each may need a different model based on cost, quality, and latency requirements.
- Enable Salesforce-hosted models
Setup > Einstein Setup > Generative AI > enable. Pick which Salesforce-hosted models to make available. xGen is the default; other Salesforce-hosted models may be available based on edition and add-ons.
- Connect Trust Layer third-party models
Setup > Model Builder > Add Connection. Pick the provider (OpenAI, Anthropic, Google, AWS Bedrock). Provide API credentials and configure the model endpoint. Salesforce applies the Trust Layer protections automatically.
- Configure bring-your-own LLM if needed
For custom hosted models, set up the BYO LLM connection. Provide the endpoint URL, authentication, and request/response format. Validate the connection by running a test prompt through the model.
- Assign models to prompt templates
Open each Prompt Template in Prompt Builder. Pick the model that fits the use case. Document the rationale; future admins need to know why each template uses its specific model.
- Monitor usage and quality
Setup > Einstein Generative AI > Usage. Review per-model token counts, response times, error rates, credit consumption. Look for templates whose model choice produces poor quality or excessive cost; adjust accordingly.
- Set credit consumption alerts
Configure alerts for credit spend approaching the entitlement. High-volume templates with expensive models can produce significant credit consumption; alerts let you intervene before overage charges.
- Review and rotate models periodically
LLM capabilities improve rapidly. Newer model versions often outperform older ones at lower cost. Schedule quarterly reviews of model choices and rotate where the trade-offs have shifted.
Salesforce-hosted, Trust Layer third-party, or bring-your-own LLM. Different cost, quality, and compliance profiles.
Data masking, retention policy, and audit settings that apply to all LLM interactions through the platform.
Each Prompt Template specifies which model it uses. Allows different templates to use different models based on use case.
- Third-party LLM credits consume per token. High-volume templates with expensive models like GPT-4 produce significant cost. Monitor consumption closely.
- Model quality varies significantly. Defaulting every template to the most powerful model wastes credits; matching model to use case optimizes cost without sacrificing quality.
- BYO LLM connections need careful authentication and rate-limiting configuration. Misconfigured endpoints can leak data or fail unpredictably under load.
- LLM capabilities evolve rapidly. The optimal model today may be outclassed by a newer release in six months. Schedule periodic reviews to stay current.
- The Einstein Trust Layer matters for compliance. Confirm your model connections route through the Trust Layer; direct API calls bypass the protections.
Trust & references
Cross-checked against the following references.
- Model Builder OverviewSalesforce Help
- Einstein Trust LayerSalesforce Help
Straight from the source - Salesforce's reference material on Model Builder.
- Model Builder DocumentationSalesforce Help
- Bring Your Own LLMSalesforce Help
- Einstein Generative AI SettingsSalesforce Help
About the Author
Dipojjal Chakrabarti is a B2C Solution Architect with 29 Salesforce certifications and over 13 years in the Salesforce ecosystem. He runs salesforcedictionary.com to help admins, developers, architects, and cert/interview candidates sharpen their fundamentals. More about Dipojjal.
Test your knowledge
Q1. What is Model Builder?
Q2. Who can use Model Builder?
Q3. What should you do with model predictions?
Discussion
Loading discussion…