Bring Your Own LLM to Salesforce: Do You Need Data Cloud?

Yes, but not through Salesforce's own bring-your-own-LLM (BYOLLM) feature. Salesforce sets up BYOLLM models in AI Models, formerly Einstein Studio, and that needs Data 360™, formerly Data Cloud, provisioned and enabled in the org.

Three routes put a model you contract for to work on Salesforce data:

1. Salesforce BYOLLM. Add the model in AI Models and use it in Prompt Builder, Apex, the Models API and Agentforce™ custom actions. Needs Data 360 provisioned and enabled.
2. Apex callout through a Named Credential. Your developers call the provider's API directly. No Data 360, and no Salesforce generative AI features: prompt assembly, masking and logging are yours to build.
3. An AppExchange app. A packaged app such as GPTfy makes the Named Credential callout for you, with grounding, masking and audit records included. No Data 360.

The rest of this page sets the three side by side. If your question is whether Agentforce itself needs Data Cloud, that is answered on Does Agentforce require Data Cloud?

BYOLLM is Salesforce's feature for adding a model you already pay for to its generative AI stack. Salesforce's documentation lists four providers you can connect with your own credentials:

• Amazon Bedrock (Anthropic models)
• Azure OpenAI
• OpenAI
• Google Vertex AI

For any other endpoint, including a model you host yourself, Salesforce offers the LLM Open Connector. The provider runs the inference, but the request still goes through Salesforce's Models API, so Einstein Trust Layer features apply to it. BYOLLM is available in Enterprise, Performance and Unlimited Editions with an Einstein for Sales, Service or Platform add-on.

Which managed models Agentforce uses by default, and how to override them per agent, is on Which LLM does Agentforce use?

Because BYOLLM sits on top of Einstein generative AI, and Einstein generative AI sits on Data 360. Salesforce Help says that to add a foundation model, "Einstein Generative AI must be enabled in your org", and that before turning Einstein generative AI on you should "make sure Data Cloud is provisioned and enabled", because Data Cloud is required for the Trust Layer. The AI Models setup page itself sits under Data 360 in Salesforce Help.

That is a tenant and a switch, not a Data Cloud programme. You do not need to ingest data, build a unified profile or set up a Data Library to use a BYOLLM model in a prompt template. The difference between provisioned and implemented is set out, with Salesforce sources, on Does Agentforce require Data Cloud?

For many orgs the tenant is not the obstacle. Salesforce lets Enterprise Edition and above provision Data 360 at no cost, though using it consumes credits. The obstacles are the Einstein add-on, the Flex Credits consumption that follows, and the team's view on adding another Salesforce data store. Those are the reasons buyers ask this question.

A fourth path, your own MuleSoft™ integration built on the MuleSoft Inference Connector, works like the Apex route if your integrations already run through MuleSoft. Salesforce's MuleSoft recipe for the LLM Open Connector is different: it ends in AI Models, so it still needs Data 360. If you want people to use ChatGPT or Claude apps against Salesforce rather than bring a model into Salesforce, see Use Claude or ChatGPT in Salesforce without Agentforce.

Two bills, sometimes three. Salesforce meters BYOLLM calls as Starter Prompts, counted in 2,000-token chunks. Its Flex Credits rate card updated 31 August 2026 charges 2 credits per chunk, the same as Basic prompts on Salesforce-managed models (Standard is 4 and Advanced 16). Agent actions are 20 credits each, and Salesforce lists Flex Credits at $500 per 100,000. Your provider then bills its inference under your own contract. If you store Trust Layer audit data in Data 360, that collection consumes Data 360 credits too.

What a Flex Credit costs, and how the two Agentforce pricing models compare, is on Agentforce pricing: Flex Credits vs Conversations. Per-user alternatives are on Agentforce alternative with fixed per-user pricing.

Your provider, under your contract. With Salesforce-managed models, Salesforce relies on its own zero-retention agreements with model partners. Its Agentforce privacy FAQ says those answers do not apply to BYO LLM: the provider is a non-Salesforce application and you must comply with your own contract with it. So training, retention, processing region and any BAA are set by your agreement with OpenAI, Microsoft, Google or AWS. The same is true of the Apex route and of GPTfy.

Masking is the other half. Einstein Trust Layer masking can apply to prompt templates, but Salesforce documents that it is disabled for Agentforce agents, so agent traffic reaches the model as retrieved. The provider-by-provider training and retention terms are on Is it safe to send Salesforce data to an LLM? and the agent masking detail is on Does Einstein Trust Layer mask agent data?

Salesforce BYOLLM fits if Data 360 is already enabled, you are building on Prompt Builder and Agentforce, and you want the model inside Salesforce's own Trust Layer and billing.

Apex and a Named Credential fit if you need one or two narrow calls, have developers to own them, and are prepared to build masking and logging yourselves.

An AppExchange app fits if you want your own model across many use cases without turning on Data 360, with masking and audit records included.

When not to choose GPTfy: you have fewer than 100 users who need AI (the plan minimum), you are standardising on Agentforce and Data 360 anyway, or you only need a single callout your developers can maintain.

GPTfy is a Salesforce-native app on AppExchange that runs inside your org and calls the model you choose through admin-controlled Named Credentials. It grounds prompts on your existing standard and custom objects with Data Context Mapping, so it needs no Data 360 or Agentforce licence.

Before each call it masks data in four layers: field-level rules, regular expressions, blocklists, and Apex for cases the first three cannot express. Raw data stays in Salesforce and only masked data reaches your AI provider. Each call writes a Security Audit Record. With bring your own model, the provider contract is your own, so you set retention terms with your provider directly.

Pricing is $20, $30 or $50 per user per month with a 100-user minimum, and model usage is billed by your provider. Apex invocation and AI agents start in the Enterprise plan.

More on bringing your own model, data masking and pricing.

Book a demo

Every statement about Salesforce on this page comes from one of these. Salesforce updates its documentation each release, so re-check it rather than relying on the date on this page.

• Salesforce Help: Bring Your Own Large Language Model
• Salesforce Help: Add a foundation model
• Salesforce Help: Set up Einstein generative AI
• Salesforce Developers blog: Bring Your Own Large Language Model in Einstein 1 Studio
• Salesforce Help: Select Agentforce model option
• Salesforce Help: Flex Credits usage types (Starter Prompts)
• Salesforce: Flex Credits rate card, updated 31 August 2026
• Salesforce: Agentforce pricing
• Salesforce: Agentforce privacy FAQ (June 2025)
• Salesforce Help: Free Data 360 account (knowledge article)
• Salesforce Help: Data masking for BYO LLM models
• Salesforce Help: Data masking limitations in Agentforce
• Agentforce Developer Guide: Models API
• Apex Developer Guide: Named Credentials as callout endpoints
• MuleSoft Docs: MuleSoft Inference Connector
• AppExchange: GPTfy listing

This comparison is based on publicly available documentation as of 9 November 2026. Features and pricing subject to change. Salesforce, Agentforce, Einstein, Data 360, MuleSoft, and related marks are trademarks of Salesforce, Inc. GPTfy is an independent product available on AppExchange and is not affiliated with or endorsed by Salesforce, Inc. beyond marketplace partner status. Microsoft and Azure are trademarks of Microsoft Corporation. OpenAI, Anthropic, Claude, Google, Gemini, Vertex AI, Amazon Web Services and Amazon Bedrock are trademarks of their respective owners. Last reviewed 9 November 2026.