The short version
- There are three ways to pair Claude with Salesforce: Claudeforce (Salesforce runs Claude inside its own trust boundary), bring-your-own model registered through Einstein Model Builder, and Claude called directly from your own AWS account.
- Amazon Bedrock keeps Claude inside your AWS security boundary: AWS IAM authentication, a choice of global or region-pinned endpoints for data residency, and no use of your prompts to train the base model.
- The Einstein Trust Layer wraps external models, so grounding, PII masking, zero data retention, and an audit trail apply even to a Claude model you host yourself.
- For regulated work, Bedrock offers a HIPAA-eligible and FedRAMP High path through AWS GovCloud, though you should confirm which specific Claude models carry that authorization today.
What 'Claude on Bedrock for Salesforce' Actually Means
Amazon Bedrock is AWS's managed service for calling foundation models, Anthropic's Claude among them, without running any model infrastructure yourself. Pairing it with Salesforce is one of the most common questions we get now, and the honest answer is that there is not one integration. There are three.
The three differ in who runs the model, whose AWS account it sits in, and how the data is governed. Getting the choice right up front saves a painful migration later, so this guide lays out all three, then gives you a way to pick.
- Claudeforce: Salesforce runs Claude for you, inside its own trust boundary, on Bedrock.
- Bring your own model: you host Claude on Bedrock in your AWS account and register it with Salesforce through Einstein Model Builder.
- Direct on Bedrock: your own application calls Claude on Bedrock and writes the result back to Salesforce.
Why Amazon Bedrock, Not the Direct Anthropic API
The first fork is not about Salesforce at all. It is whether you reach Claude through Bedrock or through Anthropic's own API. For an enterprise already on AWS, Bedrock usually wins, and the reasons are governance, not features.
On Bedrock, you authenticate with AWS IAM rather than an Anthropic API key, and every call is signed with your AWS credentials. The model runs inside the AWS security boundary, Anthropic staff have no access to that inference infrastructure, and requests and responses are logged to CloudWatch and CloudTrail like any other AWS service.
Data residency is a real control here. Bedrock offers a global endpoint that routes across regions at no premium, and a regional endpoint you can pin to a single AWS region when the data cannot leave a geography, for a small premium. Bedrock also does not use your prompts or outputs to train its base models, supports your own KMS encryption keys, and can be reached privately from your VPC over AWS PrivateLink.
undefined
Illustrative only. Model names on Bedrock change almost monthly, so we point clients at the live model access console rather than hard-coding a version. The request body itself is the same Messages format either way, which keeps application code portable across both front doors.
Way 1: Claudeforce, Where Salesforce Runs Claude for You
Claudeforce is the managed path, announced by Salesforce and Anthropic on August 26, 2026. Claude became a preferred foundation model for Agentforce, and the first large language model integrated fully inside the Salesforce trust boundary, deployed on Amazon Bedrock but operated by Salesforce.
In practice that means you do not touch AWS at all. Claude powers the Atlas Reasoning Engine, runs Agentforce Coworker and Agentforce Vibes by default, and is selectable in Agent Builder. Regulated early adopters named at launch included CrowdStrike and RBC Wealth Management, with model traffic kept inside the Salesforce boundary.
Choose Claudeforce when you want the fastest path, have no AWS estate to manage, and are happy for Salesforce to own the governance. We cover the wider announcement in AIforce and the practical rollout in Claudeforce, explained.
Way 2: Bring Your Own Claude on Bedrock, Through Einstein Model Builder
The second path keeps Claude in your AWS account and connects it to Salesforce through Einstein Studio Model Builder. Salesforce supports Amazon Bedrock as a built-in model provider, so you can register a Bedrock-hosted Claude model, then ground prompts on your own data in Prompt Builder, callable from Flow, Apex, and invocable actions.
The draw is control. The model runs in your AWS account, on your committed AWS spend, in the region you choose, under your IAM policies. Crucially, the Einstein Trust Layer still wraps the call: the request passes through Salesforce's grounding, masking, and zero-retention controls before it reaches Bedrock, and the response returns the same way.
At Dreamforce 2026, Salesforce widened this further, giving Agentforce access to the broader Bedrock model library. Choose this path when you already run on AWS, need the model in a specific account or region, or want your own procurement and billing behind the inference. The exact connection method has shifted over time, so treat any older step-by-step as a pattern and confirm the current authentication in your org.
Way 3: Claude on Bedrock in Your Own App, Writing Back to Salesforce
The third path skips Einstein entirely. Your own application, a service, a batch job, or a data pipeline, calls Claude on Bedrock directly, then writes the result into Salesforce. This fits work that does not belong inside an Agentforce conversation: bulk enrichment, document processing, back-office scoring.
The integration back into Salesforce uses the same patterns any AWS-to-Salesforce build would. Salesforce Event Relay streams Platform Events and Change Data Capture to Amazon EventBridge with no middleware, Amazon AppFlow moves data both ways, and large files can live in S3 wired into Salesforce rather than eating org storage.
Choose this when the Claude workload is a system process rather than a user-facing agent, and you want full control of the prompt, the batching, and the cost.
The Real Question: What Data Leaves Salesforce
For most buyers, especially in regulated industries, the deciding factor is not speed or cost. It is where the data goes. All three paths are built to answer that, at different layers.
Inside Salesforce, the Einstein Trust Layer enforces dynamic grounding, PII masking, toxicity detection, zero data retention by the model, and an audit trail, and it applies whether the model is Claudeforce or one you brought yourself. On the AWS side, Bedrock does not train on your data or share it with the model provider, encrypts it in transit and at rest, and can run entirely over private networking.
For the strictest workloads, Salesforce Private Connect builds a private link to your AWS account over AWS PrivateLink, so traffic between Salesforce and Bedrock or Redshift never crosses the public internet. This is the pattern we reach for in healthcare and financial services.
Regulated Industries: The HIPAA and FedRAMP Path
Bedrock carries a serious compliance footprint: ISO, SOC, HIPAA eligibility, GDPR, and FedRAMP High in AWS GovCloud, with a DoD IL4 and IL5 authorization path for Claude through GovCloud. That is what makes the regulated version of this architecture possible at all.
One caveat worth stating plainly, because we will not overclaim it: FedRAMP and DoD authorizations cover specific models at specific times, and the covered list has historically lagged the newest Claude releases. If a FedRAMP High or IL5 boundary is a hard requirement, confirm which Claude model is authorized in GovCloud today before you design around a particular version.
How to Choose
The decision comes down to who you want owning the model and the governance.
- Pick Claudeforce when you want the fastest path, have no AWS estate to run, and are happy for Salesforce to manage the model inside its boundary.
- Pick bring-your-own on Bedrock when you already run on AWS, need a specific account or region, or want your own AWS commitment and controls behind the model, while keeping the Trust Layer.
- Pick direct on Bedrock when the workload is a back-office or batch process rather than a live agent, and you want full control of prompt, batching, and cost.
Many organizations end up running two of these at once: Claudeforce for agent-facing work, and a direct Bedrock pipeline for bulk processing. That is a healthy pattern, not a contradiction.
Where to Start
Cloudsheer sits on both sides of this join. We are an Anthropic Claude Partner and a verified Salesforce Consulting Partner, and we deploy Claude through Bedrock inside Agentforce for clients in regulated industries, which is exactly the crossover our AWS practice is built around.
If you are weighing these three paths, the fastest way forward is a scoping call. Bring your compliance constraints, your AWS footprint, and the workload you have in mind, and we will map it to the right one. You can also read how we connect Claude to Salesforce for the build-level view, or explore the Claudeforce practice for the managed path.