All insights

Agentforce

Agentforce data residency: geo-aware routing and keeping model requests in your region

Agentforce sends your prompts to a large language model that may run in a different country than your org. Geo-aware routing keeps those requests close to your instance, but it falls back to the US unless you turn on one fail-closed switch, and it does not cover batch. This is what the setting does, how your region is decided, and the audit to run before a residency review.

An agent you build in Frankfurt can send a customer’s message to a model running in Virginia. The Agentforce setup screen will not mention that it happened.

A sentence like that ends a lot of residency reviews badly. The grounding data sits in your Data 360 region, the org is on a European instance, and everyone assumes the whole pipeline stays in Europe. The prompt, the retrieved records, and the model’s answer are a different hop, and by default that hop goes wherever the model can be served.

Salesforce has a control for this, Geo-Aware LLM Request Routing on the Einstein Generative AI Platform. It keeps your model requests close to your instance by default, but it still falls back to the United States unless you turn on a second setting, and it does not apply to batch. If you have a residency obligation, the fallback is the part you have to think about, not the routing.

What geo-aware routing routes

Every Agentforce answer, every Prompt Builder template, every grounded reply runs through the Einstein generative AI platform, which calls a model. Geo-aware routing decides which copy of that model gets the call.

The platform sends the request to the model servers nearest your instance, for the models that exist in that region. Salesforce documents the behaviour for OpenAI, Anthropic, and Gemini models, with Anthropic served on Amazon Bedrock and Gemini on Google’s Vertex AI. If the model you picked is hosted in your region, the request stays there.

The word doing the work is “nearest”. Routing is a proximity decision, not a residency guarantee, and the difference is the whole post. Nearest becomes in-region only when the model is available in your region and you have closed the fallback.

How your region is decided, and why Data 360 is in the sentence

You do not set the region per agent. It comes from the instance.

For instances enabled on or after June 13, 2024, the Einstein generative AI platform region matches your Data 360 region. The tidy case follows from that: provision Data 360 in the EU, and the model routing resolves to the EU along with it. For an older instance, the mapping is not something you read off a screen, and Salesforce points you to your account executive to confirm it.

The same mapping is why the Agentforce Vibes team made Data 360 a prerequisite for turning geo-aware routing on in the coding agent. The region has to come from somewhere, and Data 360 is where the platform reads it. When the setting is on, the Vibes model picker shows only the models served in your region, so a developer cannot pick a model that forces a cross-region hop without noticing.

Check your own mapping before you promise anyone a region. “We’re on an EU instance” is not the same statement as “our model requests resolve to the EU”, and only the second one survives an audit.

The default is fallback, and fallback means the US

Here is the behaviour that catches teams. For Salesforce-managed models, a request that cannot be processed in the designated region is re-routed, and the final fallback is a US region, so that the user gets an answer instead of an error.

That default is the right call for most orgs. An agent that returns “model unavailable” because a regional endpoint had a bad minute is worse than an agent that completes the request from another region, for a company with no residency rule.

For a company with a residency rule, the same default is the problem. The request you promised would stay in Frankfurt completes in Virginia, succeeds, and leaves no error for anyone to notice. The failure is silent, which is the worst kind in a compliance context.

The one switch residency teams need: fail closed

The setting that changes this is Enable In-Region Model Requests Only, in Einstein Setup. Turn it on and the final US fallback is removed for the models it covers. A request that cannot be served in your region now returns an error instead of completing elsewhere.

Fail-closed is what a residency obligation requires. You are choosing a visible error over an invisible cross-border transfer. An error is a ticket someone files and you fix. A silent transfer is a finding in next year’s audit.

Two details to get right before you flip it.

The OpenAI and Azure OpenAI path has its own handling and a toggle with slightly different wording, so read the setting text in your org rather than trusting a blog’s label, this one included. Requests still try the primary and secondary regional endpoints first; the setting only removes the last-resort US hop.

And the newest models are the ones most likely to break under a fail-closed rule. A model Salesforce launched last quarter may not be hosted in every region yet, so turning on in-region-only can turn “use the latest model” into “use the latest model, except where it errors in your region.” Pick the model and the residency posture together, not separately.

Batch does not follow these rules

The caveat that undoes a careful setup: batch processing does not honour geo-aware routing, and gives no in-region guarantee.

If you run large asynchronous jobs that call a model, a nightly classification over a data lake object, a bulk summarisation, a scored segment, those do not inherit the fail-closed switch. A residency design that only reasons about the interactive agent and forgets the batch jobs has a hole in exactly the place where the most records move at once.

List your model-calling batch work separately and treat it as out of scope for in-region-only until Salesforce documents otherwise. If a batch job must stay in-region, that is a question for your account team, not an assumption.

Government Cloud is the exception that proves the model

Government Cloud orgs do not get a toggle, because they do not need one. Agentforce Vibes routes AI requests from a Government Cloud org to a dedicated Government Cloud endpoint, with no setting to change and no fallback to civilian infrastructure.

Every residency-sensitive customer wants that shape: routing that is a property of the environment, not a checkbox a new admin can clear. The commercial geo-aware controls are the configurable version of the same idea, which is why the configuration discipline matters more. On commercial Hyperforce, the guarantee is only as good as the settings, and settings drift.

Where routing sits next to the Trust Layer

Residency is one of two questions people conflate. The other is retention, and the Einstein Trust Layer answers that one.

The Trust Layer’s zero-retention agreements with the model providers mean your prompt is not stored by the provider or used to train their model. Geo-aware routing is about where the request is processed, not whether it is kept. You can have zero retention and still have a cross-border transfer, which is why a GDPR or sovereignty review needs both answered, not one standing in for the other.

If you are also weighing whether to register your own model, the bring-your-own-LLM path changes who hosts the endpoint but not this reasoning: you still decide the region and the fallback, you just decide them on your side. And if the obligation is public-sector, the GovCloud route removes the configuration risk entirely. The EU AI Act’s transparency and risk obligations sit on top of all of this and are a separate exercise again.

The audit to run before the review

Four checks, in the order they tend to fail.

  1. Confirm the region mapping. For a post-June-2024 instance, verify your Data 360 region is the one you think it is, because the model routing follows it. For an older instance, get the mapping in writing from your account team.
  2. Decide fallback posture per obligation, not per org. Turn on Enable In-Region Model Requests Only where a contract or regulation requires it, and accept the errors it will surface. Leave it off where availability matters more than residency.
  3. Inventory the batch jobs that call a model. Each one is outside the in-region guarantee today. Write down which ones touch regulated data and raise them with Salesforce rather than assuming.
  4. Pin the model to a region-safe choice. If in-region-only is on, test that your chosen model is served in your region, and re-test when you upgrade the model, because the newest ones lag in regional availability.

None of this is exotic. The discipline is the same as any other governed data foundation: know where the data goes, choose the failure mode on purpose, and test the thing that fails silently.

Understanding the basics

Does Agentforce keep my data in my region?

The grounding data stays in your Data 360 region, but the model request is a separate hop. Geo-aware routing sends that request to the servers nearest your instance by default, and falls back to the United States if the model cannot be served in your region, unless you turn on Enable In-Region Model Requests Only in Einstein Setup.

What does Enable In-Region Model Requests Only do?

It removes the final US fallback for the models it covers, so a request that cannot be processed in your designated region returns an error instead of completing elsewhere. The setting gives you the fail-closed posture a residency obligation usually requires, at the cost of a visible error when a model is unavailable regionally.

How is my Agentforce region determined?

For instances enabled on or after June 13, 2024, the Einstein generative AI platform region matches your Data 360 region. For older instances, confirm the mapping with your Salesforce account executive, because it is not shown directly in Setup.

Does geo-aware routing cover batch jobs?

No. Batch processing does not follow geo-aware routing and gives no in-region guarantee, so asynchronous model-calling jobs are outside the fail-closed switch. Inventory them separately when you design for residency.

Choose the failure mode on purpose

The mistake is treating residency as a thing you have or do not have. On commercial Agentforce it is a setting, and the default trades residency for availability without telling you it made the trade.

So make the trade yourself. For most orgs the default is fine and the latency and uptime are worth more than a region guarantee nobody is asking for. For the orgs under a real obligation, turn on in-region-only, accept the errors, and remember the batch jobs. The feature does the routing. You own the fallback.


If you are standing up Agentforce under a residency or sovereignty obligation and want the region mapping, the fallback posture, and the batch jobs reviewed before your compliance sign-off, talk to us. A Data & AI strategy engagement is where that review usually starts.

Keep reading

All insights