Does Salesforce AI send data to OpenAI? The short answer
If you use Salesforce AI with an external model path, the useful question is not only whether data is sent. The useful question is what exact version of the data reaches the model after Salesforce applies its own masking and trust controls.
In a practical test, the interesting result was not "everything is exposed". The interesting result was that masking behavior looked much stricter than many teams expect. If your team is simultaneously evaluating a custom model route, pair this with Use Your Own AI Model in Salesforce.
How Salesforce Trust Layer changes the payload before the model call
With masking enabled, the payload that reaches the model may be materially different from the raw Salesforce record. Names, emails, phone numbers, and similar fields can be transformed before they leave the Salesforce side of the flow.
That does not remove the need for review, but it changes the risk conversation from a vague fear into a testable question: what exactly is visible after masking in your own configuration. For the broader product position, see Security.
How to verify what Salesforce AI actually sends out of your org
A good privacy review should test real examples, not only read product copy. Use a sample record with known values, enable the masking settings you plan to rely on, then inspect what the model actually receives.
You should also verify the workflow boundary. A summary, prompt, or custom model flow can have a different data path than teams assume. Related pages: Security, Salesforce AI Tools Without Einstein, Use Your Own AI Model in Salesforce, and Guides.
FAQ and next step
Does Salesforce AI always send raw CRM data to OpenAI?
That is the wrong default assumption. The actual payload depends on the workflow and on the trust or masking settings in use.
Is masking enough on its own?
Masking helps, but teams should still test their own configuration and confirm what reaches the model in practice.
Why does this matter for smaller teams?
Because privacy concerns often block useful AI workflows long before technical quality becomes the real issue.
How does DealScope approach this?
The public security position is narrower: minimize data sent to the backend and avoid transmitting personal data where possible. Start with Security, then compare workflow-specific pages like AI Follow-Up Email for Salesforce.
Next step:
If you are comparing privacy posture across approaches, continue with Use Your Own AI Model in Salesforce. If the bigger question is whether you need Einstein at all, review Salesforce AI Tools Without Einstein.