How to Bring Your Own AI Model to Salesforce Data Cloud

With BYOM for Data Cloud, you can plug in custom AI models - whether predictive, generative, or Retrieval-Augmented Generation (RAG) - and combine them with real-time, unified customer data. The result is smarter insights and faster automation.

New to the Data Cloud or BYOM world? We’ll explain exactly what it all means, the benefits (including how RAG pipelines work), and how to get started with AI models that are tailored to your business.

Data Cloud pulls customer data together from all your systems - Salesforce apps, third-party tools, and more - into one real-time, actionable view. It’s built to power smarter marketing, automation, and decision-making.

‘Bring Your Own Model’ (BYOM), found within Einstein Studio, lets you bring in external AI models and run them on this unified customer data.

So, if your data science team has fine-tuned a model using Amazon SageMaker, Google Cloud Vertex AI, or Databricks, you don’t have to start over. With BYOM, you connect it to Data Cloud and start making it work for your business.

Imagine you’ve customised an open-source large language model (LLM) like LLaMA or DeepSeek using Hugging Face Transformers, compressed it for efficiency, hosted it on SageMaker, and now you want to use it alongside your Salesforce data. BYOM makes this possible without any need to rebuild or rehost. Your model connects directly and starts delivering predictions, content, or RAG outputs instantly.

BYOM offers a host of benefits:

1. Make the most of your existing AI

If you’ve already invested time and budget into a custom model, say a trimmed-down LLaMA version trained on your business data, then naturally, you want it to deliver value. BYOM helps you do just that. By combining your model with Salesforce’s real-time customer data, you unlock insights that support your marketing goals.

2. Flexibility across platforms

Salesforce’s platform-agnostic approach means you’re not tied to one ecosystem. Whether your model runs on Amazon SageMaker, Google Vertex AI, or Databricks, BYOM supports it. It gives your team the freedom to use familiar tools while tapping into Salesforce’s infrastructure. That flexibility means you can choose the best tools for your team now, and easily switch or upgrade later without losing momentum.

3. Faster, cheaper integration with zero-copy access

Gone are the days of clunky ‘Exchange, Transform, Load’ (ETL) processes. BYOM uses what’s called a ‘zero-copy’ architecture - your model doesn’t need to move or duplicate data to work with it. Via secure APIs, it securely accesses Data Cloud in real time, cutting latency and keeping costs down. This is particularly powerful for RAG pipelines, where fast data retrieval is key.

4. Real-time personalisation with RAG pipelines

Retrieval-Augmented Generation (RAG) pipelines are where BYOM truly shines.

Let’s say you’ve trained an AI model on your product catalogue and customer interactions, then connected it to Salesforce via BYOM. That model now taps into real-time data, such as purchases, chats, or browsing history, and uses it to generate personalised responses or content instantly.

For Sales, this could mean suggesting a matching phone case right after a mobile purchase. For Support, it’s crafting instant replies using the customer’s history. And for Marketing, it’s generating emails that feel like they were written one-to-one.

What’s more, Data Cloud knows which bits of customer data are most relevant, so it feeds the best context into your model. This results in more accurate responses, higher engagement, and less manual work. With BYOM and RAG, your AI is building loyalty, boosting sales, and cutting wasted efforts. You’re ready to scale.

5. Built-in security and governance

Security isn’t an afterthought with BYOM. The Einstein Trust Layer ensures any external model you connect via BYOM has robust protection and compliance.

Whether your model is generating emails or scoring leads, everything it does is logged and safeguarded. You get full audit trails and controls to meet GDPR, PCI DSS, or other compliance requirements.

You can also mask personal data before it ever reaches your external model, which is essential for protecting identity when using third-party LLMs. Add in role-based access controls (RBAC) to ensure only authorised personnel can manage or invoke these models, and real-time monitoring for anomalies, and you’ve got a secure, compliant setup that doesn’t get in the way of innovation.

6. Lower costs, easy to scale

Hosting your own model externally (e.g. on SageMaker with pay-as-you-go pricing) can be far more cost-effective than building within Salesforce from scratch. And because BYOM works with Data Cloud’s zero-copy model, you save on data handling, too. When your needs grow, it’s simple to scale, just increase your inference capacity or add endpoints without needing to rebuild anything.

If this all sounds a bit technical, don’t panic. The process for integrating your AI model into Data Cloud via BYOM is surprisingly structured. Here’s how it works

1. Prepare your model

You’ll need a generative AI model that’s already trained on your business data, like customer feedback, product docs, or past interactions. It should be hosted on a secure platform like Amazon SageMaker, with a publicly accessible (but secure) endpoint.

2. Connect it to Data Cloud

Head into Einstein Studio and register your AI model’s endpoint - typically a URL provided by your hosting platform. You’ll set up secure authentication (like OAuth 2.0 or an API key), and from there, Data Cloud can access your model and start sending real-time customer data to it.

3. Map your data

Use Data Cloud’s tools to decide what data your generative model will use: support tickets, orders, product info, whatever is most relevant. This ensures your model always has the context it needs to generate useful, accurate content.

4. Test before launch

Once your model is activated via Einstein Studio, test it in a Salesforce sandbox environment. For example, simulate a customer query and see how the model responds. This helps you fine-tune outputs and make sure everything runs smoothly.

5. Embed in workflows

Now it’s time to put your model to work. Salesforce Flow lets you trigger the model in real-world processes, like generating a personalised support response the moment a case is created.

Using Prompt Builder, you can create templates e.g. “Case Response Generator”, that instruct your model on how to respond. For example, ‘Generate a polite, concise reply addressing the customer’s issue. Using dynamic data, these templates pull in real-time customer data and feed it into your model, which then outputs a ready-to-send reply or campaign message.

Templates come in all shapes, whether it’s filling in a specific field or writing an entire email, and they ensure your AI outputs are consistent, relevant, and aligned with brand tone.

BYOM Data Cloud integrations can supercharge industries across the board. Here are a couple of practical ways this could play out:

Healthcare: Personal follow-ups

A healthcare provider uses BYOM to connect a generative AI model to patient data in Data Cloud. A Prompt Template called “Patient Follow-Up” helps the model generate check-in messages like, “We noticed it’s been a while since your last visit, shall we book one?”

Triggered automatically via Flow when an appointment is missed, this message could be sent through Health Cloud or Experience Cloud, keeping patients engaged and improving care continuity.

Retail: Smart product suggestions

A retail brand combines Data Cloud and BYOM to recommend products based on recent purchases. A template might say: “Enjoyed your new jacket? Here’s a scarf to match.”

Salesforce Flow triggers the suggestion after a Commerce Cloud transaction, and it lands directly in a Marketing Cloud email or a Sales Cloud follow-up, helping to boost repeat purchases.

Want to explore how BYOM could work for your team?

Book a call with MarCloud. We’ll show you what’s possible, guide you through the setup, and handle the heavy lifting, so you can focus on creativity, not code.