Salesforce's unique approach to AI

A summary of Salesforce's approach to AI (as of May ‘24)?

Salesforce's vision for data and AI and how it operates across the platform is substantial. This is the highest velocity of product innovation I've seen in my last 10 years as a Salesforce consultant.

In order to help guide our teams and customers, I felt it would be best to first distill what I've learned as this rapidly moving capability continues to unfold. Perhaps this will help you conceptualize it all as well.

Overview?

Salesforce’s approach to AI uniquely fuses predictive modeling with generative AI that is available both programmatically and as part of its process framework, grounded by data and backstopped by a unique “Trust” layer.

In order to understand Salesforce’s approach to AI it is best to think of Salesforce not as a series of applications but as a customer engagement platform. This is Customer Relationship Management (CRM) in the broadest sense, as it captures all interaction points and related engagement with a customer.?Inside the organization, Salesforce uniquely bridges back-office data and process with front-office engagement and experiences. Acquisitions like Slack, MuleSoft and Tableau provide the connective tissue that enable experiences and engagement.?

In the broader context of the AI ecosystem, most of Salesforce’s AI solutions sit in the application layer. In other words, Salesforce is more focused on creating end-user value and?business impact than training high-parameter Large Language Models (LLM’s) or developing custom hardware and infrastructure on which to run them. Because businesses invest in different tech stacks, Salesforce is taking more of a “marketplace” approach to AI development, enabling businesses to choose from existing models (OpenAI, Anthropic, Vertex, etc.), or to bring their own. These will be enabled by a powerful set of developer tools (declarative and code-extended).?

Deconstructing Salesforce’s AI capabilities?

Salesforce’s AI capability set can be broken down across Machine Learning and Gen-AI. It includes Out-of-the-box features and fully customizable capabilities that can be deployed in both the flow-of-work and custom application development.?

The following graphic provides rough categorization of these capabilities.?

Out-of-the-box capabilities?

For years, Salesforce has included Machine Learning capabilities throughout its platform. Most of these have been “black box” applications used to support predictions. This “democratized” AI came packaged as platform features like “Einstein Insights,” or “Send Time Optimization;” but included little, if-any, customization capabilities.?

Recently, several Gen-AI features have been introduced. The initial group of capabilities include what will become “table-stakes” features that one would expect. For example,?the ability to draft email content in Marketing Cloud, catalog descriptions in Commerce Cloud or sales emails in Sales Cloud. The second group of capabilities involve Copilots to guide users and provide insights through key workflow related activities.?

All these capabilities exist within the “flow of work;” providing insights and guidance during regular user and?administrative activities. ?

Custom-developed capabilities?

The Einstein Copilot Studio enables the development or integration of both ML and Gen-AI capabilities, that work together, are contextually grounded by data and content and are accessible in workflows, programmatically via API, or even by other AI Copilots. The Studio is composed of 3 underlying builders.?

• Prompt Builder – create sophisticated prompts grounded by company data, using?an LLM of your choosing.?
• Action Builder – gives your Copilot the ability to invoke actions (create & edit records, invoke workflows, research answers, etc).?
• Model Builder – guides you in building a new ML model, or ingesting the outputs from common platforms data scientists use like Google’s Vertex or AWS Sagemaker.?

Data Cloud becomes a key part of Copilot orchestration, which provides grounding for both structured (data) and unstructured (i.e., documents) information.?

And all of this is kept safe by the Einstein Trust layer which prevents company and customer sensitive information from ever leaving Salesforce. As usual with Salesforce, a well-trained admin can develop in copilot studio without code, and there’s mechanisms in place to extend beyond out-of-the-box capabilities with code.?

Sophisticated grounding?

The accuracy of response you may receive from an LLM can vary widely based on how strictly it is prompted, and what context is provided to answer within. This context is called “grounding.” Salesforce is enabling multiple types of grounding that can be used in the same prompt set.?

• Field grounding - directly references fields from Salesforce and Data Cloud in the prompt?
• Dynamic grounding - can incorporate data from other parts of the platform (i.e., retrieve notes from the last sales interaction)?
• Document-based - grounding is anticipated and would enable incorporating unstructured information as part of a response (i.e., knowledge base)?

These ideas can be strung together into super sophisticated prompts that can be exposed to internal users and external consumers alike.?

When to choose building on Salesforce vs building bespoke?

There’s no one perfect answer to this question; however these scenarios will heavily favor Salesforce’s overall value proposition:?

• Primary use-cases are customer engagement / CRM focused?
• Organization is well invested in Salesforce and is evaluating or has Data Cloud?which is a required foundation for the more advanced Copilot capabilities
• Availability of pre-built integrations to key data sources?
• Availability of application developers and ability to prioritize competing workloads (will favor declarative development in Salesforce)?

Wrapping it up

This is just the beginning. I don't think the value of what Salesforce has done here should be understated. Copilot Studio is a declarative tool that enables the development of highly focused and grounded AI functionality using your company's data in a trusted environment. There are both out-of-the-box and highly customizable capabilities. And there's the ability to orchestrate the interaction between predictive modeling, generative reasoning and business workflow to create more efficient, engaged, customer interactions. This orchestration will be central to creating highly effective, industry- and company-specific AI applications.