Einstein Copilot uses generative AI plus two key techniques: grounding injects your business data into prompts so answers are relevant to your company, and actions let it execute work by mapping user requests to configured tasks like queries and flows.
What makes Einstein Copilot possible? Einstein Copilot helps you work faster by allowing you to express your questions or instructions in natural language and giving you generative AI responses that are grounded in your business data. But the key innovation that makes this experience possible is Einstein Copilot's ability to turn AI generated plans into actions that are grounded in your business data. Which means to really understand how this works, we need to talk about grounding actions and plans. Let's start with grounding. Generative AI consists of three basic parts. There's the prompt, which is the question or instruction we give to AI. There's the large language model that processes the prompt and the response, which is the answer we get from AI. But vague questions produce vague answers that have nothing to do with our business. And the solution is a technique called grounding, where we add the missing business context into the prompt. And now we get a more relevant generative AI response. And with Prompt Builder, which you've been looking at here, you can create dynamically grounded prompts that adapt generative AI to your business, which gives you access to much, much more relevant responses. And if we return to the employee experience, Grounding is why Einstein Copilot can answer questions about your business data, like summarizing the service. But grounding doesn't explain how Einstein Copilot can respond to a request to update this case, a request that involves action, which leads to our second actions. Behind every Einstein Copilot response is a library of actions that give Einstein Copilot its capabilities. This query records action is why Einstein Copilot is able to answer questions about Salesforce data. And this update record action is why it's able to modify that data. These are just some of the standard actions we're shipping. And we'll soon be shipping actions for sales, service, marketing, commerce, industries, and more. But what is an action? Well, an action is just a pointer to where work gets done. It can be based on a flow, which is used to create workflows in Salesforce or Apex, which is used to extend Salesforce with code, or a prompt template, which is what we were creating in Prompt Builder. So let's imagine I want to extend Einstein Copilot with the ability to recommend products. And I have this flow which identifies the most appropriate products for each account. It accepts an account id. Then, based on that customer's details and purchase history, it returns the products that would be most appropriate to sell. And I can equip Einstein Copilot with the ability to use this flow by just selecting it. Although I'm not quite finished. I also need to describe this flow to Einstein Copilot because this description is how Einstein Copilot will know when to use this action. And I know that feels unintuitive, but we'll get there in a moment. For now, this is it. And just like that, what we'll see is that Einstein Copilot now has the ability to make a product recommendation. But how does Einstein Copilot know which action or actions to use for any given input? And that is why we need to talk about plans. Here, inside this playground, you can inspect how your Einstein Copilot works, which includes the actions you've configured and the descriptions you've written for each action. And we know that each of these actions returns data that we can use to ground the model. So let's do this. Let's test this copilot by typing summarize the ACME account and we'll see that Einstein Copilot is able to create and execute a dynamic multi step plan that finds the account and summarizes it. And if I respond with what products can I sell for this account? Notice I'm understood in the context of an ongoing conversation, but how is any of this working? Well, Einstein Copilot is using large language models to understand my intent and our library of actions based on those descriptions. It's then using this understanding to generate a dynamic plan for responding to this input, which it maps to actions that it executes using my user permissions. And the output is then used to ground the model and generate a relevant response, which I know is a lot to absorb. So let's simplify. Imagine I say to you, here is what I was asked to do. Here is what previously happened. Here is what I could do. Make a plan and tell me what I should do. That is what is happening here. And the key innovation that makes Einstein Copilot possible is its ability to turn AI generated plans into actions that are grounded in your business data. And that is how you get a trusted conversational AI assistant that helps you work.
