This article is a simple introduction to Salesforce’s Agentforce.
What is Conversational AI?
Conversational AI is a machine intelligence which is consumed via a human like conversation experience. For eg. – chat/voice/video interaction. If you have ever used OpenAI’s ChatGPT tool its a good example of conversational AI.
What is Agentforce (Einstein Copilot)?
With Agentforce you can implement conversational AI capabilities and provide an array of AI assistant experiences to your end users, customers and stake holders on Salesforce platform. In simple terms it helps you build human like AI Agents/Assistants beyond normal chat bot capabilities.
What are prompts?
An well structured chat instruction text that will generate a predefined output using conversational AI interaction with AI assistants.
What are the building blocks of Agentforce?
While Salesforce Einstein is a much bigger offering, the Agentforce part of it consists mainly of:
- Agent Studio: Where you build multiple Agents [Consider these as hiring specialists around specific business function]
- Agents: These are the agents you configure as per your business needs within the Agent Studio. You can deploy these agents in multiple ways.
- Prompt Builder: Prompt builder helps you build prompts using available LLM or Foundational Models. Here you can make use of grounding feature which introduces Salesforce DATA into prompts for generation of business specific responses.
- Agent Topic: Let’s imagine a teacher is teaching you Mathematics. At a given timestamp he/she can offer her expertise around certain topics like Geometry. Agent Topic are exactly same, you define a scope and classification around your own defined topic with some extra topic specific instructions. It helps the Agent to further pick from a set specific actions relevant to the topic once it identifies what topic the human in conversation is talking about in current context.
- Agent Actions: These are the advanced tasks within your CRM instance that your AI Agent can perform during conversations to better assist with the human task. There are three types of Agent Actions which you can mix mash smartly:
- Apex: To invoke an apex logic like: Calling your custom Apex to finalize the cart and return confirmation.
- Flow: To invoke flow operation, similar to Apex, but via Flow.
- Prompt Template: To reuse other prompts as an action. For example a prompt to summaries things in a predefined format.
- Plan: Agentforce dynamically generates a plan based on the intentions of a conversation that it detects using its own proprietary reasoning engine. These plans help Agentforce detect and pick the correct set & sequence of topics and actions to perform. While you build you agent, you preview the plan yourself and improve it further to your business needs.
- Just like you have a Flow Builder to manage Process Automations in Salesforce, you can also build and manage your Agents in Agent Builder. Agent Builder shows you every related configuration around your Agent in a good user interface. Agent Builder can be accessed from Agent Studio.
- Other: There are also other items like Language Settings and Data Library, however for ease of reading I will keep it out of scope for now. Know that these others tools help you elevate your agent experiences further.
What is grounding?
A simple prompt will generate simple response based on the Foundational Model. We want to generate these responses with respect to your enterprise data (Salesforce + Data lake integrations i.e. Data Cloud). For same you can plug your CRM data into these prompt dynamically just like email template merge fields.
Is this secure?
Every GenAI request under Einstein offering in Salesforce goes through a security wall called as Salesforce AI TRUST layer. This Trust layer implements following security controls to protect your data:
- Salesforce MASKS the grounded data in backend such that it DOESN’T travel into Foundational/LLM Models.
- Salesforce has a strict ZERO retention agreement with available Foundational Models.
- Data grounded in prompts RESPECT the data permissions of the user in context.
Sample use cases:
- Helping with Lead qualification.
- Handling order enquiries in real-time.
- Knowledge assistance.
How is Agentforce priced?
Agentforce pricing is based on a unit of Conversation. A Conversation starts when a user sends first message in the chat OR clicks an Option provided by the Agent. Once a user clicks the END chat button, OR confirms to end the chat via conversation OR wait for a period of 24 hours from the first chat conversation message – it completes one Conversation unit. For current pricing visit: https://www.salesforce.com/in/agentforce/pricing/
View official Salesforce demo videos: HERE
Here is how your Agentforce might feel working with me: HERE.
If you are interested in thinking beyond Agentic AI use cases watch this: Here.