Returning to work from the excitement of Dreamforce, there is only one question on everyone’s mind: What is Agentforce?
From the moment Marc Benioff unveiled Agentforce and all of its glory in the main keynote, our team dug in with Salesforce experts and at hands-on workshops to understand how Agentforce works and more importantly how it can help businesses improve customer experiences.
This product and its related technology are new, so it may take an extra moment to let all the concepts sink in. Before we get into the nitty gritty, it’s important to understand what is new and powerful about this technology.
Agentforce combines the power of leading AI models with your company’s data
Agentforce will let you create autonomous agents that do work on your company’s behalf. The chief reason Agentforce is different than its predecessor, the AI chatbot, is that Agentforce has the ability to receive questions from a customer, pull data from your company’s database, add that information from your customer system into the prompt being asked and return a more accurate and relevant response to the customer. The term for this in the industry is RAG: retrieval-augmented generation.
In the past, AI could not go into your company data securely and reliably pull the information needed to help answer a question. The vision of Agentforce is to let businesses create far more reliable and helpful AI agents by giving them the ability to incorporate data from within a business’ data cloud.
Your company data will not be used to train these public AI models
If you are like me, my first concern was the security of data being sent through an AI model. I talked to many Salesforce experts about Agentforce and got clear confirmation that the models available for Agentforce (such as OpenAI GPT 4 or Anthropic Claude 3.5 Sonnet) will not train on proprietary company data. One security engineer let me know that there are many legal agreements and security measures in place so that third-party analysts and engineers cannot see company data. Please note that Salesforce is taking an open model approach to Agentforce so you can bring your own LLM as well! If that is the case, you should make sure the LLM you use is not training its model on enterprise customer data. For example, in a talk from the team working on watsonx from IBM, they made it very clear that no customer data is used to train their model. Same is true from an employee from Cohere AI.
With all this context set, lets get into the details of what Agentforce actually is.
What actually is Agentforce?
Agentforce is a new product that lets businesses build Agents with a tool called Agent Builder. These agents are autonomous, meaning that they can be triggered automatically to do certain things for customers. They can be accessible much like a support chatbot seen on many websites but they can also be accessed through other channels like phone calls.
How does Agentforce work?
Agents are configured to specify how they can be reached by customers. Once an agent is created, you can create topics that explain how the agent can help if a customer’s question or request falls within the scope of a topic. Agents can have multiple topics.
For example, an agent can be a service agent and within this agent, a business can create topics the service agent can help with, like order returns, shipping changes, or questions on a price match policy. This can augment your staff on your customer support team so that AI agents can resolve a portion of support tickets without a person needing be involved. Salesforce emphasized how helpful this can be during seasonal spikes in support tickets for certain businesses, like back-to-school season for a book store.
Setting up Topics for Agents
For each topic for an agent, your team can set up powerful actions for the AI to take to help address the topic specified. There are a few key things that need to be defined as topics for these actions to be triggered. You need to specify:
- When the topic should be triggered for the agent
- The scope of when an agent can and cannot help for the given topic
- Instructions for what the agent should do to address the topic
To make this more tangible, let’s run through a real world example. Lets imagine we have a customer service agent and we are setting up the topic of order returns (both things I defined myself). In simple terms, the purpose of the service agent is to autonomously resolve issues related to customer support that are straightforward but that usually take up the time of support reps who could be working on more important things.
When creating the order returns topic, an admin will need to specify that when a customer requests an order return, the agent will be tasked to help the customer. An admin specifies this in the classification description text field.
An admin then needs to specify the scope of the agent on this topic. For example, an admin may want to specify that if the return is a unique or unusual situation that is not clearly specified in company policy, then the ticket should be escalated to a human support rep to help out. An admin defines this in the scope text field.
Giving Agents Instructions
Next, an admin needs to specify the instructions for the agent for what to do when the topic is triggered given that the customer has specified an order return and it is not unusual or unique. These instructions may tell the agent to get the order id, check that a return is still possible and then if a return is allowed, give the customer a shipping label pdf they can use to make the return.
Finally, you have to define actions related to this topic. These actions do not need to be triggered by the agent in a linear order. Instead, these actions act as a list of all the abilities a given agent has to do in order to complete the instructions. Actions can be AI prompts, flows, apex code, or API calls. For the agent to process an order return, we may give the agent the action to be able to look up orders based on customers and records associated with the customer’s orders. This way the agent can confirm, for example, that the purchase was recent enough to still qualify for a free return.
Let the AI Agents Do the Heavy Lifting
Once all of this is set up, an AI agent like this Service Agent is able to automatically process typical order returns without human involvement, relieving a support team from a repetitive task. To set this up in your Salesforce org, Agent Builder offers a preview panel to make sure the agent can do the task you have set up with the given actions you configure. Salesforce also has some other tools for testing and monitoring to make sure autonomous agents are running smoothly.
This example was intentionally simple to explain the functionality of Agentforce but the functionality of Agent Builder, prompt builder, flow builder, data cloud and apex code can allow businesses to set up agents to solve more complex issues.
Final Thoughts
To avoid complexity, this article just covers Agent Builder but there is more to know about Salesforce’s other tool, Prompt Builder that works closely with Agent Builder. We plan to cover that tool soon.
The Match My Email team is very interested in the ways email and calendar data can help make Agentforce more powerful for customers. By storing email messages and events in Salesforce, we learned that our customers will be able to easily access that data in Agentforce. We loved getting hands on at Dreamforce and look forward to experimenting further as Agentforce gets released more broadly.
In terms of pricing, how much Agentforce will cost wasn’t quite clear despite asking many people about this. We did hear that Agentforce will cost $2 a conversation but this pricing definition has lots of room for interpretation and will surely become clearer in the coming months. To fully understand this, it sounds like we will all have to wait and see. What we do know is that the cost model could be a mixture of a subscription model and a usage model.
