Agentforce was the talk of the town in San Francisco this week during Salesforce’s annual Dreamforce conference — and for good reason! Agentforce is the kind of solution that will fundamentally change how people and organizations interact with their CRM.
But as with any new product announcement, we know there will also be a lot of questions. I experienced that firsthand this week while presenting at Dreamforce and talking to admins and developers who are excited to get started with Agentforce.
So, in this blog, I’ll outline in more detail precisely what Agentforce is, how it works, and how you can use it to drive automated actions in your org to drive value today.
How Agentforce extends Copilot and Prompt Builder
Many Dreamforce attendees who came to San Francisco expecting to hear about Einstein Copilot were understandably confused when Salesforce rolled out Agents just the week before the conference. But the differences between the legacy Einstein Copilot and the new Agentforce are important to understand.
Essentially, Agentforce Agents are a rebranding of Copilot Agents, with one critical caveat — they extend the functionality of Copilot to create more autonomous agents that are capable of doing things like summarizing/generating content, as well as taking specific actions.
Here are a few key changes in vocabulary that are good to know:
- Agentforce: The overall name for generative AI-based agents in Salesforce.
- Agent Actions/Agent Studio: Previously called Copilot Actions/Copilot Studio, these are the specific actions that an agent can perform, such as Identify a record, Summarize a record, or Write a Sales Email.
- Prompt Builder is still Prompt Builder. Use this tool to templatize generative AI prompts to various different LLMs (or an LLM of your choosing).
Like Einstein Copilot, Agents can take input from a user called an “utterance” — this is text that is entered into an Agentforce chat interface. The agent will then translate that utterance into a series of actions based on instructions that you can configure and enhance, allowing for opportunities to continue to improve performance over time. Once the Agent identifies the right plan to put together, it can then execute it and provide a response to an end user.
Understanding Agents: Topics
A key difference between Einstein Copilot and Agentforce Agents is that Agentforce adds a new layer called “Topics,” which can allow for greater flexibility and an increased number of supported actions. Topics essentially sort actions by business function, so when a user enters an utterance, Agents first determine the relevant topic and then must determine the relevant actions to address the user input.
Why add the topic layer? Including topics helps Agents avoid getting confused and calling the wrong action. By adding this layer of organization to the list of actions an agent can take, you can significantly increase the number of custom actions that can be built and supported (previously, Copilot only supported 15-20 actions, but Agentforce can support many more).
Understanding Agents: Actions
Actions remain the same in Agentforce as in Einstein Copilot. These are the tasks that an agent can carry out in executing a plan. Whenever you enable Agentforce, you will have a laundry list of standard actions that are activated right out of the box. If you’re looking for a quick win, standard actions are the way to go.
However, if you want to start to customize actions with your own data/objects — or perhaps you need to tackle a bespoke use case that is a little different from what a standard action might provide — you can also create custom actions based on Apex, Flows, Prompts, or Service Catalog items (in beta right now).
Understanding Agents: Prompts
Remember that any time you use an LLM, you need to provide some input to get an output. Typically, this input comes in the form of a prompt where you might ask a question, give some directions, or provide some request to the LLM. Here’s a good rule to follow: If you want a better answer, ask a better question. Prompt Engineering is a critical part of working with LLMs to yield better results, reduce hallucinations and irrelevant responses, and drive better behavior from agent actions. Just as Einstein Copilot and Prompt Builder were closely tied together, Prompt Builder provides the underlying layer to build, test, templatize, and maintain prompts that are used in Agent Actions.
How Generative AI enhances CRM
GenAI tools like Agentforce offer exciting opportunities to enhance your Salesforce org in many ways, including:
- Productivity. For example, agents that provide summaries to end users can greatly reduce time for bankers, sellers, or other people working in Salesforce to prepare for meetings with their customers.
- Personalization. Emails that include pertinent client information can get sellers or service agents started on the right the path to manage issues with clients before they become significant points of friction.
- Standardization. By using standardized prompt templates, flows, or apex to drive agent actions, outputs can be based on the same data points, driving a more consistent workflow. For example, an account summary that different sellers use will use the same foundational information, so each user is making decisions based on the same information. Summarizing knowledge articles also helps standardize the knowledge base you have, allowing users to directly search for and summarize information stored in Salesforce Knowledge.
- Efficiency. Agentforce provides the ability to automate key workflows, like handling service cases, allowing people to focus their energy on more complex tasks.
These benefits, however, are only realized when CRM users adopt these technologies to make better decisions, work more efficiently, or enhance their customer experience. Change enablement and training are an important part of this mix, too, because most CRM users are not going to be used to an AI-assisted workflow. Eventually, organizations that adopt Agentforce will have to start to rethink their business processes in the context of an AI-assisted world.
Getting started with Agentforce
With all the excitement coming out of Dreamforce, it’s no surprise that many organizations are looking to get started with Agentforce right away. The good news is that there are quick wins to be had using these tools.
Our recommendation is to start using standard Agent actions and see where and how those out-of-the-box actions like Opportunity summarization, sales emails, and creating close plans work. Likely, you will need to make some changes to get the full value from them, but they provide a fantastic foundation to test some minor tweaks before you build fully-custom actions.
Testing and adoption
Testing in a generative AI environment is crucial, and differs from traditional approaches. You have to ensure that AI agents can handle the various ways users might phrase the same question (i.e., utterances).
Since AI relies on probabilistic models, responses can vary with each test, making repeated testing essential to ensure consistency. Context also plays a significant role in testing. The system’s behavior can change based on the page or scenario, such as the difference between summarizing an opportunity on an account page versus an opportunity page. It’s important for agents to adapt naturally to these varying contexts.
As you gain muscle memory planning, developing, testing and deploying Agent actions, you’ll find that the more complex use cases don’t seem quite so hard. Before you know it, you’ll be using GenAI-based tools to rapidly transform your Salesforce experience from an old-school point-and-click interface into something much more powerful: an agentic, chat-based platform with a streamlined, intelligent user experience.
Our AI experts can help
Interested in learning more about Agentforce, or want some guidance in getting started? We’d love to help. Atrium specializes in AI and Analytics solutions in CRM, and we’ve helped customers see substantial gains in productivity using AI-based tools to improve their business processes. Learn more about our data science and AI consulting services.
