Today, Salesforce unveiled Agentforce, a suite of autonomous AI agents that augment employees and handle tasks in service, sales, marketing, and commerce. The decade old Einstein platform, and the Einstein Copilot offering, is now Agentforce.
“The Einstein platform has been amazing, and for the last decade, over and over again, it's been iterated on and created,” said Marc Benioff, Chair and CEO, Salesforce, in a prebriefing. “Today, it's doing 2 trillion AI transactions a week. Those are predictive transactions and generative transactions.”
The focus of Agentforce, however, is to provide Salesforce customers with a platform approach to AI. Benioff said that what they will debut at Dreamforce next week is not an LLM, not RAG, not a technique, not a data set.
“It’s none of those things. And then it's all of those things, because at the fundamental level we will demonstrate that humans and AI agents will drive customer success together in a trusted way,” Benioff said. “It's going to be easy for them, it's going to be very low cost, it's going to be low code, and it's going to be no code, but mostly it's going to be a platform. And it's going to be a platform they already have and are already familiar with. There's nothing new to install. They'll be able to take their existing Salesforce team and have a new experience.”
The following graphic illustrates the four announcements: Agentforce Agents, the Atlas Reasoning Engine, Agent Builder and the partner network. Each of the Agentforce agents are out-of-the-box and ready for customers to turn on and configure.
The Atlas Reasoning Engine is the brain of the agent – it generates a plan of action, evaluates it, refines it and then automates the actions to achieve the goals of each Agent (services, sales, etc.). Agent Builder is the newest low code builder and it is part of the overall Agentforce Studio, which includes Prompt Builder and Model Builder with which customers can train, tune or bring their own models. The Partner network is a consortium of software companies who are coming together to build agents that adhere to the same ethical guardrails.
Overall, Salesforce said that its new AI agents can analyze data, make decisions, and perform tasks like answering customer service inquiries, qualifying sales leads, and optimizing marketing campaigns, and that customers will be able to use Agent Builder to build, customize and deploy agents along their own role-dependent task flows.
In a prebriefing, Clara Shih, CEO of Salesforce AI said that there are five attributes of an Agentforce Agent: Role, Data, Actions, Guardrails and Channel (illustrated below). “This sounds like a lot,” Shih said, “But the good news is Salesforce customers already have all of these pieces in Salesforce.”
Shih continued with several examples:
So, if an Agent is assigned the role of “salesperson” then the Atlas engine will generate a plan to build a pipeline and close deals. Atlas is built on a proprietary system designed to simulate how humans think and plan. It starts by evaluating user queries, refining them for clarity and relevance. Next, it retrieves the most relevant data and builds a plan for execution. The process then refines the plan further, ensuring it’s accurate, relevant, and grounded in trusted data. This process allows Agentforce to autonomously reason, make decisions, and complete business tasks, all while delivering precise, factually accurate results.
“You don't want the agent making up their own actions. It's looking at what [Salesforce] Flows, what Apex code, what API calls from MuleSoft are allowed, what's been proven to be effective for this organization and all the teams that have been using Salesforce for the last 25 years,” Shih said. “That's where it focuses its action.”
Salesforce said that in terms of troubleshooting the Agents, it is possible to get the Agent to explain why it made a given decision. Customers will also be able to add confidence values to actions. This would allow an organization to escalate to a human supervisor if a certain task is being assessed with low confidence by the Agent.
“That's why so many customers are coming to us for Agentforce. You need to be able to seamlessly hand off to a human whenever the confidence score drops,” Shih said. “What we're seeing very conservative regulated industries do is they'll often start with our existing human in the loop use cases and once they see good results from that, those workflows will graduate to the customer self service agent.”
Agentforce Agents are not chatbots or copilots. This Salesforce article identifies multiple types of chatbots, from rules-based to AI-powered chatbots that leverage natural language processing (NLP), machine learning (ML) and generative AI. Copilots are the generative AI-powered “assistants” that provide everything from assistance writing emails, to summarizing documents and customer service calls.
What differentiates Salesforce’s Agentforce is its integrated, platform approach which includes Data Cloud. Data Cloud unifies and harmonizes all customer data and metadata across systems in real time, enabling Agentforce to operate with complete context and precision. Data Cloud improves the investment many organizations have already made in data lakes by providing a “Zero Copy” capability, which enables Salesforce customers to connect structured and unstructured data from external systems without having to copy it. This makes enterprise data retrievable by Agentforce to analyze and act on.
The five attributes Shih listed are then layered on top so that the Agents can operate autonomously, retrieve the right data on demand, build action plans for any task, and execute on these plans without requiring human intervention. When desired, Agentforce seamlessly hands off to human employees with a summary of the interaction, an overview of the customer’s details, and recommendations for what to do next.
This past July, Salesforce launched Agentforce Service Agent to help customers access self-service portals and messaging channels 24/7; in August, the company announced two AI-driven sales agents: Agentforce Sales Development Rep (SDR), which can autonomously engage with inbound leads, and Agentforce Sales Coach, which allows sellers to role-play different sales scenarios with them to uplevel their skills.
This latest announcement, Benioff stressed in the briefing, demonstrates how Agentforce is a platform in which AI has been incorporated, not one in which AI copilots have been “bolted on” to already-existing CRM or customer service platforms.
The company addressed security and confidentiality concerns by saying that Agentforce relies on the Einstein Trust Layer (a set of protocols and features aimed at siloing data and maintaining access and permission controls for different roles within a customer organization) to define the parameters and boundaries for an Agentforce deployment.
“This is very different from consumer AI in the enterprise,” Shih said. “You have data security, data privacy, company policies, different people have access to different data, and different workflows. Salesforce allows companies to easily specify those guardrails — what's in or out of scope for each different type of agent, and then to be able to easily monitor and audit that the agent is behaving as expected."
Salesforce will be offering a number of out-of-the-box agents for customers. The company described them below:
In addition to the pre-designed agents, Salesforce also unveiled three tools for helping customers create and optimize their own agents. The company described these as:
The company said Agentforce for Service and Sales will be generally available on October 25, 2024, and pricing starts at $2 per conversation; standard volume discounts will apply. In a post-briefing conversation, president and chief product officer David Schmaier said the company will be offering activation to its customers at Dreamforce next week.
