Salesforce adds skills to its AI agents and agentic platform to serve more enterprise use cases

Customer relationship management (CRM) software provider Salesforce has updated its agentic AI platform, Agentforce, to make it easier for enterprises to build more efficient agents faster and deploy them across a variety of systems or workflows.

Christened Agentforce 2.0, the second release of the agentic AI platform, which comes just two months after the first version was released, gets new features and capabilities, such as the option to switch to an updated reasoning engine, new agent skills, and the ability to build agents using natural language.

For the uninitiated, agent skills are actions or functions that a particular agent can take on behalf of a human worker, without any user intervention.

The new library of pre-built agent skills and Agentforce’s integration with MuleSoft will help enterprises cut down on the time and complexity of building a new custom agent and integrating it into a workflow, said Ritu Jyoti, general manager of research at IDC.

And the ability to build agents using natural language will extend the low code suite’s usability to a variety of users within an enterprise, Jyoti added.

New agent skills in Agentforce 2.0

As part of the new release, Salesforce said that it was adding new agent skills, such as Sales Development and Sales Coaching, for use cases related to sales. These skills join a fleet of existing skills such as lead development and personal shopper.

The new skills are expected to help enterprises nurture leads based on an enterprise sales team’s preferences, join calls with prospects, and provide instant feedback, the company said.

Other skills added for marketing, commerce, and service use cases include Marketing Campaign and Commerce Merchant skills, scheduling skills for service engagements, and new skills for field service workers.

Additionally, Adam Evans, senior vice president of product at Salesforce’s AI division, pointed out in an interview with CIO.com that these skills can be “remixed” to suit any use case.

Explaining further, Evans gave the example of the lead development skill being adapted to suit a recruitment workflow, wherein the lead development skill could be used to shortlist candidates to fill a vacant position.

However, Salesforce isn’t the only agentic AI provider that is taking the approach of launching basic agents which could be tweaked to suit a variety of use cases.

Microsoft’s corporate vice president Bryan Goode, who leads products such as Copilot Studio and Dynamics 365, told CIO.com during the launch of its AI agents that it was releasing 10 pre-built agents that would act as templates for enterprises to help them develop agents for a variety of use cases.

Salesforce, and rivals such as Google, Microsoft, AWS, and IBM, are also partnering with other software vendors, such as Workday, DocuSign, and Neuron 7, to create more agents that can be accessed via their marketplaces.

The idea is that these partner-built agents would act as templates, and enterprises need not go through the process of creating a new agent from scratch, Evans explained.

Building agents with natural language

With the latest release of Agentforce, Salesforce has added the ability for enterprise users to build agents using natural language.

“Once users describe the kind of agent they want to build inside Agent Builder, the tool automatically generates relevant topics and instructions to help build the agent,” Silvio Savarese, chief scientist at Salesforce, told CIO.com.

Agent Builder can also pull from the library of skills and actions already available to that enterprise to make the development activity faster, Savarese added.

However, the ability to develop an agent using natural language is not unique to Salesforce; rivals such as AWS, Google, Microsoft, and IBM offer the same capability in AWS Bedrock via Bedrock Agents, in Google Generative AI Studio via Vertex AI Agent Builder, in Microsoft Copilot Studio, and in IBM watsonx respectively.

But what’s new, according to Amalgam Insights’ chief analyst Hyoun Park, is Agent Builder’s ability to suggest agent topics and instructions.

“This ability builds on the deep metadata context that Salesforce has across a variety of tasks. Salesforce’s ability to recommend task-based agents rather than insisting on having users technically define the nature of the agent they are trying to create is a democratizing trait for agent creation,” Park explained.

However, The Futurum Group’s principal analyst Dion Hinchcliffe feels that, though this ability might be compelling, many enterprises won’t open up to such features right away, until they determine it is safe to do so.

Expanding further, Moor Strategy and Insights’ principal analyst Jason Andersen pointed out that this ability might not be enough to get an enterprise to switch to Agentforce from another platform.

Agentforce 2.0 gets an upgraded Atlas reasoning engine

The second release of Agentforce also gets an upgraded version of the Atlas reasoning engine, the module that Salesforce describes as the brain behind the product.

The upgrade, which includes enhanced reasoning and retrieval capabilities, will help enterprises answer more complex queries via their agents so that the burden on the contact center personnel is reduced, Salesforce CEO Marc Benioff explained to reporters during the launch.

Benioff gave the example of help.salesforce.com, the site that handles Salesforce customer enquiries, and said that after its agents were connected to the upgraded Atlas engine, nearly 80% of customer queries were being answered by AI agents, compared to approximately 50% when the first iteration of Agentforce was released.

In its efforts to illustrate the difference in reasoning capabilities between the two versions of Atlas, Salesforce gave the example of a financial advisor AI agent answering what it described as a complex question: Does an user’s investment portfolio have any hidden risks due to rising interest rates on fixed-income and interest-sensitive assets?

While the first iteration of Atlas would just be able to let the user know the details about their financial portfolio, the second version is able to break down the “complex” question and answer it within the boundaries of its defined knowledge bases, along with citations, Benioff explained.

But, he added, “Atlas still hands any query to a human agent if it cannot answer it, or reasons that it is beyond its expertise.”

However, Evans pointed out that not every AI agent needs to be connected to the upgraded Atlas reasoning engine, and enterprises will have the choice to toggle between the versions of Atlas.

He also warned that there is a tradeoff, as the second version, while being qualitatively better in reasoning, is slower, and might be better suited for use cases where the user doesn’t mind waiting a few seconds.

Some examples of such use cases, according to Evans, are answering questions on contracts or large documents, especially in the legal, insurance, and healthcare sectors.

MuleSoft, Tableau, and Slack integration in Agentforce

Salesforce has also integrated Agentforce with MuleSoft, Tableau, and Slack to make deploying it easier throughout the enterprise.

The MuleSoft integration, which uses Flow, will help users to create low-code workflows that span systems, with pre-built connectors for building multi-system workflows, IDC’s Jyoti said.

“This capability allows APIs to be turned into Agentforce actions and, via the new MuleSoft API Catalog, those APIs are made discoverable,” Jyoti explained, adding that the MuleSoft Topic Center expands on this by enabling teams to infuse Agentforce metadata into every API they build.

On the data visualization front, Agentforce’s integration with Tableau comes in the form of new Tableau Topics and Actions that will help deliver data visualizations and predictions for a deeper understanding of agent responses, and accurate, business context-rich answers using Tableau Semantics.

This ability, according to the company, will help unlock new conversational analytics use cases.

In addition, Salesforce has added Slack Actions to Agent Builder workflows in Agentforce. This will help a user or team to use agents to get updates on a project or take actions, such as joining a meeting on the user’s behalf or asking a customer for feedback, Salesforce said, adding that agents can be added to any conversation within Slack.

The agents added to Slack via Agentforce can also take advantage of Enterprise Search within Slack to draw from conversational data to enhance the relevancy of responses and actions, according to Evans.

Pricing and availability

Salesforce top executives said that there is currently no price hike for the use of Agentforce 2.0.

The full release will be generally available in February 2025, with some features coming sooner.

To begin with, Agentforce in Slack, Slack Actions in Agent Builder, Slack Enterprise Search, and natural language creation of agents in Agent Builder will be generally available in January.

Features and capabilities such as MuleSoft for Flow, MuleSoft API Catalog, Topic Center, Atlas enhanced reasoning, and RAG capabilities will be generally available in February.

Other skills, such as Sales Development and Sales Coaching, have already been made generally available, and pricing for these start at $2 per conversation.

The company expects to launch, or at least showcase, the third release of Agentforce in May.