• Salesforce has expanded its AI tools with Agentforce, a platform to help users automate tasks.
• Agentforce uses generative AI models that build on existing Salesforce products.
• This article is part of “CXO AI Game Manual” – candid talks from business leaders about how they are testing and using AI.

For “CXO AI Playbook,” Business Insider examines mini-case studies of AI adoption across industries, company sizes, and technology DNA. We asked each of the featured companies to tell us about the problems they are trying to solve with AI, who makes these decisions internally, and their vision for using AI in the future.

Salesforce is best known for its customer relationship management software, which it says is used by more than 150,000 companies, including Amazon and Walmart. It also owns Slack, a popular communications platform.

The company is expanding the availability and capabilities of its AI agents with Agentforcea platform that helps Salesforce customers build and deploy agents to automate tasks like generating written reports from sales data and creating conversation summaries in Slack.

Situation Analysis: What problem were they trying to solve?

AI agents are not new to Salesforce. In 2016, the company freed Einsteinan AI service for its CRM software. Einstein handled some of the tasks that Agentforce can accomplish, but the actions it could take were more scripted.

With the arrival and rapid advancement of generative AI models, Salesforce saw an opportunity to improve agents’ decision-making and understanding of natural language input, like email.

This initially led to Einstein GPT, a generative AI redesign of Einstein. It then evolved into Einstein Copilot, announced in early 2024. It is now called Agentforce, which offers predefined and customizable agents.

“We’ve had to recognize that our customers want to either expand the number of agents we ship or create their own customer agents,” Tyler Carlson, Salesforce’s vice president of business development, told Business Insider.

Key staff and partners

Agentforce’s core technology, which Salesforce calls Atlas Reasoning Engine, was developed by the company’s engineering team. Salesforce also allows its customers to access AI models from third-party technology partners, including OpenAI, Anthropic, Amazon and Google.

Salesforce worked with internal and external partners to showcase its AI agents. The company uses Slack to highlight AI agents accessible directly from the communication platform. (Currently, the Agentforce integration in Slack is in beta.)

“The reason Slack is interesting for these employee-facing agents is because it presents these automations and capabilities to users where they spend their time,” Carlson said.

AI in action

Salesforce claims that the Atlas reasoning engine uses what is called Prompt React to help agents reason about problems.

ReAct helps an AI agent powered by a large language model break down a problem into its component parts and then provide steps to solve it. After each action, the AI ​​agent evaluates the result to judge whether it satisfied the user’s prompt. Otherwise, the agent goes through different stages of reasoning and action until reaching a satisfactory result.

Salesforce says this can lead to fully autonomous agents that offer personalized assistance based on a user’s specific request and take actions, like scheduling an appointment or meeting, without human intervention.

The concept is similar to OpenAI’s new o1 model, which uses the chain of thought prompting technique to follow the prompts step by step.

The Atlas Reasoning Engine is a proprietary technology designed by Salesforce and has access to LLMs developed in-house by the company. But it is also designed to work with LLMs provided by partners: the company expects many agents to take advantage of this functionality.

“That’s what I find most interesting about agentic systems,” Carlson said. “They are not monolithic. It is not a single LLM.”

Allowing access to third-party LLMs can be risky because it means transmit data to an external provider. To address this issue, Salesforce instituted a policy prohibiting data retention. It also provides safeguards designed to identify and reject inappropriate responses to prompts.

Salesforce customers can also use a tool called Agentbuilder to create and deploy custom Agentforce agents that can act on triggers, such as receiving an email or making a sale, without direct prompting from employees. ‘a business. An automated agent can open messages from a general inbox and forward them to specific departments based on the agent’s understanding of the message’s intent.

Agentforce also uses generation augmented by recovery to create agents that can answer questions based on a company’s internal documents and private data. For example, this year, Salesforce and Workday created an “AI employee service agent” designed to answer employees’ questions about a company’s HR policies.

Did it work and how did they know?

Carlson said Salesforce measures the success of its software by the success of its customers, adding that the results so far have been positive. “This includes promising statistics, like seeing 90% of customer inquiries resolved by an agent,” he said.

Salesforce will also judge its success by adopting Agentforce as a platform. The company wants to see its customers widely adopt AI agents built with Agentbuilder and deploy them in their CRM systems and Slack. He also wants to see new AI agents take on larger workloads and resolve interactions to satisfy customers and customers.

“By next year, I want to see a larger, more open partner ecosystem,” Carlson said. “I want Agentforce to have thousands and thousands of agent skills, topics and more for our customers to take advantage of.”