Einstein Copilot integrates conversational AI into Salesforce

Salesforce introduced Einstein in 2016 to provide predictive AI services across its products. Just months after ChatGPT, OpenAI introduced Einstein GPT to allow natural language software questions across the platform in March.

At the Dreamforce customer conference in San Francisco today, the company introduced Einstein Copilot, which embeds this ability to ask questions in the context of whatever users are doing, regardless of product.

Clara Shih, Salesforce AI CEO, says Einstein GPT was the first attempt to spread generative AI across the platform. It was designed to provide generalized automated assistance, like writing a customer service response or a salesperson email, but many customers wanted more, so they introduced Copilot.

“We are launching Einstein Copilot, a conversational AI assistant for companies, employees, and customers to securely and safely access generative AI to do their jobs better, faster, more easily, to augment and amplify their abilities, skills, work, efficiency, and productivity,” Shih.

This allows users to ask the bot in a conversational way for information that would normally require several clicks and knowledge. A salesperson could look at their pipeline, a new customer service rep could ask how to handle a 30-day return, and a product manager could ask how to add a new product or how it’s performing in a certain geography. If Einstein Copilot has been trained to answer the question, they can ask it in plain language and it will find the information.

CRM Essentials founder and principal analyst Brent Leary says that while most software companies are embedding generative AI, Salesforce has one advantage. Einstein Copilot’s coverage of all customer touchpoints, including commerce, may set it apart from other enterprise software. That creates an opportunity to impact multiple customer interactions and employee experiences while engaging customers in need, Leary said.

Shih is open about the problems with large language models. “We know AI trust is low. There are reasons for the gap…Island data can cause hallucinations and incorrect or incomplete outputs, she said this week at a press event.

By linking Einstein Copilot AI tooling to data from its Data Cloud (introduced last year at Dreamforce as Genie) and building its own model, Salesforce believes it can reduce some of the issues we’ve seen with large language models, particularly hallucinations where the model makes up an answer when it doesn’t have enough information.

The company also introduced a ‘trust layer,’ an underlying security, governance, and privacy architecture to give customers more confidence when using Salesforce generative AI tooling internally and externally.

However, large language models are believed to have no known way to eliminate hallucinations.

Einstein Copilot is in customer beta. Salesforce did not estimate a release date. Einstein Trust Layer will be available across the platform next month, according to the company.