Today at the Dreamforce customer conference in San Francisco, Salesforce announced Einstein Copilot Studio, a tool that lets customers customize the Salesforce base Einstein GPT and Einstein Copilot offerings.
Einstein Copilot Studio consists of three elements: prompt builder, skills builder and model builder, according to Clara Shih, CEO of Salesforce AI.
“The first piece is the prompt builder, and this is for customers who want to customize the prompt templates that have been included in Einstein GPT,” Shih told TechCrunch. That means customers can add their own custom prompts for their products or brand or to include items specific to their business or market that aren’t available out of the box with Einstein GPT or Einstein Copilot.
The skills builder enables companies to add actions to prompts. “There's also Skills Builder, which means that Einstein Copilot is now no longer just accessing your data and answering questions on the data. Companies also have the ability to control and designate which workflows they want Copilot to have access to and run," she said. Some examples include running a competitor analysis or objection handling, according to Shih.
The final piece lets you bring your own model, or use one of the supported third-party offerings like Anthropic, Cohere, Databricks, Google Cloud's Vertex AI and OpenAI. “These are for more of the enterprise companies that have large data science teams. They want to bring their own models,” she said. Shih says this is an update to the bring your own model idea introduced earlier this year.
Brent Leary, founder and principal analyst at CRM Essentials, says Studio fills a gap for companies who need more than the base tooling. “The power of Studio is it's the place where all the components are accessible from across the Salesforce platforms, which is a major part of simplifying prompt and skills creation for admins that address real-word use cases specific to organizational needs,” Leary said.
Salesforce is well aware of the problem with hallucinations when a model makes up an answer, and other issues like bias and inappropriate responses that can impact large language models. Shih says they are building a system they call "the Einstein Trust Layer" to help companies manage things like security, governance and data privacy.
In addition, by using the data in Data Cloud, introduced last year at Dreamforce as Genie, it can help ensure to a degree that it’s going to give accurate answers, based on information that's in Salesforce's databases.
Shih says customers can also tune prompts and skills in Studio, depending on the nature of the question, and on how much risk is at play, whether to fully automate a response or indicate that a human must be involved. “So if it's a simple question and answer like ‘how much do we charge for a wire transfer?’ -- first of all, I haven't seen it hallucinate on that -- but the stakes are lower than enabling Copilot to do something like helping transfer a large amount of money without talking to a person.” In a case like that, she said you would probably want to set up a rule to get a person involved before the transaction can take place.
It’s worth noting that it is widely believed that there is currently no known way to completely eliminate hallucinations in large language models.
While Shih admits that AI could eliminate some jobs, she sees new jobs being added too. "I think it's a big moment in time, and there will certainly be impact to certain jobs. And there will also certainly be new jobs that are being created, such as the prompt engineer," she said. The Studio product could be part of that.
Einstein Copilot Studio will be available in pilot sometime this fall. The Einstein Trust Layer will be generally available across the Einstein platform next month, according to the company. Salesforce did not get specific about the dates.