The hype around generative AI since ChatGPT’s launch in November 2022 has driven some software vendors to rush to incorporate the technology into their applications. Despite being an early adopter of AI in general, Salesforce has taken a more measured approach to generative AI.
Following its announcement of Einstein GPT in March, the company has slowly added the generative AI functionality to a few of its applications. But behind the scenes, Salesforce has been concentrating on building a solid technical foundation for generative AI before rolling it out across the board.
“Getting the benefits of AI isn’t quite as simple as telling your employees they should just start using a generative AI bot, right?” said Clara Shih, now CEO of Salesforce AI, in a conference call on the eve of the company’s Dreamforce customer event.
Six months ago, Shih was GM of Salesforce Service Cloud and the company wasn’t ready to unleash generative AI on its customers. What’s changed since then, apart from Shih’s title, is Salesforce has rearchitected its underlying Data Cloud and Einstein AI framework to use an improved metadata framework, creating a new platform it calls Einstein 1.
That platform is the foundation for Salesforce’s next big move: readying a new generative AI interface, Einstein Copilot, that will eventually appear in all its applications. Some customers will get access to Einstein Copilot this fall, but the company hasn’t yet said when the new conversational interface will be generally available.
Einstein 1: Salesforce’s gen AI core
As with any AI, data is an essential ingredient for making generative AI work. To that end, Salesforce is leveraging Data Cloud as a central data hub for enterprise implementations of Einstein Copilot.
Data Cloud brings in enterprise data from Salesforce apps, data lakes, and warehouses, unifying it into one customer record for use across the Salesforce platform, Patrick Stokes, Salesforce’s EVP of product and industries marketing, explained in the same conference call.
“This means we now have a hyperscale data engine directly inside of Salesforce to connect all of your data,” he said. “It’s an open platform meant to support many different data providers, large language model providers, and independent software vendors.”
Although the goal of Einstein 1 is to make all of a company’s data accessible to its Salesforce applications, “the data doesn’t have to live in Salesforce,” he added.
For example, an enterprise with data in the Databricks Lakehouse Platform and in the lakehouse within Salesforce Data Cloud can now access both sets of data as if they were in a single location, with no need for ETL.
The Einstein 1 platform can support thousands of metadata-enabled objects, each with trillions of rows, and changes to any of the objects can trigger automation flows at rates up to 20,000 events per second, according to Salesforce.
The previously announced Einstein Trust Layer, seen in Salesforce AI Cloud, is also now part of Einstein 1. It serves to keep customer data within Salesforce by masking it from external large language models (LLMs); warning users of potentially toxic prompts or responses; and keeping an audit trail.
When Salesforce first announced Einstein GPT, its first iteration of generative AI technology, it only worked with its own LLMs or those from OpenAI. Now, Shih said, the company can access other models through AWS Bedrock or Google Vertex AI.
“We’ve really spent a lot of time building all of that foundation, so we can bring this into these customer applications,” Stokes said.
Access to the new Data Cloud is now included for all customers using Salesforce Enterprise Edition or above, allowing them to unify 10,000 customer profiles at no cost and to explore their data with two free Tableau Creator licenses, he said.
Einstein Copilot: Salesforce’s grand unified chatbot
After its initial tentative forays with Einstein GPT, adding limited generative AI features here and there in some of its applications, Salesforce is getting ready to go big with Einstein Copilot, a conversational AI assistant that will appear in the right-hand rail of every Salesforce application. Some of the new features are in pilot trials now, and others will be later this year.
“It’s going to be an open-ended assistant that employees and customers can simply chat with in natural language,” Shih said.
But more than that, Copilot will also be able to trigger specific Salesforce workflows.
Using Copilot Studio, enterprises will be able to control which workflows they want Copilot to have access to, linking to specific database fields they have in Data Cloud. They’ll also be able to customize the behind-the-scenes prompts that drive those workflows, personalizing them to their brand’s voice, Shih said.
Salesforce isn’t the first to come up with the idea of using generative AI to build a virtual coworker or copilot. What sets it apart, according to Stokes, is that, “Some competitors are going after productivity; we’re really going after the core customer workflows, sales, service, commerce, and marketing, where you have a direct interaction with your customer, because that’s our mission.”
Using Salesforce’s Copilot offers several advantages compared to using stand-alone generative AI tools, according to Shih. “Think about change management,” she said. “There’s no training required. You just talk to the assistant like a coworker, and the assistant will ask you if it needs clarification.” The alternative, having employees swivel to a different screen to use a generative bot, and then copy and paste, is not data secure, and is also a hurdle to adoption, she said.
Despite these advances, Einstein Copilot still won’t make you a coffee — but if your coffee machine has an API you can access from the Copilot Studio, you might be able to teach it.
Artificial Intelligence, Generative AI, Software Development, Technology Industry