Salesforce Inc. today rolled out an artificial intelligence conversational copilot assistant for its applications that can assist users in their everyday work alongside a studio that will allow companies to build their own prompts, actions and models.

The new generative AI assistant, dubbed Einstein Copilot, integrates directly into Salesforce applications and understands the context of what users are looking at, providing them the ability to “chat” with their corporate data. It works similarly to chatbots such as OpenAI LP’s ChatGPT enabling users to ask questions conversationally, receive answers and have the AI perform actions.

Coming in a pilot this fall, the Copilot is part of a larger trend of Salesforce embracing generative AI in its applications, including Einstein GPT, which the company included in its Sales Cloud, Service Cloud and Data Cloud offerings as a bot integration in a closed pilot earlier this year. The new Copilot, announced today at the company’s Dreamforce conference in San Francisco, goes far beyond simply composing emails and meeting notes to directly integrate into the user experience so that the AI can work beside them.

From the point of view of the user, the copilot should just become another tool, Muralidhar Krishnaprasad, executive vice president of software engineering at Salesforce, told SiliconANGLE in an interview. It acts to augment their work and make it less complex by helping them get knowledge work done faster and take actions on their behalf.

“Most of our users probably don’t know what AI even means for them,” said Krishnaprasad. “It’s kind of like when you drive a car, you just want that lane change thing or the blind spot monitoring tool to just work. And so, we have embedded AI into their line of work.”

For example, a customer service worker can receive information while on a call to help reduce call time because it has access to remediation data, or sales workers can get quick information on the best discounts to move old merchandise faster in readable format. Marketers can get help generating their segments, analyze campaigns to produce engaging content right out of their own data.

Industry professionals can also get direct assistance from Copilot. Finance advisers could analyze client spending and savings histories in order to offer personalized coaching and create tailored savings plans. A healthcare professional could reduce patient no-shows through historical analysis and generate appointment notes summaries, and a college administrator could use the AI assistant to provide curriculum notes based on individual student skills.

Customizing the Copilot with Copilot Studio

Administrators and managers can customize Einstein Copilot using Copilot Studio and modify how the assistant responds to users such as employees and customers to build their own apps with their own prompts and actions, which Salesforce calls “skills.”

Prompts are how a generative AI retrieves information from the vast store of data that it is trained on and generates new information when queried. The Prompt Builder allows users to build, test and make new queries that match the company’s brand and communication style in a way that requires very little expertise. For example, a marketer could use it to make a personalized message about a discount based on a customer’s location and purchase history and then deploy the same prompt for any customer in the future.

The Skills Builder allows companies to make workflow automations for creating AI-driven actions to complete tasks. These are fundamental for AI applications where the AI is going to do more than simply respond to questions and answers and may be needed to complete some sort of activity, such as changing a password or initiating a return for a customer. With skills, an AI agent can be hooked into external application programming interfaces, make calls to databases, initiate alerts and more.

For developers and power users, companies can also build and select their own AI models from whatever provider they like to fit their own business needs. The Model Builder gives them the flexibility to do that, allowing them to choose a proprietary large language model from Salesforce, or integrate a generative AI model that can be trained and fine-tuned from data in the Data Cloud without moving or copying data. Using the Model Builder, Einstein Copilot can support a “bring your own model” sort of integration with LLMs from Anthropic, Cohere Inc., Databricks Inc., Google LLC’s Vertex AI, Amazon Web Services Inc.’s SageMaker and others.

“If you’re a power user, you can go use your tools of your choice,” explained Krishnaprasad. “We have the power to bring all the data together, whether it’s through copy or reference. We have the power to then build any model that you want.”

Finally, everything is built on top of Einstein Trust Layer, a framework for AI security that is built natively into the Einstein 1 Platform that provides privacy and protection for all data that is passed through the system. It does this with a series of steps that include masking sensitive personally identifiable information and protecting the privacy of individuals and making certain that no data is retained by the AI models that it is passed through so that it will not be used for training. This prevents proprietary enterprise information from potentially being leaked.

The Trust Layer also attends to toxicity detection, meaning that it protects against bias and other potential issues. If a user prompts the AI to bring up a list of customers that might include a particular problematic bias, for example a racial bias, the system will also warn the user from the outset just in case before proceeding.

AI enhanced with Einstein 1 Platform with Data Cloud

All of this is made possible with the newly announced Einstein 1 Platform, which includes major improvements to Data Cloud that gives companies the ability to safely connect their AI-powered experiences to their enterprise data and build apps using low-code interfaces.

“A company’s AI strategy is only as good as its data strategy,” said Parker Harris, co-founder and chief technology officer of Salesforce. “We pioneered the metadata framework nearly 25 years ago to seamlessly bridge data across applications. It’s the connective tissue that fuels innovation.”

Data Cloud brings together all of a company’s metadata-enabled objects into one place at scale, allowing them to be accessed by the AI assistant safely and acted upon using a hyperscale engine. This reduces the total amount of fragmentation across enterprise data stacks between cloud, social and mobile and provides access to data all in one place. It’s the equivalent of having one language for accessing and processing data in a unified view for applications and also provides discovery, automation and analytics at scale.

Part of that analytics is how the entire platform is being used, such as feedback on the prompts and skills used in Einstein Copilot. That data closes the loop by providing valuable information about how what built is being used by the people it was designed for so that it can be refined.

“We also make sure we record all that feedback,” said Krishnaprasad. “Is your AI really helping your sales agents or your service agents? Let’s assume you have created a skill or a prompt for your salespeople to use. But maybe what you did was so bad that none of your sales have actually used it. So that’s valuable feedback.”

Users of the Einstein 1 Platform will receive automatic upgrades three times per year, and the metadata framework prevents any integrations or security models from breaking. Organizations can further customize it by adding, extending and building on top of it as Salesforce evolves.

Image: Salesforce