Discover how companies are responsibly integrating AI in production. This invite-only event in SF will explore the intersection of technology and business. Find out how you can attend here.
It’s one thing to have a basic chatbot that can provide answers to simple questions, but it’s quite another to have a fully generative AI-powered platform that can actually perform actions.
Salesforce today is expanding the availability of its Einstein Copilot making the technology generally available. Alongside the release, Salesforce is also expanding the capabilities of the platform with new Einstein Copilot Actions that enable sales teams to be more productive with the power of gen AI. Einstein Copilot was first previewed by Salesforce in September 2023 during the company’s Dreamforce 2023 conference. In February of this year, Einstein Copilot reached beta availability enabling more users to try out the technology.
A core focus for Einstein Copilot is the ability to connect to an organization’s data, which is not just about data stored on the Salesforce platform. As part of the general availability rollout today, Salesforce is also announcing its Zero Copy Partner Network, which helps organizations connect to other data sources. The Zero Copy Partner Network supports vendor technologies that use the open-source Apache Iceberg table format for data lakes.
“Bringing the product into GA one of the things we learned is that the better the context, the more complete the context, the better Einstein Copilot works,” Jayesh Govindarajan SVP of Salesforce AI, told VentureBeat.
VB Event
The AI Impact Tour – San Francisco
Join us as we navigate the complexities of responsibly integrating AI in business at the next stop of VB’s AI Impact Tour in San Francisco. Don’t miss out on the chance to gain insights from industry experts, network with like-minded innovators, and explore the future of GenAI with customer experiences and optimize business processes.
Copilot Actions gives gen AI the power to actually do things
With Einstein Copilot, organizations get a conversational gen AI interface to ask about customer relationship management (CRM) data and connected data sources.
The basic idea of having a conversational interface at this point in 2024 is table stakes for gen AI. Where Salesforce is going above and beyond is with the deep context it can provide and perhaps more importantly, with the actions it can take. Instead of just being able to summarize some data or write some content, with Einstein Copilot Actions, organizations can trigger entire workflows to optimize a sales process and hopefully close more deals in the process.
Copilot Actions allows users to register any invocable action that can be executed by Einstein Copilot, both within and outside the Salesforce ecosystem. Einstein Copilot can also break down higher-order tasks into a series of actions and orchestrate their execution to complete the task. This includes actions like workflows, API calls and custom macros that users have registered with Einstein Copilot.
Govindarajan explained that the types of tasks that can be given to Einstein Copilot have a wide range that includes very specific and singular tasks to completely ambiguous and multi-step operations. Any task, be it ambiguous or specific can be triggered by a natural language prompt.
A single-step request could be something like asking Einstein Copilot to get a piece of data. A higher-order task would be a user asking Einstein Copilot to find the best sales opportunities to work on for a given day and provide a draft email for that prospect.
The higher order task is not just a simple retrieval augmented generation (RAG) gen AI style request either. Govindarajan explained that for it to work, the system needs to understand who the user is, and what an actual sales opportunity is in a given context and time. The system also needs to understand what the best opportunity is based on closing as well as value.
How Einstein Copilot reasons to enable enterprise workflows
To reason through higher-order tasks, Einstein Copilot uses several advanced AI techniques.
Govindarajan explained that Salesforce has done a lot of work developing planners that functionally teach Einstein Copilot how to reason. Among the techniques used is a sequential planner, which breaks down a task into a series of logical steps.
Salesforce is also making use of chain-of-thought reasoning as well as density-of-thought reasoning techniques. In these approaches, the gen AI system will go step by step to reason the optimal outcome based on a prompt.
Govindarajan said that for more ambiguous tasks Einstein Copilot uses a technique known as a reactive plan. For example, if a user is trying to determine the best sales opportunity to close, there may be a need for a series of questions to narrow and define the task better. With a reactive planner, the system reacts and asks follow-up questions.
Ever wonder how gen AI is performing? Einstein Copilot Analytics can help
Improving enterprise operations over time has long been the domain of data analytics. Now Salesforce is bringing that same type of discipline to gen AI with Copilot Analytics.
Copilot Analytics provides visibility into how organizations are using Einstein Copilot. It tracks interactions between users and Copilot, including higher-order tasks, conversations, how tasks are broken down, data grounding and the actions executed. This usage data is stored and can be customized and analyzed by customers. Some key metrics it tracks include which conversations ended positively versus not, which prompts were executed and their outcomes and where gaps exist where data or actions need improvement. Customers can use these insights to identify areas for customization or tuning prompts and models to make the Copilot experience better.
Looking forward, Govindarajan said that Salesforce is working on further improving Einstein Copilot in several ways including building new smaller and more efficient gen AI models.
“As this thing takes off, we fully expect from a performance and cost perspective, there are a lot of efficiencies that we can eke out as we bring smaller models to bear,” he said. ” We are doing this in the labs today and it’s showing great promise.”
