The Salesforce platform evolves significantly every single year. With 3 scheduled releases per year (Winter, Spring and Summer), new product or feature introductions and deprecation of legacy elements is certainly expected. This is good for the platform, and good for the overall ecosystem as it helps Salesforce keep pace with the rest of the technology world.
However, this also brings with it a lot of hand wringing and head scratching for the partners and consultants in the ecosystem. Let me explain. Salesforce, in the humble opinion of the author of this blog post, has been playing a game of musical chairs with the branding of its products over the last few years. Case in point, did you know that the product Wave was rebranded as Einstein Analytics, which was then rebranded as Tableau CRM, which was then finally (and hopefully) renamed to the current brand name of CRM Analytics? What about the CDP product Genie that launched with much fanfare in late 2022? That is called Data Cloud now.
In a similar but not so similar category, there are product brand names like ‘Einstein’ that have evolved to encompass an entire suite of products. It is indeed very tough to list out all the products that are part of the Einstein umbrella in one shot.
A bit of history on Einstein. The Einstein brand name first appeared in 2016 as an AI feature that was integrated into the core CRM apps like Sales, Service, Commerce and Marketing to deliver predictive capabilities for activities like lead scoring, case classification and product recommendations. It is important to note that these capabilities were leveraging machine learning models to predict outputs.
As of pre-Dreamforce 2024, the brand name of Einstein represented features and tooling that leveraged both the predictive and generative mechanisms of Artificial Intelligence to generate outputs.
Although a complete analysis of the differences between both the noted sets of AI mechanisms is out of the scope of this blog post, a deeply simplified explanation is that:
- Predictive AI represents techniques employed by a well trained model that identify patterns from a large volume of existing datasets and predict the most probable output for a given set of inputs. The outputs can be either numeric or non-numeric like classifications. For instance, classifying a case based on similar case patterns from the past.
- Generative AI is also powered by a model (mostly very complex neural networks as of now) that has been trained on possibly millions of data points to generate original and new content on the basis of provided inputs. Combine this ability with a large language model (LLM) that uses natural language processing (NLP) and you have a powerful combination that can generate new contextual text, images, audio and even video by providing regular English language sentences as inputs. The GPT-4 model that powers ChatGPT is a well known example.
Post Dreamforce 2024, it is clear that Salesforce has decided to draw some branding distinction between legacy Einstein products that leverage predictive capabilities (like Einstein Discovery) and the latest ones that draw upon generative AI models (like Einstein Copilot).
In other words, so long Einstein Copilot. Welcome, Agentforce.
What is Agentforce?
Short answer – A powerful chatbot that leverages LLMs and relevant data retrieved from CRM or other sources to:
- produce insightful responses to user utterances.
- use reasoning like a human to take actions that automate business processes.
It is important to note that Agentforce is not the same as Einstein Bots. Einstein Bots do leverage some AI but they do not have the reasoning engine or the ability to take actions that Agentforce possesses. (Are you confused yet?)
There can be multiple ‘Agents’ in existence, although as of Sep 2024, there can only be two types of agents: Salesforce and Digital Channel.
A ‘Salesforce’ type agent is internal facing and can be configured to interact with internal Salesforce users to give them relevant answers to their questions or take actions on their behalf. There can only be one such type of agent.
For example, if a business user wants to know all the ‘Open’ opportunities that are assigned to John Smith, they can just type that as-is into the conversational panel and get a response. Additionally, if they are viewing one of those open Opportunities and would like to change the Amount to $15k and set the Stage to Closing Process, a text command will do.
In contrast, a ‘Digital Channel’ agent is external facing and can be configured to be exposed on Experience Cloud sites. Applicable example scenarios include helping customers deal with initiating product returns, changing shipping addresses on existing returns, providing accurate and contextual answers to customer issues and so on.
There are a multitude of use cases that can be built for internal and external users, not just in traditional Sales and Service environments but across other industry verticals.
How does Agentforce work?
The Agentforce workflow stack comprises five main ingredients: Agents, Reasoning Engine, Topics and Agent Actions, LLM.
We’ve already covered Agents in the above section. Let us take a look at the others.
An agent action is a no-code configuration that gives an agent the ability to perform tasks in Salesforce, like creating or modifying data and executing Flows or Apex. Agent actions can either be Standard or Custom. Any agent action relies on three configuration parameters:
- Inputs
- Outputs
- Action Instructions
Typically, custom Agent Actions can be powered by a Prompt Template, Flow or Apex. In other words, agents can take direct actions to invoke a Flow or Apex, or indirectly use Prompt Templates to invoke them.
A quick sidebar on a Prompt Template. As the name suggests, it is a reusable set of Natural Language instructions that can be given as input to any LLM. The benefit of using a Prompt Template in Salesforce is that it merges those reusable instructions with contextual data from the CRM or any external sources. LLMs are at their most efficient when enough data is provided along with detailed instructions on how to process that data, and Prompt Templates help us achieve that.
Agents are able to select and invoke specific actions through reasoning by evaluating the user utterance with the Action Instructions noted on each configured Agent Action. This is facilitated by the Planner Service or Atlas Reasoning Engine. Fun fact: The reasoning engine itself relies on a LLM internally to ‘reason’ and select actions!
A Topic is just an aggregation of Agent Actions put together that can be assigned to an agent, along with instructions on when to invoke it. For example, if an agent is helping customers subscribe to new products and also process their returns, it can be assigned two distinct Topics called ‘Subscription Management’ and ‘Return Orders’ with their set of associated Agent Actions grouped underneath each of them respectively.
Here’s a simple workflow to put all of this together: After a user posts an utterance, the Reasoning Engine understands the text and looks for the right Topic to be invoked. After selecting the right Topic, it further determines the appropriate Agent Action within that fits the user’s requirement. It then invokes the Agent Action and that subsequently calls a Flow, or Apex or a Prompt Template, depending on what is powering it. If a Prompt Template is invoked, it merges the relevant data and instructions and sends it over to a LLM to get an insightful response.
Note: In the above scenario, if an Agent Action invokes a Flow or Apex, those in turn can call a Prompt Template and therefore a LLM. As you can see, the combinations are numerous!
So, why do I need Agentforce?
Over the last couple of years, LLMs have upended the conventional perception of where AI can fit into a company’s needs. With powerful LLMs powering it and its own ability to reason and take actions within Salesforce, Agentforce can help companies derive a lot of value by eliminating manual work and doubling up as a knowledgeable and handy assistant.
As noted earlier, the use cases can range from simple information data retrieval tasks to complex automation processes. This article did not discuss Data Cloud in depth, but combining RAG (Retrieval Augmented Generation) with Agentforce’s conversational capability is an excellent choice for large enterprises that rely on disparate lakes of siloed data.
It is definitely worth exploring if Agentforce can fit into a company’s workflow. The rapid pace at which this space is evolving poses an enormous challenge not in terms of complexity with implementation but of identifying the right use cases that provide real productivity. Or put differently, Salesforce is trying to make no-code AI a popular thing, but we will reap benefits only if we use it for the appropriate reasons.
References:
https://www.salesforce.com/news/press-releases/2016/09/19/salesforce-introduces-salesforce-einstein-artificial-intelligence-for-everyone/
https://www.researchgate.net/post/What_are_predictive_AI_models#:~:text=Data%20Input%3A%20Predictive%20AI%20models,e.g.%2C%20text%20or%20images
https://www.algolia.com/blog/ai/large-language-models-llms-vs-generative-ai-whats-the-difference/
https://www.salesforce.com/blog/what-is-retrieval-augmented-generation/