Generative AI has fundamentally changed the way professionals across all industries work, and our ecosystem is no exception. Salesforce has been investing in AI capabilities for over a decade already, with the focus now on infusing genAI functionalities in their entire product suite – Service Cloud included – to benefit both admins and users alike.
Let’s take a deep dive into Service Replies changes, how the prompts can be edited, and what making use of the new features entails.
Supercharge Agent Productivity With Service Replies
First of all, if you haven’t already used this feature, you’re in for a treat! Einstein Service Replies became generally available in Winter ’24 to support service agents in their customer conversations with contextual response suggestions offered in real-time – powered by Einstein. Even though the experience is slightly different, Service Replies are also available for emails, (albeit not in real-time due to the nature of the channel).
For chat conversation though, the generative AI replies can be displayed to agents in a separate component embedded on the Case Record Page, with a new set of recommended responses being generated after each end user utterance.
For each reply, the agent has the option to post it directly into the chat window, tweak it, and mark it as not helpful – this way, your agents will always have a starting point for their replies, reducing the time it takes for the end customer to receive their response. Customers can monitor adoption by using this information, as it is logged as part of the audit and feedback data stored in Data Cloud.
While contextual replies are quite helpful, what if you could make them more specific to your org? Through the use of Service AI Grounding, responses can be based on information from your existing Knowledge Base. This has been a good starting point for grounding Service Replies, but truth be told, many Salesforce organizations need to leverage much more information from other related records.
They also need better control and a better format for the recommendations their agents receive to maximize the functionality as much as possible and truly make it their own – which is exactly what Salesforce has in store for Service Replies!
What Is Changing?
Service Replies are made available to agents through a standard Lightning Web Component (LWC); an embedded app that leverages a prompt in the backend to provide the actual recommendations based on the conversation transcript.
The challenge with this approach was that customers had no access to the prompt nor a way to suitably modify it– in practical terms, they could only leverage Service AI Grounding, while Salesforce was doing the rest in a standardized manner.
Based on this customer feedback, Salesforce took swift action, switching the hard-coded prompt to readily available Prompt Templates. You and your team can now freely modify the tone and format of the output, but more importantly, use the LiveChatTranscript or MessagingSession as resources within Prompt Builder to access fields and related records – more on that below!
Another significant change is the use of Einstein Data Libraries to allow for richer and more relevant replies by grounding them in both structured and unstructured data ingested by Data Cloud. This approach will also make the indexing much smoother for the Knowledge articles, fields, or actual documents you’ll be uploading.
Einstein Data Library Setup
To benefit from the new Service Replies updates, the first step is to enable Data Cloud, as they are built on the Data Cloud Vector Database for companies to get the most out of their unstructured data e.g. conversations, articles, or files.
Through the updated Einstein Reply Recommendations setup page, you can easily navigate to the Data Library Setup and choose an existing Einstein Data Library (EDL). Alternatively, you could create a new one that’s suitable for the Service Replies grounding, as the page follows the guided experience you’re most likely familiar with.
Since Einstein Data Libraries are also used to ground the Agentforce Service Agent, companies using both products can leverage the same EDL if they want to. Admins configuring Service Replies will need admin access to Data Cloud to manage the EDL retrievers.
Keep in mind that you can always return and modify the EDL if needed. But to make sure the replies use all the information as intended, you should thoroughly test the functionality in a Data Cloud sandbox. However, switching from using Service AI Grounding to the new EDL-powered experience will prompt the search index to be moved to Data Cloud – this change is irreversible.
The Prompt Builder Experience
By far the most exciting part for Einstein Service Replies customers is the ability to modify, experiment, and version the four new out-of-the-box Prompt Templates – this will fully transform their Service Agents’ experience while chatting with the customers. As you will notice below, there are two types of templates: Contextual and Grounded.
The Contextual prompts are based on the conversation and will generate a search query that is passed to the Grounded template. Technically, both types can be grounded in customer data, but it’s recommended that you leverage simple related record fields – for example, in the Contextual template – while the more complex resources should be used in the Grounded template.
By default, within the Service Replies embedded app, agents will receive one contextual response and up to two grounded ones. Below is an example of instructions for a Contextual Prompt Template, using basic information from the related Contact:
Assist the agent in an ongoing customer service conversation provided in the CONVERSATION TRANSCRIPT section below.
If the customer’s words may be considered unethical and inappropriate, you must not repeat the inappropriate words in the response.
The customer is {!$Input:MessagingSession.EndUserContact.Name}.
###
Here are the instructions to follow to generate the output:
1. The response must naturally fit into the flow of the conversation.
2. The response must be concise and friendly. Never say “sorry” or apologize to the customer, but sympathize with their frustration.
3. The response must not repeat things the Agent has already said.
4. Always offer to call the customer at their phone number: {!$Input:conversation_object.EndUserContact.MobilePhone} to discuss further.
5. Reference past cases that the customer has opened with us with the case number: {!$RelatedList:MessagingSession.EndUserContact.Cases.Records}. Make sure to tell the customer their case number if there is a previous case that is relevant.
Provide exactly one contextual conversational response.
###
CONVERSATION TRANSCRIPT:
{!$Input:conversation_transcript}
Within Prompt Builder, you will be able to completely transform the readily available instructions if you wish to – or simply enhance them with additional guidelines of your own – as well as additional resources. The main benefit of being able to access these prompts now, as you saw above, is the ability to ground them in any information from or related to the LiveChatTranscript or MessagingSession. While initially, these were accessible through a Conversation object, Salesforce made the decision to simplify the experience by directly allowing you to drill down into the record or related information you need from either the Messaging Session or Chat Transcript based on the Prompt Template you selected.
Note: If you have already overwritten the standard templates and used the Conversation Object prior to the January update, it may cause an error now. If you are impacted, check out this Known Issue, which also contains the workaround.
By making use of one of these two objects within the Prompt Template Workspace when laying out your instructions and data, you can get to the related Account, Case, and even Opportunity for MessagingSession, or any other available field you want to make use of.
In addition, as with any other Prompt Templates you may have created in the past, you will be able to leverage Salesforce Flows and Apex. Furthermore, you can choose the model that produces the best outcome based on the test results you will be able to review directly in Prompt Builder. You will also notice that there are ten language options available to choose from!
That’s right – you can now test your Prompt Templates for Service Replies. For contextual prompts, you’ll need to input a transcript along with either a LiveChatTranscript or MessagingSession record. For grounded prompts, the testing process also requires an LLM search query and a Retriever ID or Name for Retrieval-Augmented Generation (RAG), in addition to the transcript and record.
You can also invoke the functionality through Connect API if needed, and enjoy all the benefits of using Prompt Templates.
Enjoy Fully Customized Reply Suggestions
After going through the new setup experience, customizing, and – of course – testing the Prompt Templates, it’s time for Service Agents to enjoy their fully personalized responses within the same embedded app they are accustomed to (if they have already been using Service Replies).
This means that the suggested replies will keep being generated as the conversation continues in real-time, all while leveraging information the Service Agent might need to include in their responses. On top of this, it will also be able to edit responses as required before hitting send.
As your Service Agents start using the functionality, make sure to keep an eye on the consumption – especially if your organization was already using Service Replies before these enhancements.
Data Cloud credits are used for processes such as ingestion, querying, and indexing for Knowledge, fields, and uploaded files. On the other hand, Einstein Requests will be consumed when the LLM provides the suggested replies.
Both can be easily tracked in your production instance in their respective consumption cards – found in the Digital Wallet within Your Account App.
Note: Make sure you ask your Salesforce AE if you need help estimating potential future consumption.
Final Thoughts
There’s no denying that amidst the autonomous agents’ news and overall Agentforce enhancements, Salesforce continues to invest in functionalities across the entire platform to allow users – and in this case service agents – to focus on what matters most while still having full control over the customer experience they provide.
Einstein Service Replies alleviates the burden of agents having to go through numerous articles, records, or even files, as all of the information they need within replies is readily available to them.
