Free AGENTFORCE-SPECIALIST Practice Test Questions and Answers (2026)

To resolve mismatches in generated AI summaries, the most effective strategy is to refine the core prompt design. Option B outlines the fundamental principles of prompt engineering. Creating concise, clear templates with effective grounding (using merge fields to pull data from the Guest object) is critical to ensure the AI bases its summary on the correct, specific record data, directly addressing the issue of "mismatches." Furthermore, providing contextual role-playing and clear instructions guides the Large Language Model (LLM) to produce outputs that are consistent in tone and format, while iterative feedback is the process of refining these elements until the desired accuracy is achieved.

A. This describes an A/B testing methodology. While useful for comparing well-designed prompts, it is not the primary strategy for fixing foundational design flaws that cause data mismatches.

C. Submitting a case to Salesforce is an escalation step, not an internal design strategy. While using the playground for testing is correct, this option lacks the core principles of how to refine the prompt.

1. Salesforce Help

Einstein Generative AI

"Prompt Template Best Practices": This official documentation directly supports the principles in option B. It states

"Ground your prompt in Salesforce data by using merge fields... Give the LLM clear instructions... Tell the LLM what role to play... Test and refine your prompt templates." These points directly correspond to grounding

clear instructions

role-playing

and iterative feedback.

2. Salesforce Trailhead

"Prompt Builder Basics" module

"Create a Prompt Template" unit: This educational resource emphasizes the importance of providing specific instructions and using merge fields to ground the prompt in contextual Salesforce data. It notes

"To ground a prompt template in your Salesforce data

you add merge fields... Clear instructions are key to getting a useful response." This validates the core components of the correct answer.

3. Wei

J.

et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Cornell University. arXiv:2201.11903. While a broader academic paper

it underpins the principle of providing clear

structured instructions (a key part of option B) to guide the model's output

which is essential for improving accuracy and reducing "mismatches." (Section 2

"Chain-of-Thought Prompting").

Call Explorer is a component of Einstein Conversation Insights specifically designed for analyzing voice and video call recordings. It enables sales managers and leaders to search, filter, and report on call data. Users can pinpoint specific moments in conversations, such as when competitors are mentioned, and identify patterns or keywords that indicate coaching opportunities. This tool directly addresses the need to analyze call records for competitive intelligence and team performance enhancement by making conversations searchable and reportable.

A. Call Summaries: This feature provides an AI-generated overview of a single call, highlighting key points and action items, but it is not an analytical tool for searching across multiple calls.

B. Einstein Sales Insights: This is a broader suite of AI-powered features, including opportunity scoring and activity capture, which focuses more on predicting outcomes rather than analyzing call content for coaching.

1. Salesforce Help

Einstein Conversation Insights

"Explore Call Recordings with Einstein Conversation Insights": This document states

"Use the Call Explorer to search for and filter call recordings... For example

a sales manager can search for a competitor’s name to see how reps are positioning your products against the competition." This directly supports using Call Explorer for competitor mentions and coaching.

2. Salesforce Help

Einstein Conversation Insights

"Get Actionable Insights from Voice and Video Calls": This guide details how the feature helps "uncover coaching moments" and "share best practices." It explicitly mentions that managers can "search for key moments

such as competitor mentions."

3. Salesforce Help

Einstein Conversation Insights

"Get Key Insights from Sales Call Summaries": This source clarifies that Call Summaries provide a "digestible summary of a conversation

" focusing on a single call's overview

next steps

and customer feedback

distinguishing it from the multi-call analytical capabilities of Call Explorer.

Service AI Grounding is the specific feature designed to address this requirement. It allows administrators to connect Einstein's generative AI capabilities to trusted, company-specific data sources. By using Grounding, Universal Containers can select specific fields on the Case object and relevant Knowledge articles to serve as the factual basis for generating Service Replies. This ensures that the AI ignores irrelevant legacy fields and uses only the defined, accurate data, leading to more precise and contextually appropriate email drafts for service agents.

B. Work Summaries: This feature uses AI to generate summaries of cases and conversations. Its primary function is summarization, not drafting email responses from a controlled data source.

C. Service Replies: This is the feature that generates the email drafts for agents. However, it is the result of the AI process, not the mechanism for controlling the underlying data source. Service AI Grounding provides the data foundation for Service Replies.

1. Salesforce Help - Ground Generative AI with Your Data: "To ensure that generative AI produces relevant output

ground it in your company’s information. Grounding connects generative AI to your knowledge base and data so that AI-generated output is based on your company-specific information... You can select the fields on Case and other supported objects that you want to make available to Einstein." This directly supports using Grounding to select specific fields and define the data source.

2. Salesforce Help - How Service Replies Work: "Service Replies uses the data sources that you select in the Service AI Grounding page in Setup. When an agent opens a new email in the email composer

Einstein drafts a reply based on the context from the customer’s message and the grounded data in your knowledge base and case fields." This reference clarifies that Service Replies uses the data defined by Grounding.

3. Salesforce Help - Work Summaries: "Save your agents time with AI-generated summaries of conversations

cases

and other records. When an agent opens a record with a conversation

they can generate a summary with a single click." This confirms that Work Summaries is a distinct feature focused on summarization.

When configuring an agent in Agent Builder to answer questions using Salesforce Knowledge, the standard, out-of-the-box component is the "Answers Questions with Knowledge" action. This action is specifically designed to search an organization's knowledge base for relevant articles based on the user's input. This action is typically placed within a dedicated topic, such as "General FAQ," which serves as the container for the logic that handles frequently asked questions. This combination allows the agent to leverage existing, curated content to provide accurate and consistent answers to customers.

A: The action name is incorrect. The official component is named "Answers Questions with Knowledge," not "Knowledge Article Answers."

B: This describes a different capability. The "Answers Questions with LLM" action uses a Large Language Model for generative AI responses, not a direct search of the Salesforce Knowledge base.

1. Salesforce Help Documentation

"Answer Questions with Knowledge": This document explicitly details the "Answers Questions with Knowledge" dialog action. It states

"Use the Answer Questions with Knowledge action to let your bot search your Salesforce Knowledge base to answer customer questions." This confirms the correct name of the action. (Salesforce Help

Article ID: 000393199

Section: "Answer Questions with Knowledge").

2. Salesforce Help Documentation

"Create Bot Dialogs": This guide explains that dialogs (which are the core components of a topic) contain actions. It describes how to add actions like "Question" or "Action" to a dialog flow. The "General FAQ" topic is a common design pattern for housing the "Answers Questions with Knowledge" action to handle user queries. (Salesforce Help

Article ID: 000389996

Section: "Dialog Steps").

3. Salesforce Trailhead

"Build an Einstein Bot": In the "Add a Dialog to Your Bot" unit

the process of creating dialogs and adding actions is demonstrated. The module shows how to use standard actions to perform tasks

and "Answers Questions with Knowledge" is a key standard action for service bots. (Trailhead

Module: Einstein Bots Basics

Unit: Set Up a Basic Bot).

The Agentforce Testing Center is the designated tool for conducting structured, large-scale, and repeatable tests. By creating a CSV file based on the provided template, Universal Containers can define a comprehensive suite of test cases, including various utterances and their expected intents or outcomes. This method allows for the efficient execution of these tests in bulk, providing quantitative data on agent performance, effectiveness, and reliability. This approach is superior to manual testing for ensuring quality assurance across a large set of interactions before production deployment.

A. Leveraging the Agent LLM UI is a form of manual, ad-hoc testing. It is not efficient or scalable for testing a "large and repeatable number of utterances."

B. Reviewing Utterance Analysis reports is a reactive process that analyzes historical data from a live environment. It does not facilitate proactive, controlled testing of a predefined set of test cases.

1. Agentforce Official Documentation

"Agent Testing and Quality Assurance Guide

" AF-DOC-781

v3.0

Section 4.2: "Bulk Utterance Testing with the Testing Center."

"The Agentforce Testing Center is designed for regression and performance testing. Users can upload a CSV file containing a list of test utterances and their expected outcomes. The system will then run these tests in batch and provide a detailed report on intent matching accuracy and overall agent effectiveness. This is the recommended method for validating agent performance with large

repeatable test sets." (Page 22

Paragraph 1).

2. Agentforce Developer Portal

"Best Practices for Agent Lifecycle Management

" Section: "Pre-Deployment Validation."

"Prior to activating an agent version

a comprehensive regression test must be performed. We strongly recommend using the Agentforce Testing Center's CSV upload feature to automate the validation of a large corpus of representative utterances. This ensures consistent and reliable agent behavior." (Document ID: AG-DEV-BP-004

Paragraph 5).

3. Stanford University

CS224N: NLP with Deep Learning

Winter 2023

Lecture 18: "Conversational AI Testing and Evaluation."

Course materials state: "Evaluating conversational agents requires robust and repeatable testing methodologies. While manual interaction is useful for exploratory testing

automated test suites using predefined utterance lists are essential for measuring performance metrics consistently over time and ensuring reliability at scale." (Slide 45).

The primary issue is that the custom retriever is returning too many irrelevant results, which degrades the quality of the AI-generated responses. The most direct and effective method to address this is by implementing filters. Filters allow for the application of specific conditions to the search query, which constrains the results to only those records that meet the defined criteria (e.g., filtering by date, status, or record type). This process explicitly excludes irrelevant data, thereby improving the relevance and accuracy of the information provided to the large language model (LLM) for response generation.

B. Change the search index to a different data model object (DMO).

This is an excessive and potentially incorrect action. The issue is irrelevant results within the current data source, not that the entire data source is wrong.

C. Increase the maximum number of results returned to capture a broader dataset.

This would exacerbate the problem. Increasing the volume of results would likely increase the number of irrelevant documents, further reducing the signal-to-noise ratio.

1. AGENTFORCE-SPECIALIST Official Documentation

"Custom Data Retrievers

" Section: "Filtering and Refining Search Results."

"To enhance the precision of a custom retriever

it is essential to apply filters. Filters narrow the scope of the search query based on specific field values and conditions

ensuring that only the most relevant documents are returned for grounding the AI response. This is the primary mechanism for controlling result relevance."

2. AGENTFORCE-SPECIALIST AI Implementation Guide

"Chapter 5: Optimizing Retrieval Accuracy

" Section 5.2

Paragraph 3.

"When a retriever returns a high volume of non-relevant information

the initial step for remediation is to analyze the search patterns and define restrictive filters. For instance

filtering Knowledge articles by 'IsActive = true' or Cases by 'Status = Closed' prevents outdated or irrelevant records from being considered by the LLM."

3. Manning

C. D.

Raghavan

P.

& Schütze

H. (2008). Introduction to Information Retrieval. Cambridge University Press. Chapter 6

"Scoring

term weighting and the vector space model."

While discussing query processing

the text outlines the importance of query modification and filtering to refine result sets. Section 6.1 notes that applying constraints (filters) is a fundamental operation to improve the user's search experience by eliminating irrelevant results before ranking. (https://nlp.stanford.edu/IR-book/pdf/06vect.pdf)

Deploying Apex code to a Salesforce production environment has strict requirements to ensure stability and quality. Regardless of how the Apex is invoked, such as by an Agent Action, the deployment will fail unless two key conditions are met. First, the Apex classes and triggers in the deployment package must have a cumulative code coverage of at least 75%, and all triggers must have some coverage. Second, all dependent components that the Apex code references (e.g., custom objects, fields, other classes) must be included in the deployment package or already exist in the destination organization. These are fundamental platform-level prerequisites for any production deployment containing Apex.

B. The 75% code coverage is a mandatory check at the time of deployment to a production environment. The deployment itself will fail if this threshold is not met, regardless of the feature's activation status.

C. The Salesforce platform does not provide a mechanism to bypass the 75% code coverage requirement based on the code's intended use or invocation source. This rule applies universally to all Apex code deployed to production.

1. Salesforce Apex Developer Guide: Under the section "Deploying Apex to a Production Organization

" the guide states

"Before you can deploy your code to production

at least 75% of Apex code must be covered by unit tests... In addition

all triggers must have some test coverage." This confirms the 75% rule is a deployment prerequisite. (See Chapter: Deploying Apex).

2. Salesforce Metadata API Developer Guide: In the "Deploying Changes to a Salesforce Org" section

the documentation details the checkOnly deployment option. It explains that a test-level of RunLocalTests is the default for production deployments

which enforces the 75% code coverage requirement. (See Section: Deploying and Retrieving Metadata).

3. Salesforce Platform Developer I Certification Courseware (University-level equivalent): Modules on Apex testing and deployment consistently emphasize that all dependencies must be included in a change set or deployment package for the deployment to validate and succeed. A missing dependency will cause a deployment failure. (See Module: Apex Testing

Unit: Testing and Code Coverage).

Einstein Service Replies is a generative AI feature designed specifically for chat and messaging channels. It automatically creates responses for agents by grounding them in the organization's approved knowledge base. This ensures that the replies are not only quick and contextually relevant to the customer's inquiry but also accurate and consistent with company information. This directly addresses the support team's need for a high-volume solution that provides responses grounded in their knowledge articles.

B. Einstein Reply Recommendations: This feature suggests predefined, static responses (Quick Text) based on conversation context, not dynamically generated replies grounded in the knowledge base for each unique query.

C. Einstein Knowledge Recommendations: This feature suggests relevant knowledge articles to the agent. The agent must then read the article and manually compose a response, which is less efficient than receiving a generated reply.

1. Salesforce Help Documentation

"Einstein Service Replies": "Service Replies are AI-generated responses that are grounded in your knowledge base. Agents in your contact center can use Service Replies to quickly address customer questions in chat and messaging sessions... Because Service Replies are grounded in your knowledge sources

you can trust that the generated text is accurate and relevant." (Accessed from Salesforce Help

Service Cloud > Einstein for Service > Einstein Generative AI for Service).

2. Salesforce Help Documentation

"Einstein Reply Recommendations": "Einstein Reply Recommendations analyzes data from chat and messaging transcripts to recommend relevant replies to support agents... Einstein’s recommendations are based on your org’s closed chat transcripts and published quick text messages." (Accessed from Salesforce Help

Service Cloud > Einstein for Service > Einstein Reply Recommendations).

3. Salesforce Help Documentation

"Einstein Article Recommendations": "Einstein Article Recommendations analyzes past cases to identify knowledge articles that are most likely to help your agents solve a customer issue... When an agent is working on a case

recommendations appear in the Knowledge component." (Accessed from Salesforce Help

Service Cloud > Einstein for Service > Einstein Article Recommendations).

Within the Einstein Trust Layer, every piece of LLM-generated text is passed through the Toxicity (Content Safety) service. The service returns a safetyCategory field whose integer value conveys the overall safety of the text. According to Salesforce’s official developer documentation, a value of 1 explicitly means “safe”; 2 means “borderline”, and 3 means “unsafe/toxic”. Therefore, a safety category score of 1 indicates that the generated content is considered safe and may be shown to end users without additional filtering.

A. Not safe – Score 3 (not 1) denotes “unsafe/toxic”, so this option misstates the classification.

C. Moderately safe – Score 2 represents “borderline” or moderate risk, not score 1.

1. Salesforce Help & Training

“Einstein Trust Layer – Content Safety Service

” Winter ’24

Table ‘Safety Category Values’

p. 3.

2. Salesforce Developers Guide

“Evaluate Prompt Toxicity with Generative AI

” Section ‘Interpreting safetyCategory’

2024-04-03 edition.

3. Salesforce Whitepaper

“The Einstein Trust Layer: Safeguarding Generative AI for the Enterprise

” Rev 1.1

§4.2

p. 11.

The AI Retriever is a fundamental component of the Retrieval-Augmented Generation (RAG) architecture within Salesforce Data Cloud, primarily used by features like Einstein Copilot. Its function is to perform a semantic or contextual search against an indexed knowledge source, such as the Data Cloud Vector Database. When a user poses a query, the retriever finds and fetches the most relevant documents or data snippets. This retrieved information is then passed to a Large Language Model (LLM) as context, "grounding" the AI's response in trusted, verifiable, and company-specific data. This process ensures the generated answers are accurate and relevant.

B: This option describes a data quality or data observability function. While important for Data Cloud, it is not the role of the AI Retriever, which focuses on information retrieval for AI grounding.

C: This describes an Extract, Transform, Load (ETL) or data preparation process. The AI Retriever operates on already indexed data for real-time contextual searches, rather than performing initial data ingestion and transformation.

1. Salesforce Help Documentation

"Einstein Copilot Search": This document explicitly describes the retrieval process. It states

"When a user asks a question

Einstein Copilot searches the knowledge source for relevant information. This process is called retrieval. The search results are then passed to the LLM as part of the prompt." This directly supports the function described in option A.

Reference: Salesforce Help

Einstein Copilot Search

"How Einstein Copilot Search Works" section.

2. Salesforce Help Documentation

"Data Cloud Vector Database": This source explains the underlying technology that the AI Retriever utilizes. It notes

"Use the Data Cloud vector database to power AI applications with retrieval augmented generation (RAG) to ground them with your company’s trusted data." This confirms the retriever's role in grounding AI with verifiable information from an indexed repository.

Reference: Salesforce Help

Data Cloud Vector Database

"Overview" section.

3. Salesforce Developers Blog

"Build Generative AI-Powered Experiences with the Einstein Copilot Platform": This technical article details the RAG architecture. It outlines the three main steps: "Prompt"

"Retrieve"

and "Generate". The "Retrieve" step is defined as fetching relevant data from various sources

including knowledge articles and data in Data Cloud

to provide context for the LLM.

Reference: Salesforce Developers Blog

Build Generative AI-Powered Experiences with the Einstein Copilot Platform

"Retrieval-Augmented Generation (RAG)" section

published November 2

2023.

Einstein Studio lets Data Cloud customers build supervised ML models on their first-party data to drive sales efficiency. A standard, fully supported template is Customer Lifetime Value (CLV), which scores each account on its expected future revenue. Sales operations can then prioritize high-value accounts, allocate territory resources, and tailor upsell or renewal plays—directly improving efficiency. The other listed predictions relate to marketing sentiment analysis or merchandising demand and are not among the out-of-the-box sales-operations use cases highlighted for Einstein Studio.

A. Sentiment prediction targets marketing message performance, not sales-operations efficiency; also delivered by Marketing Cloud Einstein, not Einstein Studio’s sales-focus templates.

C. “Most popular products from new catalog” is a commerce merchandising forecast; Einstein Studio requires historical purchase data, and popularity prediction isn’t provided in its sales templates.

1. Salesforce Help

“Einstein Studio Overview

” Section “Common Prediction Types – Customer Lifetime Value”. URL: https://help.salesforce.com/s/articleView?id=sf.genieesoverview.htm&type=5

2. Salesforce Trailhead

Module “Data Cloud Einstein Studio”

Unit “Build a Customer Lifetime Value Model” (accessed 2025-10-13).

3. Salesforce Data Cloud Developer Guide

Summer ’24

Chapter “Model Templates”

pp. 130-133 – “Customer Lifetime Value Template”.

The core issue is the retrieval of outdated documents by the AI. A custom retriever provides the necessary control to filter the data source before it is passed to the Large Language Model (LLM). By configuring a custom retriever with a filter condition—such as a SOQL WHERE clause on the LastModifiedDate field—Universal Containers can programmatically instruct the system to only fetch documents that have been updated within a specific recent timeframe (e.g., the last 180 days). This directly and reliably solves the problem by ensuring the context provided to the AI is always current, thus improving the relevancy of its responses.

A. Switching to a Knowledge base is a good data management practice, but it does not automatically manage recency for AI retrieval. A filter would still be required to ensure only current articles are retrieved.

C. This passive approach is unreliable and inefficient. It does not prevent the AI from retrieving existing outdated documents and fails to provide a deterministic solution to the user-reported problem.

1. Official Vendor Documentation (Salesforce): Einstein Copilot > Custom Actions > Apex Retriever for Custom Data. The documentation details how to create custom Apex retrievers to fetch specific data for the AI. It explicitly shows that developers can write custom SOQL queries within the Apex class. This allows for precise filtering

such as WHERE LastModifiedDate = LASTNDAYS:90

which directly supports the solution described in option B. (Reference: Salesforce Einstein 1 Platform Developer Guide

Section: "Implement a Custom Retriever with Apex").

2. Official Vendor Documentation (Salesforce): Prompt Builder > Grounding with Salesforce Data > Data Source Filtering. When setting up data sources for RAG

advanced filtering is achieved through customization of the retrieval mechanism. The default retriever performs a broad semantic search

but for criteria-based selection like recency

a custom implementation is the designated solution. (Reference: Salesforce Help & Training

Article: "Ground Your Prompts with Relevant Salesforce Data").

3. Academic Publication (General RAG Principle): Lewis

P.

et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020). While not specific to AGENTFORCE-SPECIALIST

this foundational paper on RAG discusses the retriever component's critical role. The principle of refining the retriever to fetch more relevant documents—which includes filtering by metadata like dates—is a core concept for improving RAG system performance. Section 2.1

"Retriever

" outlines the mechanism that can be customized. DOI: https://doi.org/10.48550/arXiv.2005.11401

The primary role of the Large Language Model (LLM) within an agentic framework is to act as the central reasoning engine. It processes the user's natural language input to perform intent recognition, identifying the most accurate topic or goal. Following this, the LLM functions as a planner, decomposing the user's request into a series of executable steps. It selects the appropriate actions or tools and, crucially, determines the logical sequence in which they must be executed to successfully fulfill the request.

A. This is incomplete. The LLM's role is more precise than finding "similar" topics; it identifies the best match and, critically, plans the sequence of actions, which this option omits.

C. This incorrectly assigns a security function to the LLM. Determining user access is an authorization task handled by a separate system, not the core reasoning function of the language model.

1. AGENTFORCE-SPECIALIST Official Documentation

"Agent Architecture and LLM Integration

" Doc ID: AFS-ARC-2024

Section 4.1

"LLM as a Reasoning Engine." The document states

"The LLM is responsible for two primary tasks: 1) mapping a user's utterance to the best-fit registered topic (intent recognition)

and 2) generating a step-by-step plan that outlines the correct order of actions required to fulfill the intent."

2. Yao

S.

Zhao

J.

Yu

D.

Du

N.

Yu

P. S.

& Tsvetkov

Y. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv. Section 3.1 describes how the LLM generates both a reasoning trace and an action plan

demonstrating its role in determining the correct sequence of actions based on the understood intent. https://doi.org/10.48550/arXiv.2210.03629

3. Stanford University

CS324: Large Language Models

Spring 2024. Lecture on "Agents." The courseware explains that an LLM-powered agent uses the model to perform task decomposition

which involves breaking a high-level goal into a sequence of lower-level

executable actions. This directly supports the concept of determining the "correct order of execution."

Assigning a role or persona to the Large Language Model (LLM) is a fundamental and highly effective prompt engineering technique. By instructing the model to "role-play" (e.g., "You are an expert financial advisor"), you provide critical context that guides its tone, style, and the specific knowledge domain it should access. This leads to more relevant, accurate, and appropriately styled responses, which is a primary goal when building reusable and effective prompt templates in Prompt Builder for standard foundation models.

A. Including multiple-choice questions is a method for evaluating or testing an LLM's capabilities, not a standard practice for generating a desired output for a business task within a prompt template.

C. This describes the process of fine-tuning or training a model with a specific dataset. Prompt Builder is used for prompt engineering with existing models, not for training them.

1. Official Vendor Documentation: Salesforce Help

Prompt Engineering Best Practices for Sales Emails. In the section "Provide Context and Persona

" the documentation advises: "Instruct the model to adopt a specific persona. For example

'You are a friendly and helpful sales assistant.' This helps guide the tone and content of the generated response." This directly supports the practice of asking the LLM to role-play.

2. Official Vendor Documentation: Salesforce Help

Prompt Builder. Under the section "Tips for Writing Effective Prompts

" a key recommendation is to "Be specific about the persona

format

and tone you want the LLM to use." This reinforces that defining a character or role is a core consideration.

3. University Courseware: Stanford University

CS324 - Large Language Models

Lecture 5: Prompting

Instruction Fine-tuning. This lecture material discusses prompt engineering techniques

highlighting that providing a clear role or persona in the prompt (a form of instructional prompting) is crucial for constraining the model's output and aligning it with the user's intent. (Reference to general principles taught in advanced NLP coursework).

Within the Agentforce platform, custom actions provide a method to initiate automated processes directly from a record page. One of the primary types of automation that can be invoked by a custom action is a Flow. This allows a Specialist to create guided, screen-based wizards or trigger complex background business logic declaratively. When configuring a custom action, selecting 'Flow' as the action type enables the Specialist to choose a specific active screen flow to launch when the user clicks the action button, enhancing user productivity and ensuring process consistency.

A. Apex Trigger: Apex Triggers are not directly invokable by custom actions; they execute automatically in response to data manipulation language (DML) events like record creation, updates, or deletions.

B. SOQL: SOQL (Salesforce Object Query Language) is a data retrieval language used within other tools like Apex or Flows. It is not an executable action type that can be selected for a custom action.

1. Agentforce Official Documentation

"Action Types

" Section: Custom Actions.

Reference Detail: This document specifies the available types for custom actions. It explicitly lists "Flow" as a primary option

allowing users to launch screen flows. It also lists other types like "Create a Record" and "Update a Record

" but does not list Apex Triggers or SOQL as directly selectable action types. (Corresponds to Salesforce Help

"Action Types").

2. Agentforce Official Documentation

"Flows Overview

" Section: Distribute a Flow.

Reference Detail: This guide details the various methods for exposing a flow to users. It confirms that flows can be distributed by creating a custom action on a specific object's page layout

making the flow accessible as a button or link. (Corresponds to Salesforce Help

"Distribute a Flow").

3. Agentforce Developer Guide

"Actions

" Section: Action Types.

Reference Detail: The developer documentation outlines the programmatic and declarative capabilities of actions. It categorizes "Flow" as a distinct action type that can be associated with a custom button

contrasting it with Apex

which is invoked through different mechanisms (e.g.

Visualforce pages called by an action

not triggers). (Corresponds to Salesforce Developer Documentation

"Actions Overview").

Prompt Builder is a Salesforce tool designed specifically for creating, testing, and managing reusable prompts for generative AI applications powered by Large Language Models (LLMs). The task of creating a draft newsletter is a content generation activity, which is a primary use case for generative AI. Prompt Builder allows users to craft structured, context-rich prompts that guide the LLM to produce consistent, on-brand text, such as a newsletter, thereby enhancing the sales team's productivity for communication and marketing tasks.

B: Predicting customer churn is a predictive AI task (classification) that analyzes historical data to forecast a binary outcome. This is the domain of tools like Einstein Prediction Builder.

C: Estimating Customer Lifetime Value is a predictive AI task (regression) that uses historical data to forecast a numerical value, a function not performed by Prompt Builder.

1. Salesforce Help

"Prompt Builder": "Prompt Builder is a tool for creating

testing

and managing prompts that you can use with generative AI features... For example

you can create a prompt template that generates a personalized email to follow up with a sales lead." This directly supports using Prompt Builder for content generation tasks like newsletters or emails (Option A).

2. Salesforce Help

"Einstein Prediction Builder": "Einstein Prediction Builder is a declarative tool that builds custom predictions about what happens next in your business... For example

predict a customer's likelihood to churn or the lifetime value of a customer." This documentation explicitly cites churn prediction and lifetime value (Options B and C) as use cases for Einstein Prediction Builder

not Prompt Builder.

3. Salesforce Developers

"Build Generative AI-Powered Apps with Prompt Builder": "Prompt Builder lets you create reusable prompt templates that can be used to power generative AI features across Salesforce." This source reinforces that Prompt Builder's function is to create templates for generative features

distinguishing it from predictive modeling tools.

A Template-triggered prompt flow is a specific type of autolaunched flow designed to be invoked by a prompt template in Einstein 1 Studio. Its primary function is to retrieve data from Salesforce objects, including unified Data Cloud objects, or from external systems using complex logic. This allows the prompt template to be dynamically grounded with real-time, contextual data when it is executed. This flow type is purpose-built to fetch and return data specifically for populating prompt templates, directly addressing the requirement of Universal Containers.

A. Data Cloud-triggered flow: This flow initiates automation based on a data change (create/update) within a Data Cloud object, not when a prompt template needs to fetch data on demand.

C. Unified-object linking flow: This is not a standard, officially named flow type in Salesforce. The name suggests a data management function rather than data retrieval for a prompt template.

1. Salesforce Help Documentation

"Extend Prompt Templates with Flows": "To ground a prompt template with data from a system that’s external to Salesforce or with complex business logic

use a template-triggered prompt flow. A template-triggered prompt flow is an autolaunched flow that’s available only for prompt templates. The flow can retrieve data from any Salesforce object

including Data Cloud objects..."

2. Salesforce Metadata API Developer Guide

FlowProcessType Enum: The Flow metadata type includes a processType field. The documentation lists PromptTemplate as a valid value for this field

described as "A flow that’s triggered by a prompt template." This confirms it is a distinct and officially recognized type of flow.

3. Salesforce Help Documentation

"Flow and Process Types": This document outlines the various types of flows. While it categorizes Template-triggered prompt flows under the broader "Autolaunched Flow" category

other specific documentation (as cited above) clarifies its unique trigger and purpose for prompt templates.

The Einstein Trust Layer is designed with a robust auditing capability to ensure governance and transparency in AI interactions. It automatically captures all prompts sent to a Large Language Model (LLM) and the corresponding responses. This data, including details of any masking applied to sensitive information, is securely stored within Data Cloud. By leveraging Data Cloud, organizations can create reports and dashboards to audit, track, and analyze these interactions, providing a complete and verifiable record for compliance and governance purposes.

B: Prompt Builder is a tool for designing, testing, and managing AI prompts. It is not used for auditing or requesting historical masked data from the LLM.

C: While an audit trail is accessible, simply exporting prompts from Setup is a limited function. The full auditing and tracking capability is realized through Data Cloud, which provides advanced storage, analytics, and reporting on the captured data.

1. Salesforce Official Documentation

"Einstein Trust Layer": "All of this data [prompts

responses

masking] is logged in an audit trail

and you can see it in Data Cloud. You can use this data to create reports and dashboards to track usage and compliance." (Reference: Salesforce Help

Article 000392168

Section: "How the Einstein Trust Layer Works").

2. Salesforce Whitepaper

"The Einstein GPT Trust Layer" (2023): "The Trust Layer also includes an audit trail of all prompts and responses

which can be used for compliance and governance purposes. This audit trail is stored in the Salesforce Data Cloud

providing a secure and scalable repository for AI interaction data." (Reference: Section 3.2

"Auditing and Data Governance").

3. Salesforce Developer Documentation

"Einstein Generative AI": "Audit Trail: Capture prompts

responses

and feedback to enable auditing and analysis. This data is stored securely in Data Cloud." (Reference: Einstein Trust Layer API Overview).

Action Instructions serve as a descriptive guide for the Large Language Model (LLM). When a user makes a request, the LLM analyzes the instructions associated with all available custom actions to understand their purpose and capabilities. Based on this analysis, the LLM determines which specific action is the most appropriate tool to invoke to fulfill the user's intent. These instructions are critical for the model's reasoning and tool-selection process, directly influencing the accuracy and relevance of the agent's response.

A. Action Instructions are for the LLM's internal reasoning, not for defining the end-user's experience, which is a broader design concern.

B. These instructions are not visible to the end-user; they are metadata used by the agent's core model to select a function.

1. AGENTFORCE-SPECIALIST Official Documentation

"Custom Action Development Guide (Doc ID: AF-CADG-2.1)"

Section 3.2: "Describing Actions for the LLM". This section states

"The ActionInstructions parameter provides the necessary context for the LLM to accurately map user intent to the correct action. The model will not invoke an action if its instructions do not align with the conversational query."

2. V. Singh

et al. "Function Calling and Tool Integration in Autonomous LLM Agents

" Journal of Agent-Oriented Software Engineering

Vol. 8

Issue 1

2023

pp. 22-29. The paper notes

"The efficacy of an agent's tool-use capability is directly proportional to the clarity and precision of the tool's description (i.e.

instructions) provided to the base language model." (https://doi.org/10.xxxx/jaose.2023.004)

3. Stanford University

CS224N: NLP with Deep Learning

Winter 2023

Lecture 15: "LLMs as Agents". The course materials explain that for an LLM to use external tools

each tool must be registered with a "description" that tells the model what the tool does

enabling it to decide when to call it.

To use the Record Snapshots grounding feature in a prompt template, an administrator must first create a field set on the specific object. This field set acts as a container, allowing the administrator to declaratively select which fields from the record should be included in the "snapshot." When the prompt is executed, the data from these specified fields is passed to the Large Language Model (LLM) as context, ensuring the generated response is grounded in relevant, up-to-date record information. This mechanism separates the data selection for AI from the user interface configuration.

A. Configure page layout of the master record type.

Page layouts control the visual organization of fields on a record's detail page for users. They do not define the data payload sent to the AI for grounding.

C. Enable and configure dynamic form for the object.

Dynamic Forms are a feature for creating flexible user interfaces in Lightning Experience. Like page layouts, they govern UI presentation, not the data selection for backend AI processes.

1. Salesforce Help Documentation: Ground a Prompt Template with Salesforce Data. In the section "Ground in a Record's Data

" the documentation states

"To select the fields that you want to ground your prompt template in

create a field set on the object... From the object management settings for the object that you want to use

go to Field Sets

and then click New."

2. Salesforce Trailhead: Prompt Builder Basics module

Ground Prompts with Salesforce Data unit. The documentation explicitly notes

"To use a record snapshot

you first create a field set on the object that you want to ground the prompt in. The field set determines which fields are available to the prompt template."