AI-201 — Salesforce Certified Agentforce Specialist

Full question text, answer options, and explanations for this practice set — a spoiler-free alternative is the interactive quiz above for scored, shuffled practice.

1. 1. Universal Containers wants to automatically draft a short summary of every Opportunity record when a sales rep clicks a button on the record page. The summary should be written into a custom text field on the Opportunity. Which Salesforce feature is the most appropriate tool for this requirement?

   • A. A Flow with a Decision element and hard-coded summary templates
   • B. A Prompt Builder template grounded on the Opportunity record(correct)
   • C. A custom Apex trigger on the Opportunity object
   • D. A custom report with a summary formula field

   Explanation: Prompt Builder is the Salesforce feature designed for generating AI-powered text grounded on Salesforce record data — exactly this use case. A Flow with hard-coded templates (A) cannot produce natural-language summaries that adapt to record content. An Apex trigger (C) doesn't generate language, and a summary formula (D) only performs arithmetic on numeric fields, not text generation.

2. 2. A Salesforce admin needs to generate a personalized email body that incorporates the recipient's recent Case history and Account information. The output is not bound to any single field — the user will review and edit it before sending. Which prompt template type is most appropriate?

   • A. A Field Generation template tied to the Case Description field
   • B. A Flex template, because the output is free-form and reviewed before use(correct)
   • C. An Apex prompt template invoked from an after-insert trigger
   • D. A Field Generation template tied to a custom Long Text Area field on the Account

   Explanation: Flex templates are designed for free-form, multi-purpose prompts where the output is not constrained to write into a single record field — for example, drafting an email a user will review and edit. Field Generation templates (A, D) are specifically designed to populate a single record field; that's not the scenario here. Apex-only templates from triggers (C) are not the standard pattern for user-reviewed output.

3. 3. A company wants a prompt template to answer customer questions using up-to-date product documentation stored as a large library of PDF files. The information changes frequently and is too large to include directly in the prompt. Which grounding technique should the admin use?

   • A. Merge fields that reference fields on the related Product record
   • B. Embed the full text of all PDFs directly into the prompt template body
   • C. Ground the template on a Data 360 retriever that searches an indexed library of the PDF content(correct)
   • D. Hard-code the most common product Q&A pairs as static text in the prompt

   Explanation: When the source material is large, dynamic, and unstructured (PDFs, articles), the appropriate grounding technique is retrieval-augmented generation using a Data 360 retriever against an indexed library — only the most relevant chunks are fetched at prompt execution time. Merge fields (A) only return fields from a single related record. Embedding all PDFs (B) would exceed context limits and not stay current. Hard-coded Q&A (D) doesn't scale and goes stale immediately.

4. 4. Which capability of the Einstein Trust Layer specifically prevents personally identifiable information (PII) from being sent to a third-party LLM provider?

   • A. Toxicity detection
   • B. Zero data retention agreements with model providers
   • C. Data masking (PII detection and tokenization before the request leaves Salesforce)(correct)
   • D. Audit trail logging in Setup

   Explanation: Data masking is the Trust Layer feature that detects PII in the prompt and replaces it with tokens before the request is sent to the LLM, then restores the original values in the response. Toxicity detection (A) scans for harmful content but does not mask PII. Zero data retention (B) ensures the provider does not store the request, but it does not prevent PII from being sent. The audit trail (D) logs activity but does not redact data.

5. 5. Which TWO statements accurately describe security and privacy features provided by the Einstein Trust Layer? (Choose 2)

   • A. Prompts and responses are logged for auditing without retaining customer data at the LLM provider(correct)
   • B. Sensitive data such as PII is masked before being sent to the LLM and unmasked in the response(correct)
   • C. The Trust Layer fine-tunes a private copy of the model with the customer's Salesforce data each night
   • D. The Trust Layer disables all third-party model access by default and only allows Salesforce-hosted models
   • E. The Trust Layer requires every user to grant explicit consent before any prompt is sent

   Explanation: The Trust Layer provides an auditable log of prompts and responses while relying on zero-data-retention agreements with model providers (A), and it masks sensitive data such as PII before sending and unmasks in the response (B). It does NOT fine-tune the model with customer data (C) — that's the opposite of how the Trust Layer's privacy guarantees work. It does not disable third-party models by default (D); admins can manage and prevent specific models from being accessed, but third-party access is supported. There is no per-prompt user-consent requirement (E).

6. 6. In the context of Data 360 grounding for prompt templates and agents, what is "chunking"?

   • A. Splitting large source documents into smaller passages so they can be indexed and retrieved individually(correct)
   • B. Compressing the response from the LLM into a fixed character budget
   • C. Batching multiple user prompts into a single LLM request to save tokens
   • D. Dividing a Salesforce org's data into separate buckets per Business Unit

   Explanation: Chunking is the foundational Data 360 concept of breaking large source documents into smaller, semantically meaningful passages so that each one can be indexed separately and surfaced by a retriever. Compressing LLM responses (B) and batching prompts (C) are unrelated to the Data 360 chunking concept. Splitting data by Business Unit (D) is a multi-tenancy concept, not chunking.

7. 7. A company wants to ground an agent on internal HR policy PDFs so that employee questions about leave, expenses, and benefits get accurate, citation-backed answers. They want a low-code experience to upload documents and make them queryable by the agent. Which Salesforce capability is the best fit?

   • A. Manually paste the policy text into the agent's system instructions
   • B. Agentforce Data Library, which lets admins upload documents and exposes them to agents and prompt templates as a retriever(correct)
   • C. A static resource in a Lightning Web Component referenced by an Apex action
   • D. Knowledge articles indexed by Salesforce's built-in full-text search, with no Data 360 involvement

   Explanation: Agentforce Data Library is the low-code experience for uploading documents to be chunked, indexed, and exposed to agents and prompt templates as a retriever — the exact scenario described. Pasting text into system instructions (A) doesn't scale and isn't retrieval. A static resource referenced from Apex (C) is a developer workaround that doesn't provide a retriever. Plain Knowledge with full-text search (D) is an option but it isn't the document-grounded retriever the question asks for and doesn't leverage Data 360.

8. 8. In a Data 360 grounding pipeline, what is the role of a retriever?

   • A. It periodically copies records from Sales Cloud to a data warehouse
   • B. It accepts a query at prompt-execution time and returns the most relevant indexed chunks to inject into the LLM context(correct)
   • C. It transforms the raw model output before showing it to the user
   • D. It fine-tunes the underlying LLM weights with the customer's data

   Explanation: A retriever is the component that, at prompt-execution time, takes a query (often derived from the user's prompt) and returns the most relevant indexed chunks so they can be added to the LLM context as grounding. It is not an ETL/sync job (A), it does not transform the model output (C), and it does not fine-tune the model (D).

9. 9. An admin is configuring chunking for a library of long technical PDFs (40+ pages each). What is the primary trade-off they need to consider when choosing a chunk size?

   • A. Larger chunks always reduce hallucinations because they include more context
   • B. Smaller chunks give more precise retrieval but may lose surrounding context; larger chunks preserve context but reduce retrieval precision and increase token usage(correct)
   • C. Chunk size has no practical effect — only the number of chunks matters
   • D. Smaller chunks reduce the number of API calls needed during retrieval

   Explanation: The fundamental chunking trade-off is precision vs. context: smaller chunks retrieve very specifically but can miss surrounding context that matters; larger chunks preserve context but degrade retrieval precision and use more tokens in the LLM call. Option A is wrong because larger chunks can actually increase noise and hallucination risk. Option C is wrong — chunk size directly affects retrieval quality. Option D is wrong because retrieval call count is not driven by chunk size in this way.

10. 10. What does "indexing" produce in a Data 360 grounding pipeline?

    • A. A standard Salesforce report saved into a private folder
    • B. A searchable representation (typically vector embeddings) of each chunk so that retrievers can find semantically relevant chunks at query time(correct)
    • C. A database index added to a custom object's primary key column
    • D. A Power BI dashboard that summarizes ingested documents

    Explanation: In a Data 360 grounding context, indexing means generating a searchable representation (typically vector embeddings) of each chunk so retrievers can find semantically similar chunks at query time. Standard Salesforce reports (A), relational database indexes on a primary key (C), and Power BI dashboards (D) are all unrelated to the Data 360 indexing concept used for grounding.

11. 11. In Agentforce, which two elements are the basic building blocks that define what an agent knows how to do and the work it can perform?

    • A. Profiles and Permission Sets
    • B. Topics and Actions(correct)
    • C. Validation Rules and Triggers
    • D. Reports and Dashboards

    Explanation: An Agentforce agent is built from Topics (groupings of capabilities and instructions that scope a conversation area) and Actions (the concrete steps the agent can take — searching knowledge, calling a Flow, executing Apex, running a prompt template, etc.). Profiles/Permission Sets (A) control user access, not agent capabilities. Validation Rules and Triggers (C) belong to the Salesforce data layer. Reports/Dashboards (D) are for analytics.

12. 12. An organization wants to deploy an agent that helps internal HR employees look up policies, request time off, and update employee records. The agent must respect the running user's Salesforce permissions and only be available to authenticated employees. Which agent type is most appropriate?

    • A. Service Agent, configured on an unauthenticated Experience Cloud site
    • B. Employee Agent, exposed in internal channels and running under the logged-in user's permissions(correct)
    • C. Service Agent, embedded in a customer messaging channel
    • D. An Apex-only integration with no agent involved

    Explanation: Employee Agents are designed for authenticated internal users and run with the security context of the logged-in employee, which matches the requirement to respect Salesforce permissions. Service Agents (A, C) are designed for external customer-facing scenarios. An Apex integration (D) doesn't provide the conversational agent experience the use case requires.

13. 13. A team is building an agent that needs to call an external pricing microservice and pass the result back into the conversation. None of the standard Agent Actions can call this microservice. What should the builder configure?

    • A. A custom Agent Action backed by a Flow (or Apex) that calls the microservice and returns the result(correct)
    • B. Embed the microservice URL directly into the agent's topic instructions and let the LLM call it
    • C. Configure the agent to bypass the Trust Layer so it can reach the external endpoint
    • D. Create a new standard topic; standard topics automatically expose external HTTP endpoints

    Explanation: Custom Agent Actions — typically implemented as Flows or Apex (or prompt templates / API calls in supported channels) — are how a builder extends an agent beyond the standard out-of-the-box actions. The agent invokes the custom action like any other, and the Flow/Apex performs the external HTTP call. LLMs do not directly make HTTP calls from topic instructions (B). Bypassing the Trust Layer (C) is not how external calls work and would violate governance. Standard topics (D) do not auto-expose external endpoints.

14. 14. Which THREE of the following are channels Agentforce supports for connecting an agent to end users? (Choose 3)

    • A. Slack(correct)
    • B. Email(correct)
    • C. Voice (Service Cloud Voice)(correct)
    • D. Microsoft Word
    • E. Oracle E-Business Suite

    Explanation: The exam guide explicitly lists digital experience, email, voice, and Slack as channels agents can be connected to. Slack (A), Email (B), and Voice (C) are all valid. Microsoft Word (D) and Oracle E-Business Suite (E) are not Agentforce channels.

15. 15. An agent must always greet the customer by name, must never quote pricing, and must hand off to a human after three failed clarification attempts. The builder wants this behavior to be deterministic, not left to the LLM's discretion. Which approach BEST aligns with Agentforce's mechanisms for deterministic agent behavior?

    • A. Write a long system instruction that politely asks the LLM to follow each rule
    • B. Use filters, variables, and template expressions in topics/actions to enforce the conditions explicitly, rather than relying solely on natural-language instructions(correct)
    • C. Increase the LLM temperature setting so the agent becomes more decisive
    • D. Disable the Trust Layer for that agent so its responses are not reformatted

    Explanation: Agentforce provides deterministic-behavior mechanisms — filters, variables, and template expressions — that let builders explicitly enforce conditions like "never quote price" or "hand off after N attempts" rather than relying on the LLM to obey natural-language instructions. Long system instructions (A) are useful but are probabilistic, not deterministic. Increasing temperature (C) does the opposite of what's needed. Disabling the Trust Layer (D) is not the mechanism for deterministic behavior and would also remove security guarantees.

16. 16. When an Agent Action runs, whose Salesforce permissions does it execute under by default?

    • A. The System Administrator profile, regardless of who is interacting with the agent
    • B. The user identity configured on the agent (the security context of the agent), not necessarily the end user(correct)
    • C. The Guest User profile of the parent Experience Cloud site
    • D. Whichever user owns the related record at the time of execution

    Explanation: An agent runs in a configured security context (the user identity assigned to the agent), which determines the permissions available to its actions — this is what governs CRUD/FLS for the records the agent reads or writes. It is not automatically the System Administrator (A), the Guest User (C), or the record owner (D). Builders must understand this so they don't accidentally grant or block access through misconfiguration.

17. 17. A development team is building a custom mobile app (not on the Salesforce platform) and wants it to converse with an Agentforce agent the company has already built. Which approach is appropriate?

    • A. Embed the Agentforce iframe directly into the mobile app's settings screen
    • B. Call the Agent API from the mobile backend to send user messages to the agent and receive its responses(correct)
    • C. Duplicate the agent's topics and actions into the mobile app's local code
    • D. Expose the agent through a public unauthenticated Experience Cloud site

    Explanation: The Agent API is specifically designed to let external (non-Salesforce-platform) applications invoke an existing Agentforce agent, send turns, and receive responses programmatically. Embedding an iframe (A) is not how a native mobile app integrates with Agentforce. Duplicating topics/actions in client code (C) is impractical and unsupported. Exposing the agent on an unauthenticated public site (D) is a security and governance anti-pattern, not the integration mechanism.

18. 18. Which statement BEST describes the relationship between a Topic and an Action in an Agentforce agent?

    • A. Topics are individual SQL queries; Actions are stored procedures
    • B. A Topic groups related instructions and Actions that the agent uses to handle a particular subject area; an Action is a discrete capability (e.g. search knowledge, call a Flow) the agent can invoke(correct)
    • C. Topics and Actions are synonyms; the choice of name is purely cosmetic
    • D. Actions contain Topics; Topics are the smallest configuration unit

    Explanation: A Topic scopes a subject area of the conversation, providing instructions and grouping the Actions available within it; an Action is a discrete capability the agent can invoke (search knowledge, call a Flow, run a prompt template, etc.). They are not SQL constructs (A), not synonyms (C), and Actions do not contain Topics — it's the other way around (D).

19. 19. A builder needs an agent to answer customer questions from published Salesforce Knowledge articles. A standard Agent Action already exists to search Knowledge. The builder is debating whether to use the standard action or create a custom Flow-based action that wraps the same functionality. Which is the BEST practice?

    • A. Always create a custom action — standard actions cannot be trusted in production
    • B. Use the existing standard Agent Action for searching Knowledge unless a specific business need (e.g. custom filtering, logging) requires a custom wrapper(correct)
    • C. Use both: register the standard action AND a duplicate custom action so the agent can pick whichever the LLM prefers
    • D. Avoid using Knowledge in agents — only Data 360 retrievers should be used for any text content

    Explanation: Best practice is to use the existing standard Agent Action when it meets the need — standard actions are maintained by Salesforce, optimized, and reduce maintenance burden. A custom action should be introduced only when there is a specific business need not covered by the standard action (custom filtering, audit logging, etc.). Always-custom (A) is wasteful. Registering duplicate actions (C) confuses the agent's action selection. Avoiding Knowledge entirely (D) is incorrect — Knowledge is a supported, recommended grounding source.

20. 20. What is the primary purpose of Agentforce Testing Center?

    • A. To generate the production deployment package for an agent
    • B. To run a curated set of test inputs against an agent and evaluate the quality of its responses at scale(correct)
    • C. To replace the Apex test framework for any code referenced by the agent
    • D. To simulate end-user devices in different browsers

    Explanation: Agentforce Testing Center allows builders to run an agent against curated test inputs (datasets) and evaluate the quality of its responses using evaluators — at a scale beyond what manual testing can achieve. It is not a deployment packager (A), it does not replace Apex unit tests (C), and it is not a browser/device emulator (D).

21. 21. An admin needs to move a tested agent from a sandbox to production. Which Salesforce mechanism is appropriate for promoting the agent's metadata?

    • A. Manually re-create the agent configuration in production by hand
    • B. Use established deployment patterns such as Change Sets, Metadata API, or DevOps Center to migrate the agent's metadata from sandbox to production(correct)
    • C. Export production data into the sandbox and then swap the org URLs
    • D. Agents cannot be moved between orgs and must always be built directly in production

    Explanation: Agent metadata follows the standard Salesforce ALM patterns — Change Sets, Metadata API, packages, or DevOps Center — to move from sandbox to production. Manually recreating the agent (A) is error-prone and not the supported pattern. Swapping org URLs (C) is nonsensical. Building only in production (D) is the opposite of safe ALM practice that the exam guide expects candidates to know.

22. 22. Before deploying a prompt template from sandbox to production, which TWO considerations should the admin verify? (Choose 2)

    • A. All referenced merge-field objects, fields, and related grounding sources (e.g. Data 360 retrievers) exist in the production org(correct)
    • B. The production org has access to the model(s) the template is configured to use(correct)
    • C. The sandbox's Trust Layer has been disabled to allow direct LLM calls
    • D. All user passwords in production have been reset
    • E. The Apex test coverage of the prompt template is exactly 100%

    Explanation: A prompt template's behavior depends on (A) the existence of every referenced merge-field/object/grounding source in production and (B) the production org's access to the model(s) the template uses — both must be verified pre-deployment. Disabling the Trust Layer (C) is neither required nor advisable. Resetting passwords (D) is unrelated to template deployment. There is no "100% Apex test coverage" requirement for prompt templates (E); that's a code-coverage requirement for Apex, not prompt templates.

23. 23. An organization wants to monitor how their production agent is performing — for example, conversation volume, action success rates, and topics most frequently invoked. Which Agentforce capability is designed for this?

    • A. Setup > Login History
    • B. Agent analytics — built-in dashboards/metrics that surface agent usage, performance, and outcomes for management and optimization(correct)
    • C. Apex Exception Email subscriptions
    • D. The Lightning Inspector browser extension

    Explanation: Agent analytics is the Agentforce capability for managing and monitoring agents — surfacing usage, performance, and outcomes so admins can optimize them. Login History (A) tracks user sign-ins, not agent behavior. Apex Exception emails (C) only fire on uncaught exceptions in Apex code. Lightning Inspector (D) is a developer tool for Lightning components, unrelated to agent monitoring.

24. 24. After observing that an agent's responses are inconsistent on a specific topic, which optimization step BEST aligns with the agent-optimization workflow Agentforce promotes?

    • A. Delete the agent and start over from scratch in production
    • B. Review analytics for that topic, identify failing conversations, refine the topic's instructions/actions/grounding, then re-test in Testing Center before re-deploying(correct)
    • C. Increase the model's max tokens to 10x the default
    • D. Globally disable the Trust Layer to give the LLM more freedom

    Explanation: Agent optimization is an iterative loop: observe analytics, identify failing interactions, refine topics/actions/grounding, re-test in Testing Center, then re-deploy. Deleting and starting over (A) wastes work and loses learnings. Inflating max tokens (C) doesn't address the root cause. Disabling the Trust Layer (D) is a governance violation, not an optimization step.

25. 25. A company is starting their Agentforce journey with one well-scoped Service agent for customer support. Stakeholders ask whether they should immediately split the workload across multiple specialist agents communicating via open standards like MCP and A2A. What is the BEST initial recommendation per the exam guide's guidance on Single-Agent (SOMA) architecture?

    • A. Always start with multi-agent orchestration because it scales better in every case
    • B. Start with a Single-Agent (SOMA) architecture when the use case is well-scoped — it offers better scalability, control, and lower operational complexity at the outset, and adopt multi-agent protocols like MCP/A2A only when the use case genuinely requires inter-agent communication(correct)
    • C. Do not deploy any agent until the company has built its own proprietary multi-agent protocol
    • D. Skip Agentforce entirely and integrate ChatGPT directly into Service Cloud

    Explanation: The exam guide highlights Single-Agent (SOMA) architecture as the right starting point for well-scoped use cases because it provides scalability and control with lower operational complexity; multi-agent protocols such as MCP and A2A are open standards used when orchestrating across multiple specialist agents becomes genuinely necessary. Always starting multi-agent (A) is overkill and adds complexity. Building a proprietary protocol (C) contradicts the guide's emphasis on adopting open standards. Bypassing Agentforce for ChatGPT (D) is not the certified specialist's recommended path.