Free Agentforce practice test: 20 questions
Twenty questions across the six sections of the Agentforce Specialist exam, weighted roughly as the real exam is. AI Agents carries over a third of the score, so it gets the largest share here.
AI Agents
Question 1. A service agent must look up an order status and, if the order is late, offer a goodwill credit up to a fixed limit. What is the correct construction?
- A. A single free-form instruction asking the model to be helpful
- B. A topic with scoped actions for the lookup and the credit, constrained by the limit
- C. A static knowledge article describing the policy
- D. A report scheduled daily
Show answer
Answer: B
Topics define what the agent handles, and actions perform the lookup and the credit within defined limits. Free-form generation cannot enforce a monetary limit.
Question 2. An agent answers questions it was never intended to handle. What is the most effective correction?
- A. Increase the response length limit
- B. Add more knowledge articles
- C. Narrow the topic scope and sharpen the classification instructions
- D. Switch to a larger model
Show answer
Answer: C
Tightening topic scope and classification instructions keeps the agent within its intended remit. Making instructions longer without narrowing scope rarely helps.
Question 3. What determines which topic an agent selects for a given user utterance?
- A. The topic’s classification description and defined scope
- B. The number of records in the related object
- C. The length of the knowledge articles
- D. The order topics were created
Show answer
Answer: A
The topic’s classification description and scope guide selection. Record counts and article length do not influence routing.
Question 4. An action must never run without a human decision because it issues a refund. How should this be handled?
- A. Log the refund and notify a supervisor afterwards
- B. Restrict the agent to business hours
- C. Ask the customer to confirm
- D. Require explicit human confirmation before the action executes
Show answer
Answer: D
Requiring explicit confirmation before the action executes keeps the human in the decision. Logging and notification happen after the money has moved.
Question 5. An agent must access data the running user is not permitted to see. What happens?
- A. The agent returns the data with a warning
- B. Access is denied according to the security model of the executing context
- C. The agent escalates to an administrator automatically
- D. The agent returns a summary instead of the record
Show answer
Answer: B
Agent actions respect the security model of the executing context, so unauthorised data is not returned. The platform does not silently bypass sharing rules.
Question 6. Which is the strongest reason to prefer a defined action over free-form generation for a business operation?
- A. Actions are deterministic, respect permissions and produce an auditable record
- B. Actions are always faster
- C. Actions use fewer tokens
- D. Actions do not require testing
Show answer
Answer: A
An action produces a deterministic, permission-aware, auditable operation. Generated text cannot guarantee any of those properties.
Question 7. An agent must hand a conversation to a human with full context. What should the design include?
- A. A message telling the customer to call support
- B. Ending the session and creating a case with no detail
- C. An escalation path that transfers the transcript and collected context to a human queue
- D. A link to the knowledge base
Show answer
Answer: C
A defined escalation that transfers the transcript and collected variables preserves context. Restarting the conversation wastes the customer’s effort.
Prompt Engineering
Question 8. A prompt template must include the related account's industry. How should this be supplied?
- A. Type the industry into the template text
- B. Ground the template with merge fields resolved from the related record
- C. Ask the user to state the industry every time
- D. Store the industry in the agent’s name
Show answer
Answer: B
Grounding the template with record merge fields inserts live data at runtime. Typing values into the template hard-codes a single account.
Question 9. Generated summaries are inconsistent in structure between runs. What is the most effective fix?
- A. Specify the required output structure explicitly in the prompt template
- B. Raise the temperature
- C. Shorten the input records
- D. Run the prompt twice and pick one
Show answer
Answer: A
Specifying an explicit output structure in the template constrains the shape of the response. Raising temperature increases variation, which is the opposite.
Question 10. What is the purpose of grounding a prompt template in CRM data?
- A. To reduce the number of licences required
- B. To make responses longer
- C. To translate the response
- D. To supply factual, permission-aware record context to the model
Show answer
Answer: D
Grounding supplies factual, permission-aware context so the response reflects the actual record rather than the model’s general knowledge.
Question 11. A prompt template occasionally produces content that breaches tone guidelines. What control is most appropriate?
- A. Ask agents to review every response manually
- B. Reduce the number of users with access
- C. Explicit tone and prohibited-content instructions plus output masking
- D. Disable the template
Show answer
Answer: C
Explicit tone and prohibited-content instructions combined with output masking address it at both ends. Relying on reviewers scales poorly.
Data 360 Fundamentals
Question 12. Customer records exist in three systems with different identifiers for the same person. What capability resolves them into one profile?
- A. Identity resolution with configured match and reconciliation rules
- B. A scheduled export to a spreadsheet
- C. A report grouped by email address
- D. A validation rule on the account object
Show answer
Answer: A
Identity resolution matches and merges records into a unified profile using configured match rules. Simple imports duplicate rather than unify.
Question 13. Which concept describes mapping source fields into a standard structure so data from different systems can be used together?
- A. Ingestion
- B. Mapping to a harmonised data model
- C. Archiving
- D. Encryption
Show answer
Answer: B
Data mapping to a harmonised model puts disparate sources into a common shape. Ingestion moves data; mapping standardises it.
Question 14. An agent must use behavioural data that updates continuously. Which ingestion approach fits?
- A. A monthly batch import
- B. Manual upload when requested
- C. Streaming ingestion
- D. A quarterly data warehouse refresh
Show answer
Answer: C
Streaming ingestion keeps continuously changing data current. A monthly batch would leave the agent working from stale behaviour.
Question 15. Why does data quality matter more for an agent than for a dashboard?
- A. Agents process data faster
- B. Dashboards are never wrong
- C. Agents use less storage
- D. The agent acts on and asserts the data to customers, so errors become actions and statements
Show answer
Answer: D
An agent acts on the data and states conclusions to customers, so errors propagate into actions and statements rather than sitting on a chart a human interprets.
Testing, Deployment, and Maintenance
Question 16. Before releasing an agent, you must verify it handles representative utterances correctly. What is the appropriate method?
- A. A defined test set of representative utterances run in a sandbox each release
- B. The builder chatting with it informally
- C. Releasing to production and monitoring complaints
- D. Counting the number of topics
Show answer
Answer: A
A repeatable test set run in a sandbox gives comparable results release over release. Informal chatting by the builder is neither systematic nor repeatable.
Question 17. A change must reach production without disturbing unrelated configuration. What is the correct practice?
- A. Edit the agent directly in production during quiet hours
- B. Deploy through the standard sandbox-to-production release process
- C. Rebuild the agent from scratch in production
- D. Export and re-import all metadata
Show answer
Answer: B
Deploying through the standard change and release process from a sandbox preserves traceability and avoids collateral change. Editing production directly does neither.
Governance and Observability
Question 18. Compliance asks what the agent told customers about cancellations last quarter. What must be available?
- A. The number of conversations per week
- B. A list of topics configured
- C. Retained conversation transcripts for the period, with access controls
- D. The agent’s deployment date
Show answer
Answer: C
Retained conversation transcripts with access controls provide the record. Aggregate counts do not contain what was said.
Question 19. Which signal most usefully indicates that an agent is failing users in production?
- A. Escalation and abandonment rates broken down by topic
- B. Total number of conversations
- C. Average response length
- D. The number of knowledge articles
Show answer
Answer: A
Escalation and abandonment rates by topic point at where the agent fails. Total conversation volume says nothing about quality.
Multi-Agent Orchestration
Question 20. Two agents handle different domains, and a customer question spans both. What should the design provide?
- A. Ask the customer to open two conversations
- B. Merge both agents into one with all topics
- C. Refer the customer to a human immediately
- D. Orchestrated routing that hands off with context preserved between agents
Show answer
Answer: D
Orchestrated routing that passes context between agents keeps the conversation coherent. Asking the customer to start again pushes the problem back to them.
How did you do?
Sixteen or more correct suggests you are close. Below fourteen, the domain guides in this section are the fastest fix — this exam rewards knowing which construct enforces a requirement, not which one could be made to work.