SalesforceAI-2015 domains

AI-201 Exam Notes

Last-minute traps, must-know facts, and scenario tips for the Salesforce Certified Agentforce Specialist exam.

General Exam Tips

1.Read ALL answer options before selecting — many questions have a plausible-but-wrong option that sounds correct but violates Salesforce best practices
2.Scenario questions contain crucial clues in the constraints: look for words like 'without code', 'external system', 'multiple sources', 'post-deployment', and 'pre-deployment' — they signal which tool or approach to pick
3.No penalty for wrong answers — flag hard questions, keep moving, and come back. 105 minutes for 60 questions gives you about 1 min 45 sec each
4.Multiple-select questions: read the question stem carefully for 'Select TWO' or 'Select ALL that apply' — missing a required option counts as wrong
5.The exam was replaced in March 2025 from the old AI Specialist — any prep material from before March 2025 covers deprecated content that will hurt your score
6.When two answers both seem correct, the Salesforce-recommended best practice wins. Eliminate technically-valid-but-not-ideal options first
7.Many candidates report the exam is harder than expected — scenario-based questions test judgment, not just memorization. Hands-on practice in a Developer Edition org is the most-cited differentiator between passing and failing
8.Draw a quick mental diagram for complex agent configuration questions: user utterance → Atlas → topic → action → result
9.Domain weights: AI Agents (35%), Prompt Engineering (20%), Data Cloud (20%), Development Lifecycle (20%), Multi-Agent (5%). Spend study time proportionally — AI Agents alone is roughly 21 of 60 questions
10.Aim for 80%+ on practice exams before sitting the real exam — the 73% passing score has less margin for error than it looks on scenario-based questions

Quick Navigation

AI Agents Prompt Engineering Data Cloud for Agentforce Development Lifecycle Multi-Agent Interoperability

Domain 135% of exam

AI Agents

Must-Know Facts

Atlas reasoning engine pipeline: user utterance → topic selection → action planning → action execution → response delivery. Atlas selects the topic FIRST, then plans actions — it does NOT execute directly
Four agent types and their distinct use cases: Service Agent (external customer-facing, replaces chatbots without preprogrammed scenarios), Sales SDR Agent (lead qualification and meeting booking, top-of-funnel automation), Sales Coach (personalized coaching with role-playing for sales reps, rep development), Employee Agent (internal role-based collaborator for staff)
Topics define what jobs the agent can do. Instructions set decision-making boundaries for how the agent behaves within a topic. Actions are the specific tasks the agent executes
Agent User is a special Salesforce system user that the agent runs AS — it is NOT the human interacting with the agent. Data access is controlled via permission sets on the Agent User, not the end user's profile
Standard topics/actions come pre-built and TYPE-SPECIFIC with each agent type — the standards for Service Agent differ from those for Sales SDR or Employee Agent. Custom topics/actions are manually created for unique requirements
Action types: invocable Apex (@InvocableMethod), REST Apex class, autolaunched Flow, Prompt Template, MuleSoft API — each has a distinct trigger pattern and use case
Apex-based custom actions require 75% test coverage — same as any Apex in Salesforce. This is a Salesforce platform requirement, not an Agentforce-specific rule
Channel connections available: Slack, Messaging (SMS, WhatsApp), web chat, email, custom channels. Connecting to a channel does NOT automatically configure Agent User permissions — permissions must be set separately
Deterministic behavior controls: filters and variables constrain Atlas's decision range to produce predictable responses for compliance-sensitive or regulated scenarios
Classification Description on each topic is what Atlas uses to select the correct topic — a poor classification description leads to wrong topic selection, which is one of the most common agent misconfiguration issues

Common Traps

TrapThe Agent User and the end user are the same entity

RealityCompletely separate. The Agent User is a system user the agent runs AS — like a service account. The end user is the human typing messages. Agent data access is controlled by the Agent User's permission sets, not the end user's profile or sharing rules

TrapAtlas directly executes actions when it receives an utterance

RealityAtlas first selects a topic, then plans the action sequence, THEN the actions execute. Topic selection is always the first step. Questions asking 'what happens first when an utterance is received' — the answer is topic selection, not action execution

TrapAdding an agent to a Slack channel automatically grants the agent correct permissions

RealityChannel connection and Agent User permissions are configured separately. Connecting to Slack gives access to the channel; permission sets on the Agent User control what data the agent can actually read and what actions it can execute

TrapStandard topics and actions are the same across all agent types

RealityStandard topics and actions are TYPE-SPECIFIC. A Service Agent's standard topics are different from a Sales SDR's or an Employee Agent's. Knowing which standard capabilities come pre-built with each agent type is tested directly

TrapSales Coach and Sales SDR are interchangeable for sales automation

RealitySales SDR = outbound, automates lead qualification and books meetings to scale the top of the pipeline. Sales Coach = rep development tool, provides personalized coaching with role-play scenarios for existing sales reps. A question about automating outreach → SDR. A question about training reps → Coach

TrapDeterministic behavior means the agent ignores user input and always does the same thing

RealityDeterministic controls (filters and variables) constrain the RANGE of decisions Atlas can make for specific scenarios. The agent still reasons — it does so within tighter boundaries. Used when you need predictable, compliant behavior, not scripted chatbot responses

TrapAny Salesforce user can manage agent configurations in Agent Builder

RealityAgent Builder requires specific admin permissions. The Agent User's permission sets also need to be explicitly configured for each use case — security is not inherited from the builder's own permissions

Confusing Pairs

Service AgentEmployee Agent

Service Agent = EXTERNAL (customer-facing), deployed on web chat, Messaging, and email for customer support. Employee Agent = INTERNAL (staff-facing), works alongside employees with role-based permissions matching the employee's own access. Key question: who is the end user? Customers → Service. Internal staff → Employee

Sales SDR AgentSales Coach Agent

Sales SDR = top-of-funnel automation (qualifies leads, books meetings, scales outreach 24/7). Sales Coach = rep development (role-playing sessions, coaching, feedback on deal handling). Key: SDR scales the pipeline. Coach develops the people in the pipeline

Standard Topics/ActionsCustom Topics/Actions

Standard = pre-built by Salesforce, bundled with a specific agent type, minimal setup required, type-specific. Custom = built by your team using Apex, Flow, Prompt Templates, or MuleSoft for requirements not covered by standards. Key: 'available out of the box for this agent type' → standard. 'Unique business requirement' → custom

Invocable Apex ActionFlow Action

Invocable Apex = custom code for complex logic, external API callouts, or computation — requires 75% test coverage. Flow Action = declarative, no-code process automation for record updates, approvals, and multi-step processes — most common action type for admins. Key: 'complex code logic' or 'external REST callout in Apex' → Invocable Apex. 'Declarative, no code, update records' → Flow

Atlas Topic SelectionAtlas Action Planning

Topic Selection = Atlas reads the utterance and picks the most relevant topic using the Classification Description — this always happens first. Action Planning = after a topic is selected, Atlas decides which actions in that topic to execute and in what order. These are sequential steps, not parallel

Scenario Tips

If the question asks about:

When the question describes a business wanting to handle customer service inquiries without preprogrammed scripts across web and mobile channels...

Answer:

Configure a Service Agent. It is designed for external customer-facing interactions and uses Atlas reasoning rather than predefined flows, handling novel questions without scripting

Distractor to avoid:

Employee Agent sounds like it could serve anyone but is for INTERNAL staff only. Do not confuse 'serves the business' with 'is external-facing'

If the question asks about:

When the question asks how to restrict what data an Agentforce agent can access when responding to queries...

Answer:

Configure permission sets on the Agent User. The Agent User's permissions define the agent's data access boundary, regardless of what the end user can see

Distractor to avoid:

Profile-based security or sharing rules on records sound correct but apply to human users, not to the Agent User system account. The exam specifically tests this distinction

If the question asks about:

When the question asks which action type to use to call out to an external REST API with custom response transformation...

Answer:

REST Apex class action or Invocable Apex action. Use REST Apex for REST callouts. Use Invocable Apex for custom callouts with business logic. Remember: Apex actions require 75% test coverage before deployment

Distractor to avoid:

MuleSoft API action is also valid for external APIs but requires the MuleSoft integration layer. If the question specifies no MuleSoft or asks for the simplest approach with existing Apex code, choose the Apex action type

If the question asks about:

When the question asks about making agent behavior predictable for a compliance-sensitive scenario where incorrect responses could cause regulatory issues...

Answer:

Use deterministic behavior controls — filters and variables constrain the agent's decision range for specific scenarios while still using Atlas reasoning

Distractor to avoid:

Replacing the agent with a traditional Flow-based chatbot removes AI reasoning entirely. The exam expects you to use Agentforce-native controls, not abandon AI for scripted bots

If the question asks about:

When the question asks what an administrator should check first if the agent is selecting the wrong topic for user messages...

Answer:

Review the Classification Description on each topic. Atlas uses Classification Descriptions to select topics — a vague or overlapping description causes wrong topic selection

Distractor to avoid:

Checking action configuration or Trust Layer settings won't fix topic selection errors — the issue is upstream at the Classification Description level

Last-Minute Facts

1Domain weight: 35% — roughly 21 questions out of 60

2Atlas pipeline order: utterance → topic selection (uses Classification Description) → action planning → execution → response

3Apex actions: 75% test coverage required before deployment

4Agent types: Service (external customers), Sales SDR (lead qualification + booking), Sales Coach (rep coaching + role-play), Employee (internal staff)

5TIA = Topics (what jobs), Instructions (boundaries/rules), Actions (specific tasks)

6Agent User = system user the agent runs as, NOT the human end user. Permissions set via permission sets on Agent User

7Channels: Slack, Messaging (SMS, WhatsApp), web chat, email, custom — permissions NOT auto-granted on channel connection

Domain 220% of exam

Prompt Engineering

Must-Know Facts

Prompt Builder is the Salesforce tool for creating REUSABLE prompt templates — not a general text editor. Templates must be explicitly activated before they can be used by agents or users
Three template types: Sales Email (generates email drafts for sales reps, scoped to email content), Field Generation (auto-fills a SPECIFIC record field with AI output, tied to one object + one field), Flex (general-purpose with custom inputs, not tied to a single field)
Template activation is a REQUIRED separate step — a created template that is not activated cannot be used. Creation and activation are two distinct lifecycle stages
Three grounding techniques: merge fields (raw record field values pulled at runtime), Flow merge fields (data processed/calculated by a Flow before insertion), RAG from Data Cloud (semantic search across unstructured knowledge bases)
Merge fields pull data at RUNTIME — the data is dynamic and specific to the record context at the time the template executes, not static at template creation time
Prompt templates can serve as ACTIONS in Agent Builder — a Prompt Template action is one of the five action types an agent can execute
Best practices for reducing hallucinations: ground prompts in relevant data, set a specific output format, include guardrails, provide clear and specific instructions — vague prompts increase hallucination risk
Each AI feature (including an individual Agentforce agent) can use only ONE Data Library for RAG grounding at a time
User roles in Prompt Builder control who can create, edit, activate, and use templates — not everyone with Salesforce access can manage prompt templates. This is a common governance oversight
Field Generation templates can encounter token limit errors when record data is variable-length or excessively long — Flex templates handle variable-length inputs more gracefully in these cases

Common Traps

TrapField Generation templates can be used to generate any type of content

RealityField Generation is tied to a SPECIFIC object and SPECIFIC field. It produces output saved to that field. If the use case is not about populating a specific record field with AI output, use Flex

TrapOnce you create a prompt template, it is immediately available for use

RealityTemplates must be ACTIVATED as a separate step after creation. An inactive template exists in the system but cannot be invoked by agents or users. Exam questions test this lifecycle distinction explicitly

TrapMerge fields and Flow merge fields are interchangeable — both pull Salesforce data

RealityMerge fields = raw field values directly from a record (no transformation). Flow merge fields = data that has been processed, aggregated, or calculated by a Flow before insertion. Use Flow merge fields when the data requires any calculation, aggregation, or cross-object lookup

TrapAdding more context to a prompt always improves it

RealityExtremely long prompts can confuse the model and reduce response quality. Best practice is specific, relevant grounding + clear format instructions + guardrails — not exhaustive context dumps. Prompt length has a sweet spot

TrapFlex templates are a drop-in replacement for Field Generation templates because Flex is more flexible

RealityField Generation and Flex solve different problems. Field Generation auto-populates a specific field on a specific object and saves output directly to that record. Flex is for custom outputs not bound to a single field. They are not interchangeable

TrapA Prompt Builder template is only used directly by human users — agents use their own internal prompts

RealityPrompt Template is one of the five action types available in Agent Builder. Agents can execute prompt templates as part of their action sequence. This is how AI-generated content is triggered as part of agent workflows

Confusing Pairs

Sales Email TemplateFlex Template

Sales Email = specifically generates email drafts for sales reps, tightly scoped to email content, appears in Draft Email actions on record pages. Flex = general-purpose, accepts up to five custom inputs, produces varied output for any use case not covered by other types. Key: 'generate an email for a sales rep' → Sales Email. 'Generate a summary, briefing, or custom text output' → Flex

Field Generation TemplateFlex Template

Field Generation = tied to one object + one field, AI output is saved directly to that field on the record, token limit tied to field configuration. Flex = custom inputs at creation, outputs not saved to a specific field, handles variable-length inputs better. Key: 'auto-fill this specific field on the record' → Field Generation. 'Custom flexible output not mapping to a single field' → Flex

Merge Fields (grounding)Flow Merge Fields (grounding)

Merge fields = pull raw values directly from a record's fields at runtime ({!Account.Name}). Flow merge fields = data is first processed by a Flow and the result is inserted (sum of related records, formatted date ranges, aggregated lookups). Key: 'raw field value from a record' → merge field. 'Calculated or aggregated data' → Flow merge field

Merge Field GroundingRAG Grounding (Data Cloud)

Merge fields = structured, known data from specific fields on a known record. RAG = searches across a large unstructured knowledge base using semantic similarity to find the most relevant context dynamically. Key: 'pull this specific field value into the prompt' → merge field. 'Search knowledge articles or uploaded documents for relevant context' → RAG

Scenario Tips

If the question asks about:

When the question asks about automatically generating a text summary on an Opportunity record using data from that specific record...

Answer:

Field Generation template in Prompt Builder. It is designed to auto-fill a specific field on a record with AI-generated content

Distractor to avoid:

Flex template is tempting because it is 'flexible' — but Flex is for cases NOT tied to a specific field. Field Generation is the precise tool when the output maps to a single record field

If the question asks about:

When the question describes needing data from multiple related child records (e.g., sum of line item amounts or a count of open cases) included in a prompt...

Answer:

Use Flow merge fields. The Flow aggregates or calculates the related data, then the Flow merge field inserts the computed result into the prompt template

Distractor to avoid:

Standard merge fields sound close but they only pull raw field values from a single record. They cannot aggregate or compute across related records — that requires a Flow

If the question asks about:

When a Field Generation template is working for most records but failing with token limit errors on records with unusually long text fields...

Answer:

Switch to a Flex template. Flex handles variable-length inputs more gracefully because its inputs are defined with custom controls and it is not bound to a fixed object/field structure

Distractor to avoid:

Adding more merge fields to the Field Generation template is the wrong direction — more merge fields add tokens, which increases the chance of hitting the limit, not resolves it

If the question asks about:

When the question asks about making a prompt template available for agents or users to actually use in their workflows...

Answer:

Activate the template. Activation is a required step after template creation. An unactivated template exists in the system but cannot be invoked regardless of user permissions

Distractor to avoid:

Assigning the template to a user or publishing it sounds logical but activation is the specific gate — this is a Prompt Builder lifecycle concept the exam tests repeatedly

Last-Minute Facts

1Domain weight: 20% — roughly 12 questions

2Template types: Sales Email (email drafts for sales reps), Field Generation (fills specific record field, tied to object+field), Flex (general-purpose, up to 5 custom inputs)

3Activation is required before a template can be used — creation alone is not enough

4Merge fields: raw record data at runtime. Flow merge fields: data processed by a Flow first

5Each AI feature / agent can use only ONE Data Library for RAG grounding

6Prompt templates can be used as actions in Agent Builder — one of five action types

7Reducing hallucinations: ground in relevant data + specific output format + guardrails in instructions

Domain 320% of exam

Data Cloud for Agentforce

Must-Know Facts

Data Cloud was rebranded to Data 360 in late 2025 — same product, different name. The exam may use either name; treat them as identical
Data Library is the storage layer that manages knowledge sources for agent grounding. Each AI feature (including an individual Agentforce agent) can use only ONE Data Library at a time
Data Library sources: Salesforce Knowledge articles, uploaded files (text, HTML, PDFs with size limits), and Data Cloud objects. Not all data types are supported — structured CRM records use merge fields, not the Data Library
Chunking: breaking unstructured documents into semantically meaningful pieces for indexing. Chunk size and field selection choices directly impact retrieval quality — this is manually configured, NOT automatic
Indexing: creating vector embeddings (numerical representations) from chunked content. Which fields you include during indexing determines what the retriever can find — wrong field selection = irrelevant results
Three search index types: keyword (exact lexical matching, precise for numbers and codes), vector (semantic similarity via embeddings, understands meaning), hybrid (combines both keyword precision with vector understanding)
Enriched indexes: a vector or hybrid index enhanced with automatically extracted metadata, entities, and question-answer pairs to improve RAG retrieval accuracy — an ENHANCEMENT on top of an existing index type, not a fourth index type
Two retriever types: individual (queries one index, simpler, faster) vs ensemble (combines results from multiple indexes, broader coverage, higher complexity and latency). Use ensemble ONLY when data is spread across multiple sources
Only ONE retriever version is active at a time — multiple versions can exist but only one is live. Updating a retriever changes query rules; changing what is indexed requires updating the index itself
RAG has two phases: offline (chunking and indexing documents into the vector store — happens once when data is added or updated) and online (retrieving relevant chunks at query time to ground the prompt — happens every time a user interacts with the agent)

Common Traps

TrapAn agent can use multiple Data Libraries for different topics

RealityEach AI feature (agent) can use only ONE Data Library. However, a single Data Library CAN contain multiple data sources (Knowledge + PDFs + Data Cloud objects). The constraint is one library per agent, not one source per agent

TrapEnsemble retrievers are always better than individual retrievers

RealityEnsemble retrievers are more comprehensive but add complexity and latency. Use them ONLY when relevant data genuinely spans multiple source indexes. For single-source scenarios, individual retrievers are simpler, faster, and preferred

TrapVector search is the best choice for all retrieval scenarios

RealityVector search understands semantic meaning but STRUGGLES with exact numbers, model codes, product SKUs, and domain-specific terms. For technical documentation containing both product codes and conceptual explanations, hybrid search is correct

TrapUpdating a retriever configuration changes what data is retrievable

RealityRetriever configurations define query rules and do NOT modify the underlying index. To change what is retrievable, you must update the search INDEX itself (re-configure and re-index). These are two entirely separate configuration layers

TrapData Cloud and Data 360 are separate products

RealityData 360 is just the new brand name for Data Cloud after the late 2025 rebranding. Functionally identical. Either name on the exam means the same technology

TrapChunking happens automatically when you add a data source to a Data Library

RealityChunking and field selection must be deliberately CONFIGURED by an administrator. Poor chunk size or wrong field selection leads to irrelevant retrieval — this is the most common root cause of agents returning bad answers from a Data Library

TrapEnriched indexes are a third index type alongside keyword and vector

RealityEnriched is an ENHANCEMENT applied to an existing vector or hybrid index — it adds automatically extracted metadata, entities, and question chunks. You cannot create an 'enriched-only' index; you first create vector or hybrid, then enrich it

Confusing Pairs

Keyword SearchVector Search

Keyword = exact lexical matching, precise for product codes, numbers, and specific terms — has no semantic understanding. Vector = semantic similarity via embeddings, understands meaning and context — struggles with exact terms. Key: 'find exact model number SKU-123' → keyword. 'Find content about troubleshooting connectivity problems' → vector

Vector SearchHybrid Search

Vector only = great for meaning, poor for exact terms. Hybrid = combines keyword precision WITH vector semantic understanding. Key: when content has BOTH specific terms (codes, numbers, proper nouns) AND conceptual text (explanations, procedures) → hybrid is the right answer

Individual RetrieverEnsemble Retriever

Individual = queries one search index, simpler to configure, lower latency. Ensemble = queries multiple indexes and merges ranked results, needed when relevant data is spread across multiple Data Library sources. Key: 'data in one source' → individual. 'Data spread across Knowledge AND uploaded files' → ensemble

ChunkingIndexing

Chunking = breaking documents into semantic pieces (happens first, offline phase). Indexing = creating vector embeddings from those chunks and storing them (also offline phase, happens after chunking). Both happen offline before any user query. Changing field selection during indexing requires re-indexing

Data LibraryData Cloud Data Streams

Data Library = storage layer specifically for agent GROUNDING knowledge sources (Knowledge articles, PDFs, uploaded files). Data Cloud data streams = general data ingestion pipeline for unifying customer data. They serve different purposes and are not interchangeable in exam scenarios

Scenario Tips

If the question asks about:

When the question describes an agent that needs to search product documentation containing both model numbers (SKU-4892) and conceptual troubleshooting steps...

Answer:

Configure hybrid search. It combines keyword precision (for exact model numbers) with vector semantic understanding (for troubleshooting concepts). Neither keyword alone nor vector alone handles both requirements

Distractor to avoid:

Vector search alone sounds like the modern AI answer but it struggles with exact product codes and model numbers — a common trap in Data Cloud questions

If the question asks about:

When an agent is returning irrelevant results from the Data Library despite having a well-configured retriever...

Answer:

Investigate chunking configuration and field selection during indexing. Poor chunk sizes or wrong indexed fields are the most common causes of irrelevant retrieval results

Distractor to avoid:

Trust Layer blocking sounds plausible but would prevent responses entirely, not return irrelevant ones. Permission errors cause access failures, not bad content

If the question asks about:

When the question asks an agent to pull data from both a Salesforce Knowledge base AND a set of uploaded PDF product manuals...

Answer:

Configure an ensemble retriever. It combines results from multiple indexes so data from both Knowledge articles and PDFs can be searched and ranked together

Distractor to avoid:

An individual retriever pointing to Knowledge alone would miss the PDF data entirely. You need ensemble when data genuinely spans multiple sources

If the question asks about:

When the question asks about what happens when you update the retriever configuration after seeing irrelevant search results...

Answer:

Updating the retriever configuration only changes query rules. To fix what is actually retrievable, you must update the search INDEX itself — reconfigure which fields are indexed and re-run the indexing process

Distractor to avoid:

Editing the retriever config is the wrong layer — retriever and index are separate. The exam tests this two-layer architecture distinction

Last-Minute Facts

1Domain weight: 20% — roughly 12 questions

2Data Cloud = Data 360 (rebranded late 2025) — exact same product

3One Data Library per AI feature/agent — hard limit, but one library can have multiple data sources

4Search types: keyword (exact), vector (semantic), hybrid (both combined). Enriched = vector or hybrid + auto-extracted metadata/entities — not a separate type

5Retrievers: individual (1 index, faster) vs ensemble (multiple indexes, needed for multi-source scenarios)

6Only ONE retriever version active at a time

7RAG offline phase: chunk + index. Online phase: retrieve + ground prompt

8Updating retriever config does NOT change the index — these are separate configuration layers

9Chunking must be manually configured — it is NOT automatic when adding data sources

Domain 420% of exam

Development Lifecycle

Must-Know Facts

Agent lifecycle stages: ideation, building, testing (Testing Center), deployment, and observation (Analytics + Utterance Analysis). The cycle is ITERATIVE — observation feeds back into ideation and building improvements
Agentforce Testing Center: PRE-deployment tool for BATCH testing entire agent configurations. Tests topic selection accuracy, action execution correctness, and overall agent behavior at scale before going live
Utterance Analysis: POST-deployment tool that analyzes REAL production conversations to identify where the agent struggled with intent, gave poor responses, or missed new topic opportunities
Agentforce Analytics: POST-deployment monitoring providing adoption metrics, usage patterns, deflection rates, escalation rates, and response quality insights. Built on Data Cloud
Digital Wallet: tracks CONSUMPTION/COST of Data Cloud and Agentforce usage across both sandbox and production environments. It is NOT a performance or quality metrics tool
Sandbox environments support Agentforce development and UAT. The Einstein Trust Layer audit trail is active in sandboxes — security testing works in sandbox just like in production
Agent configurations must be EXPLICITLY DEPLOYED from sandbox to production — they do not sync automatically. Treat agent configurations as deployable metadata
Agentforce Observability: the broader umbrella that includes Agentforce Analytics PLUS live health monitoring, step-by-step agent reasoning traces (watching the agent reason), and consumption tracking

Common Traps

TrapThe Agentforce Testing Center and Utterance Analysis are both pre-deployment validation tools

RealityTesting Center = PRE-deployment (validate configurations in sandbox using synthetic test cases). Utterance Analysis = POST-deployment (analyze real production conversations that happened after go-live). The time distinction is critical — exam questions use 'before going live' vs 'after deployment' to signal which tool to choose

TrapDigital Wallet measures agent performance metrics like accuracy and response quality

RealityDigital Wallet measures CONSUMPTION (resource usage and cost tracking across environments). Performance metrics come from Agentforce Analytics. Cost and usage tracking → Digital Wallet. Performance and quality → Analytics

TrapSandbox and production agent configurations are automatically synchronized

RealityThey are completely separate environments. Configurations built in sandbox must be explicitly deployed to production. Nothing moves automatically — this is the same as any Salesforce metadata deployment. Exam tests knowledge of this explicit step

TrapTesting Center tests individual prompts or individual actions, not the full agent

RealityTesting Center validates the ENTIRE AGENT CONFIGURATION — including topic selection logic, action sequencing, and end-to-end agent behavior — across batch test scenarios. It is not a single-prompt or single-action tester

TrapUtterance Analysis automatically creates and adds new topics to the agent

RealityUtterance Analysis surfaces conversation insights and problem areas — it is an analysis tool. Humans then use those insights to DECIDE whether to create or modify topics. It surfaces opportunities; it does not auto-configure the agent

Confusing Pairs

Agentforce Testing CenterUtterance Analysis

Testing Center = PRE-deployment, validates synthetic test cases at batch scale in sandbox, confirms topic/action correctness before going live. Utterance Analysis = POST-deployment, analyzes real user conversations in production, identifies where the agent struggled. Key: before deployment → Testing Center. After deployment → Utterance Analysis

Agentforce AnalyticsDigital Wallet

Agentforce Analytics = performance and adoption metrics (accuracy, usage patterns, deflection rates, response quality). Digital Wallet = consumption and cost metering (how much AI resource is being used and what it costs). Key: 'how well is the agent performing?' → Analytics. 'How much am I spending on AI consumption?' → Digital Wallet

Sandbox Testing (environment)Agentforce Testing Center (tool)

Sandbox = the environment for development and manual UAT testing (where you work). Agentforce Testing Center = a specific tool that runs AUTOMATED BATCH validation of agent configurations at scale inside that environment. Sandbox is the place; Testing Center is the tool you use inside it

Agentforce AnalyticsAgentforce Observability

Agentforce Analytics = adoption and accuracy metrics dashboard (deflection rates, usage trends, quality scores) — a reporting view for performance. Agentforce Observability = the broader umbrella including Analytics PLUS live health monitoring, step-by-step reasoning traces, and consumption tracking. Key: 'performance metrics report' → Analytics. 'Watch the agent reason in near-real-time or comprehensive monitoring' → Observability

Scenario Tips

If the question asks about:

When the question asks about identifying conversations where the agent failed to understand customer intent AFTER deployment...

Answer:

Utterance Analysis. It analyzes post-deployment production conversations to surface intent failures and topic coverage gaps

Distractor to avoid:

Testing Center is tempting because it sounds like 'testing' — but it is a PRE-deployment validation tool. After the agent is live in production, Utterance Analysis is the correct tool

If the question asks about:

When the question asks about validating that an agent correctly selects topics for 200 test scenarios BEFORE going live...

Answer:

Agentforce Testing Center. It enables batch testing of complete agent configurations before production deployment, covering topic selection and action execution across many test cases simultaneously

Distractor to avoid:

Utterance Analysis sounds like it involves testing conversations but it only works with real production conversations that happen AFTER deployment — it cannot be used pre-deployment

If the question asks about:

When the question asks about tracking how much the Agentforce deployment is costing across both sandbox and production environments...

Answer:

Digital Wallet. It provides consumption metering and cost visibility across all environments including sandbox and production

Distractor to avoid:

Agentforce Analytics is about performance quality and adoption, not cost. Observability dashboards show health and reasoning, not spend

If the question asks about:

When the question describes using observations from production conversations to improve the agent and feed back into agent configuration...

Answer:

Utterance Analysis provides the post-deployment insights. The development lifecycle is iterative — Utterance Analysis observations should feed back into the ideation and building phases for configuration improvements

Distractor to avoid:

Testing Center is for pre-deployment validation of new configurations, not for gathering production feedback that drives improvement

Last-Minute Facts

1Domain weight: 20% — roughly 12 questions

2Lifecycle stages (in order): ideation → build → test (Testing Center) → deploy → observe (Analytics + Utterance Analysis) → back to ideation

3Testing Center: PRE-deployment, batch test synthetic cases, validates topic selection and action execution

4Utterance Analysis: POST-deployment, real production conversations, identifies intent failures and new topic opportunities

5Digital Wallet: consumption and cost tracking only — NOT performance metrics

6Agentforce Analytics: adoption metrics, accuracy, deflection rates, usage patterns — NOT cost

7Sandbox ≠ production — configurations must be explicitly deployed, no automatic sync

8Trust Layer audit trail IS active in sandbox environments — security testing works pre-production

Domain 55% of exam

Multi-Agent Interoperability

Must-Know Facts

Model Context Protocol (MCP) is an OPEN STANDARD originally created by Anthropic — not a Salesforce-proprietary protocol. Salesforce adopted it for agent-to-external-tool communication
MCP architecture: client-server model. Agentforce acts as the MCP CLIENT. External tools expose MCP SERVERS. The agent calls the MCP server to interact with external systems
MCP actions in Agentforce execute WITHIN the Einstein Trust Layer boundary — PII is masked before reaching any LLM, even through external MCP connections. The Trust Layer is not bypassed
AgentExchange: Salesforce's curated marketplace of VETTED MCP servers and pre-built agent components. No-code deployment through Agent Builder. It is NOT an open marketplace — servers are security-reviewed before listing
Agent API: programmatic interface for EXTERNAL SYSTEMS to invoke Agentforce agents. Used when a custom application outside Salesforce needs to call an agent. This is INBOUND (external → Agentforce)
MCP direction: agent calls external tools (OUTBOUND, Agentforce → external). Agent API: external calls the agent (INBOUND, external → Agentforce). These are opposite directions

Common Traps

TrapMCP is a Salesforce-developed protocol for connecting Salesforce agents to external tools

RealityMCP was created by Anthropic and is an open industry standard. Salesforce adopted it — they did not create it. The exam specifically tests this distinction because it matters for multi-vendor agent ecosystem questions

TrapMCP connections bypass the Einstein Trust Layer because they route data to external systems

RealityMCP actions execute WITHIN the Einstein Trust Layer boundary. PII masking applies before anything reaches an LLM through MCP. The Trust Layer is enforced at the Salesforce side before data leaves the Trust Layer perimeter

TrapAgentExchange is like AppExchange — anyone can publish an MCP server to it

RealityAgentExchange servers are VETTED and CURATED by Salesforce. They are reviewed for security and reliability before listing. It is not an open marketplace — the curation is a deliberate enterprise trust and security feature

TrapAgent API is used for agent-to-agent communication within Salesforce

RealityAgent API is for EXTERNAL programmatic access — custom applications OUTSIDE Salesforce invoking Agentforce agents. Internal agent collaboration within Salesforce uses different mechanisms. Agent API = inbound external access

Confusing Pairs

AgentExchangeAppExchange

AppExchange = general Salesforce marketplace for apps, components, and solutions (open to vetted third-party publishers, broad catalog). AgentExchange = curated marketplace specifically for vetted MCP servers and agent components, security-reviewed before listing. Key: 'install a Salesforce app or managed package' → AppExchange. 'Add a pre-built MCP server connection for an agent' → AgentExchange

MCP (Model Context Protocol)Agent API

MCP = protocol for Agentforce agents to connect TO external tools/systems (OUTBOUND, agent as client, agent initiates). Agent API = interface for external applications to connect TO Agentforce agents (INBOUND, agent as server, external system initiates). Key: 'agent needs to use an external tool' → MCP. 'External app needs to invoke an agent programmatically' → Agent API

Scenario Tips

If the question asks about:

When the question asks how to connect an Agentforce agent to an external project management tool to create tasks, without writing custom code...

Answer:

Use AgentExchange to find and deploy a vetted MCP server for that tool. AgentExchange provides no-code deployment of pre-built MCP connections through Agent Builder

Distractor to avoid:

Salesforce Connect sounds similar but is for external DATA SOURCE connections (reading data from external systems), not for agent tool interactions or taking actions in external systems

If the question asks about:

When the question asks what ensures sensitive customer data is protected when an Agentforce agent uses MCP to interact with an external system...

Answer:

The Einstein Trust Layer masks PII before data reaches any LLM through MCP. Trust Layer is applied within the Agentforce side of the pipeline, not bypassed by the MCP connection

Distractor to avoid:

The external system's own security is a real consideration but is NOT the Salesforce mechanism that ensures data protection. Trust Layer is the answer the exam expects

If the question asks about:

When the question asks which company originally developed the Model Context Protocol...

Answer:

Anthropic created MCP. Salesforce adopted and implemented it. This is a factual question testing whether candidates know MCP is an open standard, not a Salesforce invention

Distractor to avoid:

Salesforce sounds like the correct answer because they prominently use and promote MCP — but they adopted it, they did not create it

Last-Minute Facts

1Domain weight: 5% — roughly 3 questions. Small weight but potentially easy points — do not skip this domain

2MCP created by Anthropic, adopted by Salesforce — open standard, not Salesforce-proprietary

3MCP actions run INSIDE the Einstein Trust Layer boundary — PII masking still applies through MCP

4AgentExchange = vetted/curated MCP server marketplace — NOT open like AppExchange. Salesforce security-reviews all entries

5MCP direction: Agentforce → external tool (outbound, agent as client)

6Agent API direction: external app → Agentforce agent (inbound, agent as server)

7MCP architecture: Agentforce = MCP client. External tools = MCP servers

Feeling confident?

Put your knowledge to the test with a timed AI-201 mock exam.