Hands-on AI + Salesforce builds with real-life scenarios — each project tells you the problem, the stack and exactly what you'll learn.
Build, don't just read
Every brief links to a full hands-on article. More builds land as the AI push continues.
Scenario: A team wants to ask plain-English questions about live CRM data without opening Salesforce. Activate a standard Hosted MCP server, wire an External Client App over OAuth, and connect an AI assistant as a custom connector — with object permissions, field-level security and sharing rules enforced on every tool call.
You'll learn: Hosted MCP vs DX MCP, scoping access by choosing the right standard server, and how to verify an answer is genuinely grounded in org data.
Scenario: The standard servers give an assistant your records, not your logic. Create a custom Hosted MCP server in Setup and publish an autolaunched Flow — one that returns an Account with its related Contacts and Cases — as a tool any MCP client can discover and call, with no code at all.
You'll learn: Creating and activating a custom server, publishing a Flow as a tool, and writing the tool description that decides whether it ever gets called.
Scenario: An assistant will happily ask for every account in the org. Add a parameterised Apex invocable action to the custom server that takes industry, revenue and open-case criteria, clamps what it will return, enforces CRUD and FLS, and hands back a computed health summary rather than raw records.
You'll learn: Designing a tool contract through annotations, defending against unbounded agent requests, and making failure an explanation the assistant can act on.
Scenario: A rep wants a renewal-risk brief without opening Salesforce. Ground a Flex prompt template with an Apex data provider that resolves an Account by name and gathers its contacts, open cases and open opportunities, then publish it on a custom MCP server twice — as a user-invoked prompt and behind a global invocable wrapper so the model can call it mid-conversation.
You'll learn: Why an MCP prompt alone never fires, writing grounding text for a model rather than a parser, and making a lookup miss loud enough that the LLM refuses to invent.
Scenario: Admins spend hours translating business rules into validation formulas. Build an Agentforce-powered tool where you type "Opportunity amount must be over 1000 for Closed Won" and it generates and deploys the rule via the SOAP API.
You'll learn: Prompt-to-formula generation, deploying metadata programmatically, and guarding AI output before it touches production.
Scenario: A manufacturer's customers live in a Next.js portal, not Experience Cloud. Embed an Agentforce agent behind the portal's chat widget using the Agent API — sessions, streaming responses and Case transcripts included.
You'll learn: Client-credentials OAuth, the Agent API session lifecycle, and keeping Salesforce credentials server-side.
Scenario: A retailer's returns agent needs live data from SAP and a courier API. Expose both through MuleSoft MCP servers so the agent discovers and calls them as governed tools — no custom integration actions.
You'll learn: The MCP tool contract, enterprise MCP registry governance, and swapping backends without touching the agent.
Scenario: Support answers live in hundreds of PDFs and knowledge articles. Ground an agent on that content with Agentforce Data Library so it answers from real sources instead of hallucinating.
You'll learn: How Data Library auto-builds chunks, search indexes and retrievers, and how to write grounded prompts.
Scenario: Wire the open-source @salesforce/mcp server into your AI coding assistant so "run the failing tests and explain the coverage gap" executes against your scratch org — a real agentic-DevOps workflow.
You'll learn: MCP client configuration, safe org toolsets, and where AI genuinely speeds up Salesforce development.
Scenario: Refunds must follow policy — no exceptions, no LLM improvisation. Build a refund topic where Agent Script enforces eligibility checks deterministically and the LLM only handles the conversation.
You'll learn: Hybrid reasoning, deterministic transitions, and testing probabilistic agents in the Testing Center.
Scenario: Tier 2 engineers lose 10 minutes per escalation reading case history. Auto-generate a structured handoff summary with a grounded Prompt Builder template fired from a record-triggered Flow — zero code.
You'll learn: Grounded prompt templates, invoking them from Flow, and designing AI output people actually trust.
Scenario: Sales wants a one-paragraph account brief before every call — on the record page, in a nightly digest and as an agent action. Build it once in Apex with the Models API and reuse it on all three surfaces.
You'll learn: Calling LLMs through the Einstein Trust Layer from Apex, grounding with USER_MODE queries, and reuse patterns.
Scenario: Most support tickets are the same handful of "where's my order / how do I reset this" questions. Stand up an Agentforce service agent on an Experience Cloud help site and Messaging, grounded on Knowledge, that resolves those itself and hands off to a human through Omni-Channel when it can't.
You'll learn: Deploying an agent to web and messaging channels, grounding answers on Knowledge, and designing a clean escalation-to-human handoff.
Scenario: An agent that can only talk is half a product — the value shows up when it can act. Build a custom Agentforce action that logs a warranty claim through a Flow, with typed inputs, validation and a confirmation step, so the LLM never writes junk into your records.
You'll learn: Designing agent actions with input/output types, choosing Flow vs invocable Apex, and guarding writes so AI stays inside policy.
Scenario: Sales ops fields "what's the status of the Acme renewal?" in Slack a hundred times a day. Deploy an employee-facing Agentforce agent in Slack, grounded on CRM data through Data 360, so the team self-serves answers without ever opening Salesforce.
You'll learn: Employee vs customer agent design, deploying an agent to Slack, and grounding it on live CRM data with Data 360.
Scenario: Speed-to-lead wins deals, but reps can't answer every web inquiry in minutes. Configure an Agentforce SDR agent that engages inbound leads over email, answers product questions from grounded content and books meetings — escalating qualified leads to a human seller.
You'll learn: Configuring autonomous outreach guardrails, grounding responses on approved content, and handing off qualified leads to sellers.