Before you start: prerequisites

Before touching Agent Builder, confirm you have the following in place. Skipping this step is the most common reason implementations stall in week 3.

  • Agentforce license: Confirm with your Salesforce AE that Agentforce for Service is provisioned on your org. Check Setup → Company Information → feature licenses.
  • Service Cloud: Agentforce Service Agent requires Service Cloud. If you are on Sales Cloud only, you will need to add Service Cloud licenses.
  • Einstein setup: Go to Setup → Einstein → Einstein Setup and confirm Einstein is enabled. Agentforce runs on the Einstein platform.
  • Knowledge base: Your agent will be far less useful without a populated knowledge base. You need at least 20-30 published knowledge articles covering your most common customer questions before go-live.
  • Data quality check: Run a quick audit on the objects your agent will touch — Cases, Contacts, Accounts. Fields like Status, Priority, Contact Email, and Account Name should be 90%+ populated.

Step 1: Access Agent Builder

Navigate to Setup → search for "Agents" → click Agents. You will see the Agent Builder interface — Salesforce's no-code/low-code tool for configuring Agentforce agents.

Click New Agent and select Service Agent as the template. This gives you a pre-built starting point with common service actions already wired up — significantly faster than building from scratch.

Step 2: Define your agent's identity

Every Agentforce agent has a role description — a natural language prompt that defines what the agent is, what it does, and what it should never do. This is the most important configuration you will write.

A well-written role description looks like this:

"You are a customer service agent for [Company Name]. Your job is to help customers resolve issues with their accounts, orders, and subscriptions. You have access to case history, order records, and the knowledge base. You can create cases, update case status, process refunds up to $100, and provide order status. You cannot access billing information beyond the last 4 digits of a payment method. If a customer is angry or requests to speak with a manager, transfer immediately to a human agent."

Notice what this does: it defines scope (what the agent handles), capabilities (what actions it can take), limits (what it cannot access), and escalation triggers (when to hand off).

Step 3: Configure topics

Topics are the categories of requests your agent can handle. Think of them as the agent's specializations. Each topic has its own instructions and action library.

For a typical B2B service agent, start with these topics:

  • Order Status: Customer wants to know where their order is or when it will arrive
  • Account Management: Password resets, contact information updates, subscription changes
  • Technical Support: Product issues, error messages, troubleshooting
  • Billing Enquiries: Invoice questions, payment status, refund requests
  • General Information: Product questions answered from the knowledge base

For each topic, write a clear description — the LLM uses this description to decide which topic applies to an incoming request. Vague topic descriptions cause misclassification. Be specific.

Step 4: Build your action library

Actions are what the agent can actually do. Agentforce comes with a library of standard actions — querying records, creating cases, sending emails — and you can build custom actions using Flow or Apex.

For each action, you need to define:

  • Name and description: The LLM reads this to decide when to invoke the action. Write it like an API description — precise and unambiguous.
  • Input parameters: What information does the action need? Be explicit about required vs. optional fields.
  • Output: What does the action return to the agent? The agent uses this to formulate its response.

Start with 5-7 core actions rather than trying to build everything at once. The most valuable actions for a service agent are typically: Get Case History, Create Case, Update Case Status, Search Knowledge Base, Get Order Status, and Process Simple Refund.

Step 5: Set up channels

Where will your agent interact with customers? Agentforce supports messaging channels (web chat, SMS, WhatsApp), email, and voice. For a first deployment, we recommend starting with web chat only — it is the easiest to test and the most forgiving of early-stage agent behavior.

Go to Setup → Messaging Settings → create a new Messaging Channel and associate it with your agent. You will get an embed code to add to your website.

Step 6: Configure handoff rules

This is where most first-time implementations underinvest. Handoff rules define when your agent transfers a conversation to a human — and getting this wrong in either direction is costly.

Too aggressive handoffs: the agent escalates everything and your human agents see no volume reduction. Too few handoffs: customers get stuck in loops with an agent that cannot actually resolve their issue.

The handoff triggers we configure for most clients:

  • Customer explicitly requests a human agent
  • Agent confidence score falls below threshold (configure in Einstein settings)
  • Issue has been open for more than 10 minutes without resolution
  • Customer expresses frustration (sentiment detection)
  • Request involves an action outside the agent's permission scope
  • VIP account detected (check account tier field)

Step 7: Test before you launch

Agent Builder has a built-in testing interface — use it extensively before any live traffic. Run through at minimum:

  • Your 20 most common case types from the last 90 days
  • 5 edge cases that previously caused escalations
  • 3 adversarial tests (try to get the agent to do things outside its scope)
  • 2 tests with incomplete or ambiguous information

Document every failure. For each one, decide whether to fix it via role description refinement, a new guardrail, or a new handoff condition. Do not launch until your pass rate on the 20 common cases is above 90%.

Step 8: Phased launch

Never route 100% of traffic to a new agent on day one. Use Salesforce's routing rules to send a controlled percentage:

  1. Day 1-2: 10% of new chat sessions routed to agent. Review every transcript.
  2. Day 3-5: 25% if no critical failures. Review flagged transcripts only.
  3. Week 2: 50%, then 100% if deflection rate and CSAT are on target.

Set up a simple dashboard in Salesforce to track: deflection rate (% resolved without human), escalation rate, average handle time, and CSAT. These four metrics tell you everything you need to know about agent performance in the first 30 days.

Common mistakes to avoid

  • Vague role descriptions: "Help customers with their questions" is not a role description. Be specific about scope, capabilities, and limits.
  • Too many topics at launch: 3-5 well-defined topics outperform 15 loosely defined ones every time.
  • Skipping knowledge base prep: An agent with a poor knowledge base will hallucinate. Invest in your articles before you invest in your agent.
  • No monitoring setup: If you are not watching deflection rate and CSAT from day one, you will not catch regressions before they affect customers.
  • Big bang launch: Always phase your rollout. The cost of a bad launch far exceeds the cost of a slower ramp.