Create Your First Agentforce Agent — A Step-by-Step Guide

AI agents in Salesforce are powerful — but to make them work, you need a structured approach. This guide gives you 10 clear steps (with bullet-point actions) to create your very first Agentforce Service Agent.

What you’ll build (example)

• A Service Agent that:

  • Answers FAQs

  • Fetches account/order/case details from CRM

  • Takes actions (e.g., Create Case, Send Email, Book Appointment)

  • Escalates when confidence is low

Prerequisites (5-minute checklist)

• Enable Agentforce and the Default Agent in Setup

• Assign the correct permission sets (e.g., Agentforce Service Agent Builder)

• Use a Sandbox environment with masked/seeded test data

• Confirm Einstein Trust Layer is active for data security

Step-by-Step: Build Your First Agent

1) Define the agent’s job (outside Salesforce)

• Do this in a planning doc (Google Doc, Confluence, Jira epic, or BRD/AI Playbook)

• Run a workshop with business + IT stakeholders

• Capture:

  • Purpose: What problems it solves (e.g., answer FAQs, fetch orders)

  • Guardrails: What it must not do (e.g., never share PII)

  • KPIs: How success is measured (containment %, CSAT, AHT, escalation rate)

• Output: A one-pager “Agent Job Description” to guide the build

2) Create the agent in Salesforce (2 minutes)

• Go to Setup → Agentforce Agents

• Click New Agent

• Select Service Agent template

• Enter Name + Description

• Click Create

3) Connect the agent to data (10–20 minutes)

• Identify the CRM objects it needs (Accounts, Contacts, Cases, Orders)

• Give field-level security only for required fields

• (Optional) Link Data Cloud for unified customer profiles

• Ensure least privilege access — avoid overexposing sensitive data

4) Add actions (15–30 minutes)

• Choose 3–5 safe starter actions (Flows/APIs), e.g.:

  • Create Case

  • Update Case

  • Send Email

  • Book Appointment

  • Log Call

• Use Flow for low-code, auditable automation

• Add approval steps for sensitive actions

5) Shape the agent’s behavior (10–15 minutes)

• Write a purpose/system prompt (tone, role, limits, escalation rules)

• Define response style (e.g., concise bullets + next action)

• Narrow scope to one domain for first release

• Example starter purpose prompt:

  • “You are the Service Agent for [Brand]. You answer FAQs, retrieve order/case details, and can only perform these actions: [list]. If confidence < 0.6, ask clarifying questions. Never reveal PII.”

• 
  

6) Apply trust & compliance guardrails (5–10 minutes)

• Confirm Einstein Trust Layer is enabled

• Verify field-level security & data sharing rules

• Turn on logging of prompts/outputs (privacy-safe)

• Review data compliance policies before launch

7) Test the agent in Builder (15–30 minutes)

• Use Preview/Test mode in Agent Builder

• Run scenarios:

  • Happy path: “Where is my order 12345?”

  • Edge case: “Delete my data”

  • Policy test: “Give me John’s phone number” (should refuse)

• Check that guardrails hold and outputs match the purpose

8) Evaluate & iterate (ongoing)

• Track key metrics:

  • Containment rate

  • Avg. handle time

  • CSAT

  • Escalation rate

• Refine prompts (add examples, tighten scope)

• Add one new action at a time

• Keep a version log of changes

9) Deploy to users (10 minutes)

• Start with a pilot group (small support team)

• Deploy in Salesforce Console first

• Share a 1-pager training guide

• Run a 10-minute huddle to explain when to use/escalate

10) Plan week-2 enhancements

• Add a new safe action (e.g., Update Case)

• Extend training set (20 resolved cases as examples)

• Connect one external API (through Flow with guardrails)

• Monitor KPIs to validate improvements

Wrap-up

By breaking down the process into these 10 bullet-pointed steps, you avoid scope creep and reduce risk. The magic lies in Step 1 — defining the agent’s job before touching Salesforce. Once you build, keep iterating, measuring, and expanding safely.