Agentforce implementation guide: how to implement Agentforce, step by step

This guide started life as SynconAI's Agentforce eBook in early 2025. It has been updated for the current product names and moved here, and it keeps the same eleven-part shape: what Agentforce is, what it can do, where it fits, how to run the project, how the security model works, a step-by-step Setup walkthrough, then tailoring, comparison, what comes next, and how to start.

If you want the phase-by-phase runbook with an exit criterion for every gate, read implementing Agentforce: the sequence, and the gate between each step alongside this. If you would rather have the work done with you, that is our Agentforce implementation service.

#1. What is Agentforce?

Agentforce is Salesforce's platform for building and running AI agents. It became generally available in October 2024, and Salesforce now groups its AI portfolio under the Agentforce 360 name. An Agentforce agent reads a request in everyday language, works out which of its topics the request belongs to, calls the actions it is permitted to use, and either completes the work or hands it to a person with a summary of what it found.

It differs from a scripted bot in one important way. A bot follows dialog paths someone drew in advance. An agent plans the steps at run time, inside a boundary you configure: the topics it covers, the instructions it follows, the actions it may call and the data its running user can see.

Its role in the Salesforce ecosystem

Agentforce sits on top of the platform you already run rather than beside it:

• It adds AI to existing Salesforce clouds, so a service agent works on the same Case and Knowledge records your people use.
• It gives one place to build AI automation across functions, instead of a different bot per team.
• It connects to the wider Customer 360 data model, and to Data 360 where an agent needs unified customer context.
• It is designed to assist people in sales, service and marketing, with escalation built in, not to replace the person who owns the outcome.

Why companies adopt it

The reasons in the original guide still hold, with one caution attached to each:

• Routine work handled automatically, so people spend time on the cases that need them. The caution: only routine work that is well documented.
• A better customer experience, through answers that use the customer's own records. Only as good as those records.
• Round-the-clock coverage without adding shifts, for the intents the agent is allowed to handle.
• Decisions grounded in Salesforce data, provided that data is current and has one owner per fact.
• Scale that does not grow headcount in step with volume, once the agent has proven itself on a narrow scope.

Where SynconAI fits

SynconAI is a Salesforce Select Partner and OpenAI Select Partner that designs, implements and runs Agentforce for teams in the United States and Australia. The work covers readiness assessments, implementation plans, integration with existing Salesforce orgs, training and change management, and ongoing tuning after launch.

#2. Key features and capabilities

Agentforce is a set of building blocks rather than a single feature. The ones that matter in almost every implementation:

Integration with Salesforce and other apps

Inside Salesforce, Agentforce works natively with Sales Cloud, Service Cloud, Field Service and Marketing, uses your existing records and processes, and calls the Flows you have already built. Outside Salesforce, actions can call APIs and MuleSoft integrations to reach ERP, billing and other SaaS systems, and employee agents can work in Slack. Connections to tools such as Microsoft Teams or Google Workspace depend on the connector and edition, so confirm them before you design around them.

AI, automation and reporting

• Atlas Reasoning Engine. The reasoning layer that decides which topic and actions fit a request and in what order.
• Retrieval augmented generation. The agent retrieves approved content first and writes the answer from it, which is why the quality of that content matters more than the prompt.
• Deterministic automation. Steps that must not vary, such as a refund calculation, belong in a Flow or Apex action that the agent calls. The agent decides; the action does.
• Analytics. Agentforce reporting covers conversation volume, escalations and quality, alongside the dashboards you already run on Cases and Opportunities.

The original guide listed "adaptive learning" as a feature. It needs a correction: an Agentforce agent does not quietly retrain itself on your conversations. It improves because people review transcripts and change its instructions, topics, actions and grounding content. Plan for that review work.

#3. Real-world applications

Sales

• Lead qualification against criteria your sales team already uses, so reps work the best prospects first.
• Tailored outreach and follow-up drafted from the account history, sent by a person.
• Meeting preparation: a summary of the account, open opportunities and recent cases before a call.
• Objection handling from approved answers, so new reps sound like experienced ones.
• Meeting scheduling without the back-and-forth email.

Our Agentforce for sales teams article covers the patterns in more depth.

Customer service

• Always-on answers for the questions your Knowledge base already answers well.
• Triage and routing that sends complex inquiries to the right person with context, while the agent handles the routine ones.
• Case summaries and drafted replies for human agents.
• Support in several languages where the content and the model support it.
• Knowledge gaps surfaced from questions the agent could not answer, which is often the most useful report in the first month.

For service specifically, see Agentforce customer service automation.

Field operations

• Real-time access to customer, asset and product details for technicians.
• Appointment booking and rescheduling through Field Service.
• Technician preparation: job history and likely parts before arrival.
• Guided troubleshooting from approved repair content.
• Job summaries written up after the visit for the technician to confirm.

Industry examples

• Manufacturing: order status, shipping and product specification questions answered from ERP data through integration actions; service and maintenance requests routed with asset context.
• Retail and e-commerce: order tracking, returns questions and product guidance from the catalog and order systems.
• Healthcare: appointment reminders, administrative questions and intake support. Any agent near clinical or protected health data needs a specific compliance review before design starts.

A longer list, ranked by how ready teams usually are, is in Agentforce use cases.

#4. Best practices for an Agentforce implementation

The steps the original guide set out are the right ones. Here they are with what each one has to produce.

1. Define clear objectives. Pick the business goal, the measures you will judge it by, and how it supports the wider strategy. Output: a one-page brief with the measures and a list of what is out of scope.
2. Assemble a cross-functional team. IT, the business owner of the process, the people who do the work today, and someone from risk or compliance. Name one project lead. Output: a RACI.
3. Assess needs honestly. Map the current process and its pain points, then rank use cases by value, data readiness and risk. Output: a shortlist with one use case chosen to prove first.
4. Prepare your data. Audit the knowledge and records the agent will answer from, remove contradictions, give each fact one owner, and set governance so it stays that way. Output: one authoritative source per intent.
5. Start with a pilot. One use case, one audience, a realistic timeline and agreed exit criteria. Output: a pilot plan with a rollback.
6. Train people. Role-based training for the people who will work beside the agent and for the admins who will tune it. Output: owners who can change it safely.
7. Roll out gradually and keep tuning. Add channels and audiences in phases, monitor quality and escalations, and adjust. Output: a backlog and a review cadence.

Common mistakes to avoid

• Treating change management and adoption as an afterthought.
• Underestimating data preparation and integration work. This is the most common reason pilots stall; see Agentforce implementation mistakes.
• Setting expectations from a demo rather than from what your editions allow today.
• Launching without a scored test set, which leaves nothing to compare against after the next change.
• No named owner after go-live.

Keeping performance on track

• Review quality, escalation rate and handoff health on a set cadence.
• Re-run the regression set after every change to instructions, topics, actions or the model.
• Keep grounding content current: an out-of-date Knowledge article becomes a confident wrong answer.
• Plan capacity for consumption costs as volume grows, using the current Salesforce rate card. Our Agentforce pricing and implementation cost article explains the cost lines.

#5. Data privacy and security in Agentforce

Agentforce runs inside Salesforce's security model, and that model is where most of the real decisions sit.

The Einstein Trust Layer

Salesforce describes the Einstein Trust Layer as a sequence of gateways and retrieval mechanisms between the model and your users. The controls it names are:

• Secure data retrieval, which respects the permissions of the user the agent runs as.
• Dynamic grounding, which adds relevant, permitted data to the prompt so answers come from your sources.
• Zero data retention with the model providers, so prompts and responses are not kept by them.
• Toxic language detection on generated content.
• An audit trail of prompts, responses and feedback.

The original guide listed personal-detail masking among these. Salesforce now notes that data masking for LLMs is currently disabled for agents, so do not design an agent on the assumption that it applies. Confirm which controls apply to each agent you build.

What the Trust Layer does not decide

An agent sees what its running user can see under the sharing model, plus whatever knowledge and content visibility rules expose. Those are two separate paths with separate configuration. If the running user is a copy of a broad human profile, the agent inherits that breadth. Build a dedicated running user with a narrow permission set, and review knowledge visibility separately. Our Agentforce security best practices article goes further.

Regulation

Salesforce publishes its own compliance programs for regimes such as GDPR, CCPA and HIPAA. Whether your agent meets them depends on your configuration, your contracts with Salesforce and the data you let it touch. Involve your privacy and compliance owners at design time, not at go-live.

Securing data day to day

• Run Salesforce Health Check and review the agent's permission sets on a schedule.
• Classify data by sensitivity and keep the most sensitive fields out of the agent's reach unless a use case needs them.
• Require a named human approval inside the action for refunds, deletions, credits and entitlement changes.
• Monitor event logs and set alerts for unusual activity.
• Keep backups and a tested rollback for agent configuration.

#6. Step-by-step Agentforce setup

This walkthrough follows the Salesforce Trailhead quick start the original guide was built on, which uses the Coral Cloud sample org and an Agentforce Service Agent. Setup labels move between releases, so match each step by meaning if a label differs in your org. Do all of it in a sandbox or a Trailhead playground first.

Step 1: prepare

• Confirm your editions and Agentforce entitlements with your Salesforce account team.
• Decide which permission sets the agent's running user will need, and no more.
• Create a dedicated sandbox for building and testing.

Step 2: turn on Einstein and Agentforce

1. Open Setup from the gear icon.
2. In Quick Find, search for Einstein Setup and turn on Einstein.
3. Refresh the browser.
4. In Quick Find, search for Agents and turn on the Agentforce toggle.

Step 3: publish the Experience Cloud site (service agents on a portal)

If the agent will answer customers on an Experience Cloud site, the site must be published:

1. In Setup, open All Sites.
2. Click Builder next to the site (in the Trailhead example, coral-cloud).
3. Click Publish, then confirm.

Step 4: assign permission sets to the agent user

1. In Setup, open Users and select the agent user (in the example, EinsteinServiceAgent).
2. Under Permission Set Assignments, click Edit Assignments.
3. Add the agent permission set (in the example, Service Agent Permissions) and save.

This is the step that decides what the agent can see. Keep it narrow.

Step 5: create the agent

1. In Setup, open Agents and click New Agent.
2. Choose Agentforce Service Agent and click Next.
3. Select the starting topics (for example, General FAQ).
4. Name the agent (for example, CC Service Agent).
5. Select the agent user from Step 4, then click Next and Create.

Step 6: configure topics and actions

The agent opens in Agentforce Builder, previously called Agent Builder.

1. Click New, then New Topic, and fill in the topic details (for example, Experience Management): the scope, the instructions, and what is out of scope. Click Next, then Finish.
2. Select the new topic. Under This Topic's Actions, click New and choose Add Action.
3. Choose the action type, for example Flow, and select the Flow (for example, Get Experience Details).
4. Map the inputs and outputs, then click Finish.

Write the instructions as you would brief a new colleague: what the topic is for, what to ask for before acting, and when to hand off.

Step 7: test, then activate

1. Use the conversation preview in Agentforce Builder to try real requests, including ones the agent should refuse. Read the reasoning trace to see which topic and action it picked.
2. Build a test set from real requests in Testing Center and record the results, so every later change can be compared against them.
3. Adjust instructions, topics and actions until the results meet the measures you agreed in your objectives.
4. Click Activate, then connect the agent to its channel for a limited audience first.

Beyond the first agent

Once the first agent is stable, the work shifts to identifying the core actions each team needs, documenting topics and workflows so they can be owned, and tailoring agents for specific industries and roles. The runbook in implementing Agentforce covers the gates for that stage.

#7. Tailoring Agentforce to your business

Industry and role

• Industry terms and processes. Instructions, grounding content and actions written in the language your industry uses, with its regulations in mind.
• Role-based agents. Separate agents or topics for service, sales, field and internal teams, each with its own running user, rather than one agent that can do everything.
• Brand and voice. Instructions that set tone and the phrases the agent should and should not use, checked in testing alongside accuracy.

APIs and automation

• API integration through Apex, external services and MuleSoft to reach systems outside Salesforce.
• Multi-step automation with conditional logic in Flow, called by the agent when a request needs it.
• AI-driven routing and decisions that remain reviewable: the agent proposes, the rules and the approvals decide.

For the boundary between the two, see Agentforce vs Salesforce automation.

Working across Salesforce clouds

• Sales Cloud: qualification, meeting preparation and opportunity updates.
• Service Cloud: triage, case summaries, routing and self-service. See Agentforce and Service Cloud.
• Marketing: audience and content assistance on Marketing Cloud Next, where licensed.
• Data 360: unified customer context when an answer depends on records spread across several systems. See grounding Agentforce with Data 360.

#8. How Agentforce compares with other options

Where it stands out

• Native to Salesforce. It uses the records, permissions, Flows and Knowledge you already have, rather than syncing a copy to another platform.
• Reasoning plus control. The agent handles unstructured requests, while actions keep the steps that must not vary deterministic.
• Extensible. Apex, APIs and MuleSoft let it reach systems outside Salesforce.

Strengths and weaknesses

The honest weaknesses: complex tailoring has a learning curve, consumption costs need forecasting, and an agent still needs human oversight and a named owner. Where the paths can be listed in advance, a Flow is usually the better answer.

#9. What comes next for Agentforce

Salesforce ships Agentforce changes every release, so treat roadmap items as announcements until they are generally available in your edition. The directions the original guide anticipated are the ones that have continued:

• Better language understanding, with longer, more complex conversations and broader language coverage.
• Voice and multi-modal interaction, with voice already offered where licensed.
• More autonomy with more control, as agents take on longer multi-step work and Salesforce adds tools to test and observe it.
• Deeper data, as Data 360 brings more of a customer's context to the agent at the moment it answers.

SynconAI's part in this is practical: tracking each release against what clients can actually run, building industry patterns we can reuse, and training client teams so they can own and extend what we build with them.

#10. Conclusion and next steps

Key takeaways

• Agentforce brings AI agents into the CRM you already run, working on the same records and permissions your people use.
• A successful implementation is mostly planning, data preparation, testing and ownership.
• Security and compliance are decided by configuration you can review: the running user, sharing, knowledge visibility and approvals.
• Tailoring and integration let agents fit specific industries and teams.
• Capability will keep growing release by release, so build the habit of testing every change.

How to get started

1. Assess and plan. Run a readiness assessment and agree a plan with clear goals and measures.
2. Implement and train. Build the first use case in a sandbox, train the people who will own it, and pilot it with a limited audience.
3. Tune and grow. Review results on a cadence, gather feedback, and add topics and channels once the measures hold.

Further learning

• Salesforce documentation and Trailhead: the Agentforce modules, including the quick start this walkthrough follows and the Einstein Trust Layer module.
• SynconAI Insights: the Agentforce grounding checklist, Agentforce implementation best practices and the articles linked above.
• The community: Salesforce user groups and the Trailblazer Community.

Work with SynconAI

SynconAI is a Salesforce Select Partner and OpenAI Select Partner delivering Agentforce consulting and implementation across the United States and Australia. Delivery is architect-led, starts with one use case and a pilot with exit criteria, and the team that builds the first agent stays on to run it. See our Agentforce consulting and implementation service, or compare engagement options on our pricing page.