Services / Salesforce / Agentforce
Agentforce

The platform for building and running AI agents.

Agentforce is Salesforce's platform for AI agents that handle customer and employee work around the clock and hand over to people when needed. Unlike a scripted chatbot, an agent interprets a request, plans, and chooses from the tools it has.

How an agent is built →Trust and guardrails

Salesforce's Agentforce page describes its employee assistant as formerly Einstein Copilot.

Illustration of an agent builder, simplified. Sample data.
What it does

Agents that act, within limits you set.

Agents analyse data, make decisions and take actions on tasks such as answering service questions, qualifying sales leads and optimising marketing campaigns. Salesforce describes five attributes that define one.

01Role
The agent's purpose.
02Knowledge
The data the agent needs in order to do its job well.
03Actions
What the agent is able to do: the goals it can fulfil.
04Guardrails
The guidelines the agent operates within.
05Channels
The places where the agent does its work, such as CRM, WhatsApp, Messenger or your website.
How it works

How an agent is built, tested and run.

Four views of Agentforce, from Salesforce's own pages and guides.

An agent is made of topics and actions

From Salesforce's implementation guide for a service agent.

  • Topics are the jobs the agent can do. A topic has a classification description (the kinds of request it covers), a scope (the limits of the job) and instructions (how to do it).
  • Actions are the tools the agent uses, such as getting or creating a case. They can be extended with Flow, Apex and prompt templates.
  • Newer Salesforce material also describes subagents, actions and instructions in Agent Builder.
What you’re looking at

A job description

Salesforce suggests writing a topic's scope like a job description: what the agent can do and what it must not.

Illustration of a screen, simplified. Sample data.
Trust and guardrails

How Salesforce keeps agents on a lead.

From Salesforce's trusted-AI and Agentforce pages. Salesforce calls this the trust layer, and one of its own pages calls it the Einstein Trust Layer.

01Secure data retrieval and grounding
The model works from your company data, such as Data 360 or internal knowledge, and people see only what their permissions allow.
02Zero data retention and data masking
Salesforce says prompts and responses are not stored or used to train third-party models, and that personal or proprietary details are swapped for placeholders before reaching the model.
03Toxicity detection and guardrails
Generated content is scanned, and flagged output is filtered, blocked or marked before the user sees it.
04Guardrails on the agent
Safeguards you define and protections Salesforce manages keep agents to their instructions, block off-topic or inappropriate conversations and reduce hallucinated or biased answers.
05An audit trail
Salesforce's how-it-works page lists an audit trail of agent actions and outputs for governance reviews.
06People stay in control
Deploying and activating a new version are separate steps, and Salesforce says long-running agents work under human oversight.

Salesforce's pages differ slightly in which of these they list, so check the page for the feature you rely on.

Getting started

Start simple, then build, test, repeat.

Advice from Salesforce's guide to building a service agent.

  1. 01Begin with a templateStart from a standard topic and its standard actions, then customise.
  2. 02Keep instructions minimalBegin with one instruction and add more as you simulate conversations. The guide's advice: less is more.
  3. 03Put business rules in actionsLogic that is specific to you, such as refund rules, belongs in a flow-based action, not in a free-text instruction.
  4. 04Be explicitAgents take words such as always and never literally, and do not rely on the visual order of instructions.
  5. 05Fix the weakest linkIf the answer is wrong, check the topic chosen, the action chosen, the action's output and then the tone, in that order.

Decisions to settle before you buy

  • Whether the agent serves customers or employees.
  • How predictable your usage is, which decides between consumption and per-user models.
  • How much commitment you want: prepaid, a baseline, or pay as you go.
  • How you will track consumption and set alerts.
Pricing

How Agentforce is paid for.

Salesforce describes consumption and per-user models. Rates change, so this page does not quote them.

OptionWhat it is
Flex CreditsPay per action an agent performs, from a shared pool used across teams and channels. Voice actions cost more than standard ones. Unused credits do not carry over to a later term.
ConversationsPay per customer conversation, which suits agents that face customers. It cannot be used in the same org as Flex Credits.
Per-user modelsAn Agentforce user licence for employees, add-ons for Sales, Service, Field Service and industries with unmetered employee use, and a separate licence for Agentforce Coworker.
Entry pointSalesforce Foundations is a free tier that includes the builder tools.

Agentforce pricing on salesforce.com

© 2026 Mawkish Technologies. This guide, its diagrams and illustrations are protected by copyright. You’re welcome to learn from it and to quote it with credit and a link. Reproducing it, by hand or through AI, and publishing it as your own is not allowed. Read our copyright notice

Questions

Straight answers.

What is Agentforce?

Salesforce's platform for building, deploying and monitoring AI agents that work around the clock across channels and hand over to people when needed.

How is an agent different from a chatbot?

A scripted chatbot follows a fixed path. An agent interprets the request, makes a plan and chooses the tools it has, and it can be started by a data change as well as by a person.

How do we test an agent before it goes live?

Salesforce describes generating test cases, simulating conversations in a planner and checking how well guardrails are followed, in a sandbox, because agents can change CRM data.

Is our data used to train other companies' models?

Salesforce says prompts and responses are not stored or used to train third-party models, and that personal details are masked before a prompt reaches the model.

Can we mix pricing models?

Flex Credits and Conversations cannot be used in the same org. Per-user add-ons are a separate option.

Who builds agents?

Admins and developers. Agent Builder is low-code, and agents can also be built in code.

Checked against Salesforce's Agentforce pages, architects' guidance and implementation guide in November 2026. The implementation guide is from 2024 and refers to topics; newer pages also say subagents.

Across Salesforce

Go to one of the five main products.

Each of the five has its own page, and you can visit them in any order. Choose one to go there.

You are currently viewing Agentforce, a supporting product

Also on the Salesforce platform

How the products fit together

Talk to us

Talk it through with us.

Tell us what you want Salesforce to do for you, and we will help you work out where to start.

  • Your goals and the systems involved
  • Which Salesforce products fit, and which can wait
  • A sensible first step

Salesforce, Agentforce, Data 360, Slack, Tableau, MuleSoft and other Salesforce product names are trademarks of Salesforce, Inc. Other names are the property of their owners. Screens shown are illustrations with sample data. Product descriptions follow Salesforce's own published material and can change, so check salesforce.com for current details.

Our Ravbrook Loop is the name we give to how we handle Salesforce Agentforce: one accountable partner, from the first meeting to long after go-live.

© 2026 Mawkish Technologies. This guide, its diagrams and illustrations are protected by copyright. You’re welcome to learn from it and to quote it with credit and a link. Reproducing it, by hand or through AI, and publishing it as your own is not allowed. Read our copyright notice