Integration Services
Business Consulting & Advisory
AI Solutions
AppExchange Development
Tailored Salesforce Engineering
Managed Services
Implementation
Key Takeaways
- Salesforce Agentforce is the AI agent layer of the Agentforce 360 Platform. Its agents can autonomously handle approved tasks across Salesforce and connected systems, with escalation paths for exceptions.
- Agents work in a loop: they identify the request, retrieve data, choose an action, and check the result. Agent Script locks critical steps to fixed business rules.
- Einstein Copilot was renamed Agentforce (Default) in the Spring ‘25 release with no change in functionality. It is now one agent inside a larger platform, not the whole product.
- Data 360 grounds agent answers in your company’s own trusted records, and agents inherit the Salesforce permissions and sharing rules already in place.
- The Einstein Trust Layer adds zero data retention with third-party model providers, prompt injection and toxicity screening, and an audit trail that admins can turn on in Data 360.
- Agentforce pricing combines consumption models (Flex Credits or per-conversation fees) with per-user add-on licenses, so cost depends on whether agents serve customers or employees.
- Salesforce’s 2026 releases added Agentforce Coworker for employees, job-ready agents, and a long-horizon runtime (in pilot) for goals that span days or weeks.
What makes an autonomous AI agent different from the chatbot your company deployed a few years ago?
The short answer is action. A chatbot answers questions. An agent completes work. Salesforce Agentforce is the AI agent layer of Salesforce’s Agentforce 360 Platform. Its agent’s reason over business data, decide the next step, and can autonomously handle approved tasks across Salesforce and connected systems, with escalation paths for exceptions.
Adoption is now measured in completed work, not just signed deals. In its second quarter fiscal 2027 results, published August 26, 2026, Salesforce1 reported 7.0 billion Agentic Work Units delivered to date across Agentforce and Slack. Agentic Work Units are Salesforce’s measure of discrete tasks completed by AI agents. This guide explains what Salesforce Agentforce is, how its agents’ reason and act, and how the platform has changed since its 2024 launch. It also covers where Einstein Copilot fits today, what Agentforce costs in 2026, and which trust and governance controls keep agents inside approved limits.
Table of Content
- Everything You Need to Know About Salesforce’s Agentforce
- What is Salesforce Agentforce?
- Are Agentforce and Einstein Copilot the Same?
- From Generative AI to Autonomous AI Agents
- How Does Agentforce Work?
- What Can Agentforce Do? Business Use Cases in 2026
- How Do You Build an Agentforce Agent?
- How Much Does Agentforce Cost in 2026?
- Trust and Governance: How Does Agentforce Keep AI Safe?
- What Does the Future Hold for Salesforce Agentforce?
- The Final Word
What is Salesforce Agentforce?
Agentforce is Salesforce’s platform for building, deploying, and managing autonomous AI agents that reason over business data, decide the next step, and act across business workflows. Its agents can autonomously handle approved tasks, with escalation paths for exceptions. They can also start work when an event occurs, such as a new case or an updated record, instead of waiting for someone to type a prompt.
Salesforce’s product architecture places Agentforce in a stack of connected layers, all running on its Trust Layer:
- Agentforce 360 Platform (Agentforce Builder and the agent runtime): where teams build, test, and run agents with Agentforce Builder, Agent Script, and the reasoning engine.
- Data 360: unifies customer and business data, so agents ground their answers in trusted company records.
- Customer 360 apps: the sales, service, marketing, and commerce applications where agents carry out tasks.
- Slack: the shared workspace where employees ask questions and agents to take action.
Since Dreamforce 2026, a new interface layer called Airforce sits above these systems. It brings Salesforce data, workflows, and agents into Claude (Claude force), Slack (Slack force), and the Lightning interface (Agentforce Coworker), so people can work with agents from the tools they already use. For example, consider a customer who wants to return an item they recently bought and get a refund. It’s a busy Wednesday and all your customer service reps are occupied. An Agentforce service agent checks the order data and the company’s return policy, issues the refund if it falls within approved limits, and escalates the case to a human representative when the request falls outside those guardrails. This separates Agentforce from traditional chatbots, which can only respond to direct questions.
Salesforce claims the platform can deflect up to 72%2 of routine customer questions while improving resolution times and customer satisfaction scores. Gartner named Salesforce a Leader in its 2026 Magic Quadrant for Conversational AI Platforms. For a closer look at the platform and what a Service Agent rollout involves, read our guide to what an Agentforce Service Agent implementation takes.
Agentforce Version History: Major Releases From 2024 to 2026
Agentforce has changed quickly since its 2024 launch. Salesforce ships Agentforce changes in three seasonal releases each year (Spring, Summer, and Winter). It announces its biggest capabilities at two events: TDX in the spring and Dreamforce in the fall. The table lists each major milestone, the newest first.
| Date | Release | What it added |
|---|---|---|
|
Sep 2026 |
Dreamforce 2026: AIforce |
An interface layer that brings Salesforce data and agents into Claude (Claudeforce, beta), Slack (Slackforce), and Lightning (Agentforce Coworker) |
|
Sep 2026 |
Dreamforce 2026: agents and models |
Seven job-ready agents, including Fin (six GA, Hunter GA planned for Nov 2026); a long-horizon runtime and multi-agent orchestration; the Koa CRM reasoning model (pilot); Gemini in the Agentforce Reasoning Engine (GA); Agent Optimizer (rolling out from Oct 2026) |
|
May 2026 |
Agentforce Coworker |
An AI teammate in the Lightning search bar that routes requests to deployed agents (beta in May; auto enabled for eligible orgs from Aug 2026) |
|
Apr 2026 |
TDX 2026: Headless 360 |
Salesforce capabilities exposed as APIs, MCP tools and CLI commands for agents; Agentforce Experience Layer; Agent Script released as open source |
|
Oct 2025 |
Dreamforce 2025: Agentforce 360 |
Agentforce Builder, Agent Script, Agentforce Voice, Data 360, Slack as the agent interface; Data Cloud renamed Data 360 |
|
Jun 2025 |
Agentforce 3 |
Command Center for agent observability, native MCP support, AgentExchange expansion |
|
Mar 2025 |
Agentforce 2dx |
Proactive agents embedded in workflows, developer tools, multimodal experiences |
|
Feb 2025 |
Spring ’25 release |
Einstein Copilot renamed Agentforce (Default) |
|
Dec 2024 |
Agentforce 2.0 |
Prebuilt skills library, Slack agents, MuleSoft integrations, upgraded Atlas Reasoning Engine |
|
Oct 2024 |
Agentforce GA |
First generally available release, with Agentforce Service Agent and SDR agent |
|
Sep 2024 |
Dreamforce 2024 launch |
Agentforce announced as Salesforce’s agentic AI platform |
In short, Agentforce launched in October 2024 and became the home of the former Einstein Copilot assistant in early 2025. In October 2025, it became the agent layer of Agentforce 360. In April 2026 Headless 360 exposed platform capabilities to agents as APIs and MCP tools, and in September 2026 the platform gained the AIforce interface layer.
Are Agentforce and Einstein Copilot the Same?
No. Agentforce is not just a new name. Einstein Copilot was Salesforce’s conversational assistant. In Spring ’25, it became Agentforce (Default), one agent within the wider Agentforce platform for building and running many agents. Many people are still confused between Agentforce and Copilot, so it helps to look at how the two relate today.
1. How Einstein Copilot Fits into Agentforce Today
Salesforce renamed Einstein Copilot for Salesforce to Agentforce in February 2025. Salesforce’s decision to rename the product comes as part of a broader effort to streamline its AI offerings. The official release notes confirmed there were no functionality changes, only updated UI elements and permission labels.
Salesforce stated that “this change won’t impact your implementations. Your agent is still available to help your Salesforce users with everyday business interactions, embedded right in the flow of work.” Custom actions, permissions, and integrations built for the Einstein Copilot agent type kept working under the Agentforce label without changes.
2. Where Einstein, Prompt Builder, and Model Builder Sit Now
Einstein remains the name behind Salesforce’s predictive AI features, such as predictions and recommendations, some of which date back to 2016. Agentforce is now the umbrella for Salesforce’s agentic and generative AI, and the Einstein name persists in a few core components.
The most important of these is the Einstein Trust Layer, which kept its name and sits underneath Agentforce rather than beside it. Every generative call an Agentforce agent makes passes through it, so organizations use both because one runs inside the other.
Prompt Builder lets admins and developers create custom, reusable AI prompts without coding. Model Builder connects to different AI models or builds predictive models trained on Data 360 data. Both tools remain active and are used inside Agentforce agents.
3. Why Salesforce Moved from Copilots to Agents
At Dreamforce 2023, Salesforce announced Einstein Copilot Studio, which includes Copilot, Prompt Builder, and Model Builder. However, they didn’t bring the growth Salesforce expected. One of the reasons could be that many businesses aren’t ready for AI.
In fact, in its Dreamforce 2024 keynote3, Salesforce shared that over 82 billion flows are launched each week compared to only 122,000 prompts. Flow has been around for years, so the comparison isn’t perfect. But on average, there’s less than one prompt per Salesforce customer per week.
That slow uptake of prompt-based tools shaped what came next. Generative features now sit under Agentforce and Prompt Builder, while the Einstein name persists in the Trust Layer and Einstein Studio. The shift reflects Salesforce’s move from assisted AI that waits for human prompts to autonomous agents that detect events, reason about them, and take approved actions.
Unlock the Power of Agentforce in Your Salesforce Ecosystem
From Generative AI to Autonomous AI Agents
Generative AI produces content when prompted. AI agents go further: they reason over a goal, choose actions, and complete approved tasks across systems. They escalate to people when a case falls outside their guardrails. This shift from producing content to completing work is what Salesforce and others call agentic AI.
The difference grew out of large language models with natural language processing capabilities, starting in 2018 with BERT and OpenAI’s GPT. These models could understand open-ended instructions and plan, but they worked only with text. They could not change systems or carry out tasks. In short, generative AI could answer questions, but it could not take action.
Autonomous AI agents close that gap. They combine multi-step reasoning with tool use and ongoing execution. They keep context over time and handle errors by adjusting their approach. What sets them apart is their ability to reason, review outcomes, adapt plans, and pursue goals within approved limits. To see how this works step by step, read how Agentforce agents reason and act across enterprise workflows.
| Aspect | Generative AI assistant | Copilot | Autonomous AI agent |
|---|---|---|---|
|
Trigger |
Waits for a human prompt |
Waits for a human prompt in the workflow |
Detects events or conditions |
|
Scope |
Single-turn response |
Multi-turn conversation within one task |
Multi-step workflow across systems |
|
Actions |
Generates text, code, images |
Retrieves data, surfaces insights |
Updates records, fires automations, calls APIs |
|
Human role |
Reviews and applies output |
Approves suggested actions |
Sets guardrails, handles exceptions and escalations |
Agentforce spans all three columns. Prompt Builder covers generative tasks, and Employee agents work alongside people inside Salesforce. What distinguishes the platform is the third column: its agents can detect events, reason about them, and take approved actions, while people stay in charge of the rules and the exceptions.
How Does Agentforce Work?
Agentforce agents run a continuous reasoning loop: classify the user’s intent to a subagent, retrieve relevant data, reason about the next action, carry out that action, observe the result, and repeat until the goal is met or a handoff occurs. It operates on Salesforce’s core platform, so it works with the tools your business already uses.
Plan Each Step with the Atlas Reasoning Engine and Agent Script
Atlas is the orchestration layer between large language models and Salesforce tools. It uses Reasoning and Acting (ReAct) prompting, running a loop of reason, acting, and observing until the user’s goal is met. Because the model explains its reasoning before it selects an action, this approach reduces hallucinations.
What sets Agentforce apart is that it makes interactions feel human-like through natural language understanding. When users input a query, Agentforce doesn’t just look for keywords; it understands the intent behind the message.
Agent Script gives builders deterministic control over agent behavior by blending natural language instructions with programmatic expressions. Builders define if/else conditions, transitions, and variables in script, so critical workflows follow exact business policies instead of relying on the LLM’s interpretation alone.
Ground Answers in Trusted Data with Data 360
Agentforce agents ground their responses in structured and unstructured data through retrieval-augmented generation. Data 360 provides semantic search across customer records, so agents can retrieve the right context without anyone writing precise queries.
Grounding sources include standard Salesforce objects, custom objects, external objects via Salesforce Connect, and Agentforce Data Libraries. Data Libraries automate setup steps such as pushing data streams to Data 360, mapping objects, and creating search indexes. Throughout, agents respect the permissions of the user or process that invoked them, and field-level security and sharing rules still apply.
Actions: Flows, Apex, APIs, Prompt Templates and MCP Tools
Actions are how agents get work done. Agentforce turns existing platform capabilities into agent actions: autolaunched flows handle multi-step record operations, invocable Apex classes manage complex logic, prompt templates handle generative tasks, and MuleSoft APIs connect external systems.
Developers expose Apex methods to agents with the @InvocableMethod annotation. Flow-based actions and Apex invocations both accept dynamic prompts and optional session IDs to support multi-turn conversations. Model Context Protocol (MCP) support works in both directions: agents can call tools on registered MCP servers, and external AI assistants such as Claude or ChatGPT can invoke Agentforce.
Channels: Slack, Web and Mobile, Agentforce Voice, and Headless Access
Once built, agents can be deployed on several surfaces. Users mention agents in Slack channels or direct messages and get results in the full conversation context. Agentforce Voice handles phone interactions using natural language processing and recognizes when to escalate a call to a human.
The Agent API allows headless agent calls with no user interface. External systems call agents through synchronous HTTP requests or asynchronous server-sent events that stream responses. Headless agents can also run background processes triggered by Apex or scheduled jobs.
Human Handoff and Escalation Paths
Every channel needs a clear path back to a person. Dynamic escalation enforces business rules before a conversation moves to a human representative. Builders control when handoffs happen through escalation settings on each subagent and escalation instructions, and they can require the agent to collect information or complete specific actions before escalating.
Contextual handoffs pass the conversation history, relevant customer data, and the reason for escalation to the human agent. Representatives get the full picture instead of starting the interaction over.
What Can Agentforce Do? Business Use Cases in 2026
Agentforce agents can support work across service, sales, marketing, commerce, operations, field service, and employee support. Salesforce offers prebuilt agents for common business jobs, and teams can customize them in Agentforce Builder to match their data, policies, and workflows.
| Business function | Problem it solves | Typical tasks | Salesforce-supported agent or capability |
|---|---|---|---|
|
Customer service |
High case volume, after-hours support |
Returns, refunds, order status, feedback, complaints, and service requests |
Agentforce Service Agent, Agentforce Voice, Fin |
|
Sales |
Slow lead follow-up, thin pipeline |
Lead qualification, outreach, helping sales representatives prioritize their follow-ups |
Piper (inbound lead qualification), Hunter (outbound prospecting), Sales Coach |
|
Employee productivity |
Time lost searching and updating records |
Finding answers, updating records, summarizing accounts |
Agentforce Coworker, Agentforce in Slack |
|
HR and IT |
Repetitive policy and access requests |
Inquiries about benefits, policies, and onboarding processes; password resets or software installations |
Employee agents (Paige) |
|
Commerce |
Guided shopping and order support |
Answering product questions, checking stock levels, and assisting with order processing |
Agentforce Commerce agents |
|
Marketing |
Slow campaign build |
Building campaigns from a brief |
Campaign Agent in Marketing Cloud Next |
|
Back office and supply chain |
Manual systems work spanning multiple tools |
Supply chain management, inventory tracking, and logistics |
Marshall, Systems Integration Agent |
|
Field service |
Technicians away from screens |
Hands-free job updates and lookups |
Agentforce Voice: Field Voice in the Salesforce mobile app |
|
Industries |
Sector data and compliance needs |
Patient inquiries, appointment scheduling; account inquiries, transaction details, and loan applications |
Agentforce for Financial Services, Health, Automotive, Consumer Goods |
How Do You Build an Agentforce Agent?
The Agent Builder is where Agentforce agents are built. It’s a low-code tool that allows you to create custom agents tailored to your specific business needs. No matter the level of customization, it’s easy to get started. First, you need to define the purpose of your agent followed by a set of instructions.
Agentforce lets you add unlimited instructions to your agents. The more specific you are, the more effective your agent will be. Look at this set of detailed instructions:
As you can see, this is a Salesforce Agentforce service agent being developed using simple instructions. Once done, you must set up the actions you want the agent to take:
And that’s it! You have your own customer service agent created in minutes!
However, what sets Agentforce apart is that it makes interactions feel conversational and human-like through natural language understanding. When users input a query, Agentforce doesn’t just look for keywords; it understands the intent behind the message.
How Much Does Agentforce Cost in 2026?
Agentforce offers four main pricing models, so organizations can match how they pay to how they deploy agents. Prices below reflect Salesforce’s published list rates as of October 2026.
| Model | Best When | Main Trade-Off |
|---|---|---|
|
Flex Credits |
$500 per 100,000 credits |
Usage-based agent actions |
|
Per conversation |
$2 per conversation |
External customer-facing agents |
|
Agentforce add-ons (Sales, Service, Field Service) |
$125 per user per month |
Unmetered employee use |
|
Agentforce Industries add-ons |
$150 per user per month |
Industry clouds |
|
Agentforce 1 Editions |
From $550 per user per month |
Bundled add-on plus 2.5M Flex Credits per org per year |
Flex Credits are the consumption unit for Agentforce actions. Each standard action uses 20 credits and each Agentforce Voice action uses 30 credits, which works out to $0.10 and $0.15 per action at $500 per 100,000 credits.
Conversation-based pricing charges $2 per conversation for customer-facing agents. A conversation covers up to two hours on messaging channels and 10 minutes on voice. Help Agent resolutions cost $2 (or 400 Flex Credits), with no extra charge for the number of actions or Data 360 queries in that session. Flex Credits and per-conversation pricing can’t be used in the same org.
For employees, the Agentforce User License costs $5 per user per month but requires Flex Credits purchased separately. Add-on licenses at $125 or $150 per user per month include unmetered usage for covered workflows.
Flex Credits can be bought in three ways. Pre-Purchase covers usage upfront for the full contract term. Pre-Commit sets a baseline with monthly billing in arrears. PayGo bills monthly for actual use, with no upfront commitment. Enterprise Edition customers and above also get Agentforce builder tools and 100,000 Flex Credits per org per year at no cost through Salesforce Foundations.
Trust and Governance: How Does Agentforce Keep AI Safe?
Agentforce runs on the Einstein Trust Layer, respects existing Salesforce permissions, and logs every interaction for review. Salesforce names trust as its number one value and builds security and governance into the agent architecture rather than adding them later.
“AI safety is no longer mainly a model issue, but rather a system and deployment issue. AI systems aren’t just generating text now. They are influencing decisions, triggering processes, accessing data, and interacting with other systems. That means safety must draw from disciplines like cybersecurity, risk management, and safety engineering, not just model evaluation.”
– Francesca Rossi, IBM Global Leader for Responsible AI and AI Governance
According to Salesforce’s Trust and Agentforce guidance, the main safeguards are:
- Zero data retention with third-party LLM providers: external model providers don’t store prompts and responses that pass through the Trust Layer or use them for model training.
- Dynamic grounding in trusted company data: agents retrieve information from Data 360 and other authorized sources, which reduces hallucinations.
- Prompt injection and toxicity detection: the system screens for prompt injection attempts and flags potentially harmful responses.
- Complete audit trail: prompts, responses, and trust signals are logged in Data 360, giving full traceability for compliance reviews.
- Inherited Salesforce permissions: agents respect user licenses, field-level security and sharing settings, so they only access data they’re authorized to use.
- Trusted URLs: agents only share links on an approved allowlist and redact the rest.
- Protection for sensitive data: Trust Layer data masking is turned off for agents to keep answers accurate, but data agents access, including personal information, stays protected in transit and is not stored or used for training by external LLM providers.
Agentforce is covered by SOC 2 and SOC 3 reports, holds ISO 27001, 27017, and 27018 certifications, and is HIPAA-eligible. Good governance also depends on how each agent is set up: define an approved task scope for every agent, set escalation rules that enforce business policies, and require human review before high-risk actions. Our view on what governed Salesforce AI consulting should deliver in 2026 covers these controls in more depth.
What Does the Future Hold for Salesforce Agentforce?
Salesforce has made it clear that automation and AI-driven support are central to its future vision. The announcements at Dreamforce 2026 show where Agentforce is heading next.
- Purpose-built agents now arrive ready for specific jobs in sales, service, and supply chain, with subagents, actions, and Salesforce context already configured. Teams no longer start from a blank canvas for common use cases, which speeds up deployment.
- Long-horizon runtime lets agents pursue goals across days and weeks rather than single conversations. Agents keep context, adapt plans mid-workflow, and coordinate multi-touch follow-ups, with built-in checkpoints that keep humans in the loop. General availability is planned for November 2026.
- Agent Optimizer watches performance in production, surfaces friction in instructions and workflows, and recommends specific fixes. Teams move from spotting problems to fixing them faster, and quality improves as real usage grows.
- Multi-agent orchestration coordinates specialized agents around shared outcomes. Work routes to the right agent, context carries between steps, and handoffs span systems and departments without breaking the customer experience.
The move toward AI doesn’t mean replacing your people; it means augmenting them. Agentforce takes on the grunt work, so your employees are able to focus on delivering a better, more personalized experience.
The Final Word
Agentforce is more than just a tool; it’s a comprehensive support system for your entire business. From assisting customers to helping employees, Agentforce works around the clock to provide quick, reliable service.
By reducing manual tasks, it frees up valuable time for your team to focus on what matters most. If you’re looking for Agentforce services that help both inside and outside your business, a Salesforce Einstein Services Provider might help you get started.
References:
- [1]: Salesforce
- [2]: Salesforce Ben
- [3]: Salesforce Ben
