Agentforce is Salesforce's platform for building and running AI agents that work on your Salesforce data and take actions, not just answer questions. An agent can resolve a case, update a record or qualify a lead, following instructions and permissions your team sets, and hand over to a person when it should. It is part of Salesforce's Agentforce 360 offering.
This guide is written for the people who will run it: Salesforce admins, architects and platform owners. It covers what an agent is made of, how it works step by step, how it differs from Einstein Bots, chatbots and copilots, which agents Salesforce sells today, how the Einstein Trust Layer protects your data, what Agentforce costs and what to get ready first. Every Salesforce fact links to an official Salesforce source, with the date we checked it.
How does Salesforce describe Agentforce?
On its Agentforce page (checked 10 November 2026), Salesforce calls Agentforce "the AI agent platform that delivers 24/7 autonomous support at enterprise scale" and says over 18,000 companies already run on it. The same page lists Agentforce as formerly Einstein Copilot, so if your org used Copilot, this is its successor. When Salesforce announced general availability on 29 October 2024, it described Agentforce as a new layer on the Salesforce Platform.
In practice, Agentforce gives you three things in one place:
- A reasoning engine, the Atlas Reasoning Engine, which Salesforce calls the brain inside Agentforce. It interprets the request and plans the steps.
- Access to your business data, from CRM records, knowledge articles and Data 360, so answers are grounded in your facts rather than general internet knowledge.
- Actions built from things you already have in Salesforce, such as flows, Apex, prompt templates and APIs, so the agent can do the work, not only describe it.
Agentforce sits inside a wider package. In its 13 October 2025 announcement, Salesforce described Agentforce 360 as four parts: the Agentforce 360 Platform for building agents, Data 360 as the unified data layer, Slack as the place people and agents work together, and the Customer 360 apps such as Sales and Service. Data 360 is the new name for Data Cloud; the product is the same. Our Salesforce Data Cloud (Data 360) guide explains how it works.
What is an Agentforce agent made of?
Every agent is built from the same parts. If one is weak, the agent is weak, whatever model sits underneath. Trailhead's introduction to Agentforce agents names five key components (role, knowledge, actions, guardrails and channels) and describes topics, each with its own instructions and actions, in its section on the reasoning engine. Newer Trailhead content for Agentforce Builder calls topics "subagents", so you will see both terms.
| Building block | What it is | Example for a service agent |
|---|---|---|
| Role | The agent's purpose and the job it should do | Help customers with orders and returns |
| Knowledge | The data the agent needs: company articles, CRM records and external data | Order records, return policy articles |
| Topics or subagents | Groups of related jobs; the reasoning engine sorts each request into one | "Order status", "Returns and refunds" |
| Instructions | Plain-language rules on how to decide and reply, and what not to do | "Offer a return label only for orders under 30 days old" |
| Actions | Tasks the agent may run, built from flows, prompt templates, Apex or APIs | Look up an order, create a return case |
| Guardrails | Limits on what the agent can do and when it hands off to a person, plus built-in Trust Layer protections | Hand over to a person if the customer mentions legal action |
| Channels | Where the agent works: website, CRM, Slack and more | Web chat and WhatsApp |
Salesforce's how it works page says admins and developers define "jobs to be done, instructions, actions, and guardrails" in the builder, and that you can turn existing flows, prompt templates, Apex and APIs into agent actions. Trailhead's Agentforce Builder quick start shows subagents, actions and instructions written in Agent Script, and a Preview panel where you can watch how the agent plans and runs actions.
For a Salesforce team, the practical point is that most of this is configuration you already understand: permission sets, sharing, flows, knowledge and data quality. The new skill is writing clear instructions and testing how the agent reasons with them.
What is an Agentforce agent made of?
- InstructionsJob, tone and limits
- ReasoningAtlas plans the steps
- DataRecords, Data 360
- ActionsFlows, Apex, APIs
- GuardrailsHandover to a person
How does an Agentforce agent work, step by step?
Every request follows the same broad path. Knowing it helps you see where design choices, and most problems, sit.
- A request arrives. A customer, lead or employee asks something in a channel such as web chat, Slack or a phone call, or a system event triggers the agent.
- The agent works out the job. The Atlas Reasoning Engine classifies the request into a topic (subagent) based on the descriptions you wrote.
- It plans. The engine splits the request into smaller tasks, reads that topic's instructions and chooses from the actions the agent is allowed to use.
- It grounds itself in your data. Actions and retrieval pull the relevant records and knowledge, so the answer reflects your policies and the customer's actual account. The agent sees only what its running user is allowed to see.
- It acts. The agent runs the action, for example updating a record, starting a flow or booking a meeting, with the access of the user it runs as. Salesforce Help explains that a service agent usually runs as a dedicated agent user, while an employee agent runs as the logged-in user.
- It checks and replies, or hands off. The response passes through the Einstein Trust Layer, and if a guardrail says a person must take over, the agent escalates with the conversation context.
- You monitor it. Salesforce's how it works page describes testing tools that generate test cases and monitoring that shows how agents perform.
Steps 2 and 3 are where most quality issues start. Vague topic descriptions send requests to the wrong place; loose instructions let the agent improvise. Our guide to Agentforce governance, security and testing covers how to test those steps before launch.
How is Agentforce different from Einstein Bots, a chatbot or a copilot?
The first table uses only Salesforce's own descriptions: the Trailhead comparison of Agentforce for Service and Einstein Bots and the how it works page for traditional chatbots, both checked on 10 November 2026. Salesforce describes traditional chatbots only by their conversation trees, so the other cells in that column say so.
| Question | Agentforce agent | Einstein Bots | Traditional chatbot |
|---|---|---|---|
| How does it decide what to say or do? | Reasons over natural language instructions and plans steps | Follows designed dialogs and trained intents | Needs pre-determined conversation trees |
| How is it set up? | Prebuilt subagents and actions | Manually training natural language models and designing dialogs | Not described separately |
| How is it kept up to date? | Adding and refining natural language instructions | High maintenance to keep dialogs and intents current | Not described separately |
| Follow-up questions and exceptions | Responds to follow-ups and unplanned exceptions | Responds only to the most recent question | Not described separately |
| Where it still fits | Varied requests that need actions on CRM data | Trailhead says it may suit regulated work that needs explainable, deterministic flows | Not described separately |
A copilot sits between the two. Salesforce's guide to AI agents describes chatbots as static and linear, and copilots as tools that assist a person rather than execute the work. The second table is our summary of that difference.
| Question | Rule-based chatbot | AI copilot | Agentforce agent |
|---|---|---|---|
| Who completes the task? | The customer or an employee, after the bot hands off | The person using it | The agent, using permitted actions; it hands off to a person when rules say so |
| Where does its data come from? | Fixed answers written in advance | The record or screen the user has open | CRM records, knowledge, Data 360 and connected systems |
| Typical Salesforce example | A scripted FAQ bot on a website | An assistant that drafts an email for a rep to send | A service agent that checks an order and processes a return |
Einstein Bots are not wrong for every case. If a process must follow the same path every time and be explained step by step, a deterministic flow can still be the safer choice. And because an agent acts, it needs tighter design, testing and permissions than a chatbot. For the service options side by side, see Salesforce chatbot options compared.
What can Agentforce do? Use cases by team
Each use case below matches a job Salesforce has built an agent for, and each links to the guide that covers it:
- Customer service: order status, returns, appointment changes and case triage over voice, SMS, WhatsApp and web chat, with handover to a person and the context. See Agentforce for customer service.
- Sales: answer and qualify inbound leads across websites and inboxes, book meetings, and research and reach out to new prospects. See Agentforce for sales.
- IT and HR: answer employee requests in Slack, portals and work tools. See employee agents for IT and HR.
- Commerce: help shoppers compare products and check out in chat.
- Supply chain and manufacturing: run back-office supply chain processes and order exceptions. See Agentforce for manufacturing.
- Healthcare: patient summaries, scheduling and member questions in Health Cloud. See Agentforce Health.
Types of Agentforce agents: which does Salesforce offer today?
Besides the agents you build yourself, Salesforce now sells named agents for specific jobs. Starting from one shortens design work, but you still configure its data, actions and guardrails for your org. In its 11 September 2026 announcement it listed these statuses:
| Agent | Job | Status stated by Salesforce |
|---|---|---|
| Casey | Help agent for customer service over voice, SMS, WhatsApp and web chat | Generally available |
| Paige | IT and HR service requests in Slack, portals and work tools | Generally available |
| Carter | Shopper agent that helps people compare products and check out in chat | Generally available |
| Marshall | Supply chain agent for back-office processes | Generally available |
| Piper | Inbound lead agent across websites and inboxes | Generally available |
| Fin | Customer agent for complex customer experience workflows | Generally available |
| Hunter | Outbound sales agent from research to outreach | Pilot; general availability planned for November 2026 |
The same announcement says customers deploy Piper in 45 days on average. As of 11 November 2026 we could not find a Salesforce page confirming that Hunter has reached general availability, so check before you plan around it. The Dreamforce 2026 recap tracks these statuses.
What changed in Agentforce 360 (Dreamforce 2026)
Agentforce 360 is the name Salesforce now uses for its whole platform. Its 13 October 2025 press release says Agentforce 360 was available globally that day, with further features to follow in pilot and beta. The same release describes Agent Script, which runs on the configurable Atlas Reasoning Engine and lets builders set rules, tool access and decision logic, and says customers can use models from Anthropic (via Amazon Bedrock), Google's Gemini and OpenAI.
| Date | What Salesforce announced |
|---|---|
| 29 Oct 2024 | Agentforce generally available as a new layer on the Salesforce Platform |
| 13 Oct 2025 | Agentforce 360 available globally; Agent Script and hybrid reasoning; wider model choice |
| 11 Sep 2026 | Seven named, job-ready agents; six generally available, one in pilot |
At Dreamforce 2026 (15 to 17 September 2026) the main change was packaging: Salesforce now sells the named, job-ready agents in the table above, so teams no longer assemble every agent from scratch. It also announced AIforce, a way for any AI interface to use Salesforce capabilities, and its own CRM reasoning model, Koa, both without availability dates as of 10 November 2026. Data Cloud now goes by Data 360. The Dreamforce 2026 recap lists what is live, in pilot or upcoming.
How does Agentforce keep data safe?
Two layers work together: controls Salesforce provides and controls you configure. Agentforce runs through the Einstein Trust Layer. Salesforce's how it works page lists dynamic grounding, zero data retention and toxicity detection, and Trailhead adds secure data retrieval and an audit trail that tracks each prompt. One detail matters for architects: Trailhead says data masking for LLMs is currently disabled for agents, while it remains available for embedded features such as Einstein Service Replies, so do not count on it to hide sensitive fields from an agent.
The Trust Layer does not decide what the agent may do in your org. That comes from you:
- The agent user's permissions. Give the agent the least access it needs, through its own permission sets.
- The actions you expose. Every flow or Apex action is something the agent can trigger. Review each one as you would a new integration.
- Escalation rules. Write down when the agent must hand over to a person.
- Testing and monitoring. Test in a sandbox before launch and review live conversations after it.
How much does Agentforce cost?
Salesforce sells Agentforce in several ways. The list prices below are from its pricing page on 11 November 2026, in US dollars. Salesforce notes the page is for information only and subject to change, and your contract price may differ.
| Model | List price | Fits best when |
|---|---|---|
| Flex Credits | $500 per 100,000 credits; a standard action uses 20 credits (about $0.10) and a voice action 30 | You want to pay per action across several agents |
| Conversations | $2 per conversation | You run a customer-facing service agent and count work in conversations |
| Agentforce add-ons (Sales, Service, Field Service and Operations) | $125 per user per month | You want per-user pricing for employees who work with agents |
| Agentforce Industries add-ons | $150 per user per month | You use an industry cloud |
| Max Edition | From $550 per user per month, including the Agentforce add-on and 2.75M Flex Credits per org per year | You are standardizing a large org on Agentforce |
| Salesforce Foundations | $0 add-on for Enterprise Edition and higher | You want to try Agentforce before a wider purchase |
Salesforce's FAQ adds that unused Flex Credits do not roll over to the next term and that Flex Credits and Conversations cannot be used in the same org. Its how it works page says Salesforce Foundations includes credits for the first 1,000 conversations with Agentforce for Service. For a fuller breakdown, see Agentforce pricing explained, or estimate effort with the Agentforce cost calculator.
Agentforce best practices: what should a Salesforce team prepare first?
Most of the work in a good Agentforce project happens before anyone opens the builder. Check these points first:
- Pick one narrow, measurable job. A high-volume, low-risk task with a baseline you can compare against, such as one contact reason or inbound lead response time.
- Check the data. Know which objects, fields and knowledge articles the agent will read, who owns them, and whether they are complete, current and free of duplicates.
- Reuse existing automation. Flows and Apex that already do the task safely become your first actions.
- Set access and limits. Decide which user the agent runs as, give it the smallest set of permissions that does the job, and agree what it must never do and when it must escalate.
- Write and test instructions. Write realistic conversations, edge cases and misuse attempts, with pass criteria, and test with them before any customer sees the agent.
- Name an owner. Someone reviews conversations, handovers and errors after launch, tunes instructions and approves changes.
Before you build, work through the Agentforce readiness checklist. Our guide to grounding Agentforce with Data 360 covers when you need Data 360, and the Agentforce implementation timeline covers timing.
Is Agentforce right for your business?
| Strong fit | Wait or fix first |
|---|---|
| Your customer, sales or employee data already lives in Salesforce | Core data is spread across systems with no plan to connect it |
| You have repeatable, high-volume requests with clear rules | Every request needs expert judgment or is rare |
| Knowledge articles and policies are current and owned | Policies live in people's heads or out-of-date documents |
| Flows or Apex already handle the task safely | The process itself is not agreed or documented |
| You can name an owner to monitor the agent | No one has time to review conversations after launch |
If most of your answers sit in the right-hand column, start with data and process work. An agent follows your data and rules exactly, gaps included.
How does Techila help with Agentforce?
Techila Global Services is a Salesforce Summit Partner headquartered in Frisco, Texas, founded in 2012 by Chitiz Agarwal. Our 350+ Salesforce professionals hold 500+ Salesforce certifications and have delivered 700+ projects, including 15+ Agentforce programs. Our AppExchange listing shows a 4.82 rating from 19 reviews.
Our Agentforce work is architect-led and starts with your data and a single measurable use case:
- Agentforce readiness assessment: a written review of the data, actions and permissions a first agent needs.
- Agentforce pilot: one agent, one job, tested in a sandbox and measured against a baseline.
- Agentforce implementation services: use case selection, design, build, testing, rollout and monitoring across Sales, Service and industry clouds.
If you are choosing between firms, our guide on how to choose an Agentforce implementation partner gives a scoring table you can use on any partner, including us.
Frequently asked questions
How does Salesforce Agentforce work?
An Agentforce agent reads a request, uses the Atlas Reasoning Engine to break it into smaller tasks and plan the steps, pulls the facts it needs from Salesforce records, knowledge and Data 360, then runs actions such as flows, Apex or APIs within its instructions and guardrails. It hands over to a person when its instructions say it should.
What is the Atlas Reasoning Engine?
The Atlas Reasoning Engine is the part of Agentforce that decides what an agent does. Salesforce calls it the brain inside Agentforce: it reads a request, sorts it into the right topic, plans steps using only the actions the agent is allowed to use, and pulls in the data needed to complete the task.
What is the difference between Agentforce and Einstein?
Einstein is the older name for Salesforce's AI features, and it still appears in names such as the Einstein Trust Layer and Einstein Bots. Agentforce is the agent platform: Salesforce lists it as formerly Einstein Copilot. Compared with Einstein Bots, which need trained language models and designed dialogs, Agentforce agents work from natural language instructions and can handle follow-up questions.
Can Agentforce agents take actions, or only answer questions?
They can take actions. In Agentforce, an action is a task the agent may run, built from flows, Apex, prompt templates or APIs. The reasoning engine plans which permitted actions to use for a request. Agents can only use the actions you give them, and they run with the access of the user they operate as.
Is ChatGPT an agentic AI?
ChatGPT is a general-purpose AI assistant from OpenAI. Whether any AI tool counts as agentic depends on what it is set up to do: plan steps and take actions in real systems, not only answer. Agentforce is built for that inside Salesforce, where agents act on your CRM data under your permissions, sharing rules and guardrails.
Do you need Data 360 to use Agentforce?
Not always. An agent can start with the CRM records, knowledge articles and actions already in your org. Salesforce lists Data 360 as a source of trusted business data for agents, and it becomes important when the agent needs data from outside Salesforce, unified customer profiles or unstructured content such as PDFs.
Which Agentforce agents are available now?
On 11 September 2026 Salesforce announced seven job-ready agents. It listed six as generally available: Casey, Paige, Carter, Marshall, Piper and Fin. It listed Hunter, the outbound sales agent, as in pilot with general availability planned for November 2026. As of 11 November 2026 we could not confirm on Salesforce's pages that Hunter has reached general availability.
How much does Agentforce cost?
Salesforce offers several models. On 11 November 2026 its pricing page listed Flex Credits at $500 per 100,000 credits, with a standard action using 20 credits, $2 per conversation, and Agentforce add-ons from $125 per user per month. Salesforce Foundations is a $0 add-on for Enterprise Edition and higher. The right model depends on volume, so model your use case before you buy.
How long does it take to launch a first Agentforce agent?
It depends on data readiness and scope more than on configuration. A narrow first use case with clean knowledge content and existing flows moves much faster than one that needs new integrations. Salesforce reports that customers deploy its Piper inbound lead agent in 45 days on average; treat that as Salesforce's figure, not a promise for your org.
Is Agentforce secure and compliant with data regulations?
Agentforce runs through Salesforce's Einstein Trust Layer, which Salesforce describes as including dynamic grounding, zero data retention with third-party model providers, toxicity detection and an audit trail. Trailhead notes that data masking for LLMs is currently disabled for agents. Compliance with laws such as GDPR, CCPA or HIPAA depends on how you configure and use Agentforce: least-privilege permissions for the agent user, reviewed actions, test evidence and monitoring. Review each use case with your compliance team.
Talk to a Salesforce architect
Share the systems and goals involved to get an assessment.
Talk to our Agentforce team. We reply within 1 business day.
