For a decade, "AI in the CRM" mostly meant chatbots that followed scripts and predictive scores that sat quietly on a record page. Agentforce is a different idea. Instead of scripting every turn of a conversation, you give an AI agent a job, the information it needs and a set of actions it is allowed to take, and it works out how to get the job done.

This guide explains what Agentforce is, how the pieces fit together, where it delivers value today, what it costs and how to launch a first agent without putting customers or data at risk. It is written for business and technology leaders who already use Salesforce and want a clear picture before they commit budget.

What is Agentforce?

Agentforce is Salesforce’s platform for building, testing and running autonomous AI agents that operate inside the Salesforce ecosystem. Salesforce introduced it at Dreamforce in September 2024, and it has since become the centre of the company’s AI strategy.

An Agentforce agent can understand a request written in everyday language, look up the relevant records, decide which steps are needed and carry them out, for example checking an order, rebooking an appointment, updating a case or drafting a follow-up email. Crucially, it does this using the same data, permissions and automation your team already relies on in Salesforce.

Agents can face customers (on your website, in a mobile app or on messaging channels) or support employees (inside Salesforce, Slack or other tools). Salesforce also provides prebuilt agent templates, such as service and sales development agents, that you can adapt instead of starting from a blank page.

How does Agentforce work?

Every agent is built from a small number of components. Understanding them is the fastest way to understand what an agent can and cannot do.

Topics and instructions

A topic is a category of work the agent handles, such as "Order status", "Returns" or "Appointment changes". Each topic has a description that tells the agent when it applies, plus instructions written in plain language: what to ask for, what policies to follow and when to stop. Good topics are narrow and unambiguous. Vague topics are the most common cause of agents picking the wrong path.

Actions

Actions are the things an agent is allowed to do. They are built from tools Salesforce teams already know:

• Flows for record lookups, updates and multi-step business processes
• Apex for complex logic or callouts that Flow cannot handle cleanly
• Prompt templates for generating summaries, replies or recommendations
• APIs and MuleSoft for reaching systems outside Salesforce, such as an ERP or order platform

An agent cannot do anything that has not been exposed to it as an action. This is the single most important safety property of the platform, and it is why action design deserves most of your build effort.

The Atlas Reasoning Engine

The Atlas Reasoning Engine is the part that plans. It reads the user’s message and the conversation so far, chooses the matching topic, decides which actions to call and in what order, checks the results and either responds or asks a clarifying question. This loop is what separates an agent from a scripted bot.

Grounding in your data

Agents are only as useful as the information they can reach. Grounding can come from standard and custom Salesforce objects, your Salesforce data model, Knowledge articles and, for broader or unstructured data, Salesforce Data Cloud. Grounding lets the agent answer with your policies and your customer’s actual records instead of generic model knowledge.

The Einstein Trust Layer

Requests to the underlying large language models pass through the Einstein Trust Layer. It supports secure data retrieval that respects user permissions, masking of sensitive data before prompts leave Salesforce, toxicity detection, audit trails and zero data retention agreements with model providers. It reduces risk, but it does not replace good design. An agent with overly broad permissions is still a risk.

Agentforce vs Einstein Bots vs a generic AI chatbot

Capability	Einstein Bots	Generic AI chatbot	Agentforce agent
How conversations are handled	Scripted dialogue trees	Free-form answers from a model	Goal-driven reasoning within defined topics
Access to CRM records	Yes, through designed steps	Usually none, or a custom integration	Yes, through permitted actions
Can take actions (update, book, refund)	Only scripted ones	Rarely, and risky without controls	Yes, limited to the actions you expose
Handles unexpected phrasing	Weak	Strong	Strong
Security and permissions	Salesforce permissions	Depends on the vendor	Salesforce permissions plus the Einstein Trust Layer
Effort to maintain	High as flows grow	Low to start, hard to govern	Moderate: topics, actions and tests

What can Agentforce agents do? Practical use cases

The best early use cases share three traits: the request is frequent, the answer depends on data already in Salesforce, and a mistake is recoverable. Here are examples we see working well by team.

Customer service

• Order, delivery and case status questions answered from live records
• Appointment booking, rescheduling and cancellation
• Returns and warranty eligibility checks against written policy
• Summarising a long case history before it reaches a human agent

Sales

• Answering inbound product questions and qualifying leads around the clock
• Researching an account and drafting personalised outreach for a rep to approve
• Updating opportunities and logging next steps after a call

Marketing and commerce

• Guided product discovery on a storefront
• Drafting campaign briefs and audience segments for human review

Internal operations

• Employee questions about HR policies, IT access requests and onboarding tasks
• Field service technicians asking for asset history or parts availability

What you need in place before you start

Most failed AI pilots fail on foundations, not on the AI. Before building an agent, check the following:

1. Licensing and edition. Confirm with your Salesforce account executive which editions and Agentforce packages you need for your channel and use case.
2. Clean, reachable data. Duplicate contacts, inconsistent order statuses or outdated Knowledge articles lead directly to wrong answers. A Salesforce org health check is a sensible first step.
3. A documented process. If your team cannot write down how a return is approved, an agent cannot follow it either.
4. Reusable automation. Existing Flows and Apex that are bulk-safe and well tested can become actions quickly. Legacy Workflow Rules and Process Builders should be migrated to Flow first.
5. A dedicated agent user with least privilege. The agent should see only what each topic needs.
6. A human escalation path through Omni-Channel, with clear rules for when to hand over.
7. A test plan that includes awkward, ambiguous and adversarial requests, not just the happy path.

How much does Agentforce cost?

Pricing has changed several times, so treat any figure you read online as a starting point for a conversation rather than a quote. At launch in late 2024, Salesforce priced customer-facing agents per conversation. During 2025 it added a consumption model based on credits spent per action ("Flex Credits"), along with per-user options for employee-facing agents. Your total cost has three parts:

Cost element	What drives it
Agentforce licences or consumption	Channel, number of conversations or actions, and whether agents serve customers or employees
Supporting Salesforce products	Service Cloud, Messaging, Data Cloud or Knowledge if not already licensed
Implementation	Number of topics, complexity of actions, integrations, data clean-up, testing and change management

A focused pilot with a handful of topics and actions built on existing automation is a modest project. Costs rise with the number of external systems the agent must reach and the rigour of testing your industry requires. Our Salesforce implementation cost guide breaks down the implementation side in more detail.

A six-step plan to launch your first agent

1. Pick one high-volume, low-risk use case. Look at case reasons or chat transcripts and find the top three repetitive request types.
2. Map the process and the data. For each request, list the questions to ask, the records to read, the decisions to make and the updates to write.
3. Build and test the actions first. Make each Flow or Apex action reliable on its own, with clear inputs and outputs, before connecting it to the agent.
4. Write narrow topics and precise instructions. Include what the agent must never do, such as promising refunds outside policy.
5. Test at scale. Use Salesforce’s agent testing tools to run large batches of realistic and deliberately tricky utterances, then fix topics, instructions and actions until results are consistent.
6. Launch with a human in the loop. Start on one channel, monitor transcripts daily, measure containment, escalation reasons and customer satisfaction, then expand.

Common mistakes to avoid

• Starting with the hardest problem. Complex, high-stakes processes make poor pilots.
• Giving the agent broad permissions "to be safe". It is the opposite of safe.
• Overlapping topics that make the agent guess which one applies.
• Ignoring Knowledge quality. An agent will confidently repeat an outdated article.
• Measuring only deflection. Track resolution quality and escalations too, or you will optimise for frustrated customers.
• No owner after go-live. Agents need someone reviewing transcripts and improving instructions every week.

Is Agentforce right for your business?

Agentforce is a strong fit if you already run core customer processes in Salesforce, handle a meaningful volume of repetitive requests, and have reasonably clean data. It is a weaker fit if your key data lives outside Salesforce with no integration plan, or if your processes are still changing week to week. In that case, fix the foundations first; the agent will be better and cheaper to build afterwards.

If you are weighing Agentforce against a custom large language model application built outside Salesforce, our guide to RAG versus fine-tuning explains the alternative architecture, and our overview of generative AI use cases with real ROI will help you prioritise.

How Groviya helps

Groviya’s Salesforce team designs and delivers Agentforce pilots end to end: use-case selection, data and process readiness, action development in Flow and Apex, integration through MuleSoft or APIs, testing and post-launch optimisation. Explore our Salesforce implementation services or talk to an expert about a scoped pilot for your team.