Agentforce is Salesforce's platform for building AI agents that work inside your Salesforce org and on channels such as web, phone and apps. Unlike a Flow, which runs steps someone designed in advance, an agent interprets a request, decides which of its permitted actions to use, and carries them out. It draws on your CRM records, knowledge and, where connected, data from Data 360. What it can do well depends almost entirely on the data, permissions and actions you give it.
What Agentforce is made of
Salesforce describes an AI agent as software that uses generative AI to decide what to do and which tools to use, rather than following a script written beforehand. In Agentforce, that decision-making comes from the Atlas Reasoning Engine, which Salesforce says breaks a request into smaller tasks, evaluates each step and proposes a plan until the answer or action is complete.
Agents are configured in Agentforce Builder, a low-code tool. The work an agent can do comes from actions you connect: Flows, Apex, prompt templates, MuleSoft APIs and knowledge articles, plus prebuilt actions and templates for common jobs. Salesforce has also introduced Agent Script, a language that mixes natural-language instructions with programmatic logic for the parts of a conversation that must behave the same way every time. Underneath sits the Einstein Trust Layer, which Salesforce says grounds prompts in your data, masks sensitive information before it reaches the language model, and applies zero data retention with third-party models.
How agents differ from Flow and chatbots
| Flow | Traditional chatbot | Agentforce agent | |
|---|---|---|---|
| How it decides | Follows the path a builder drew, step by step | Matches input to scripted intents and dialogs | Reasons over the request and chooses among permitted actions |
| Handles unexpected phrasing | Not applicable; it is triggered by events or buttons | Poorly, unless the variation was anticipated | Well, within the scope and instructions it is given |
| Takes action in Salesforce | Yes, exactly as designed | Usually only through handoffs or simple lookups | Yes, through Flows, Apex and other actions you connect |
| Predictability | Fully predictable | Predictable but brittle | Varies; needs testing, guardrails and review |
| Best for | Rules that never change: approvals, updates, routing | Narrow, high-volume menus | Requests that vary in wording but have clear, allowed outcomes |
These are not competing choices. A well-built agent often calls existing Flows as its actions, so the reasoning happens in the agent and the record changes happen in automation you already trust. If a process has one correct path every time, a Flow is still the cheaper, more reliable tool.
Where agents fit: service, sales and employee support
Salesforce positions Agentforce across three broad areas. In service, agents answer questions, resolve cases, manage orders and troubleshoot, with a person taking over when the request falls outside what the agent should handle. In sales, sales development agents engage prospects, answer questions, handle objections and book meetings. For employees and IT, agents resolve help desk tickets and handle routine requests such as access resets.
Our own projects show how differently these look in practice. For a biotech company, an agent grounded in 19 knowledge articles drafted email replies and summarized cases, producing send-ready drafts for 30% of question-type cases while staff kept the final say. A telecom-infrastructure contractor used Document AI to read checkbox-heavy inspection forms and an agent to answer natural-language questions about project cycle times and delays. A business lender triaging 24,000 leads a month started with lead scoring and a Data Cloud roadmap, with sales development agents planned on top of that foundation.
What an agent needs to work
- A defined job: the requests it handles, the ones it refuses, and when it hands off to a person.
- Accurate records in the objects it reads, because the agent will repeat whatever the data says.
- Current, owned knowledge articles for any question-answering use case.
- Permissions scoped to the objects, fields and actions its job requires, and nothing more.
- Actions that are already reliable, since an agent calling a broken Flow simply breaks faster.
- Data from outside Salesforce, through Data 360 or integration, if the answers depend on it.
- Test conversations drawn from real history, and someone reviewing live output after launch.
Data 360 (formerly Data Cloud) matters when the agent needs context the CRM does not hold, such as orders, usage or documents. If the use case runs on records already in Salesforce, it may not be needed at first. Our guide to when you need Data Cloud covers that decision, and the Agentforce readiness checklist goes through data, security and governance in detail.
How pricing is structured
Salesforce publishes several ways to pay for Agentforce, and they suit different patterns of use. Consumption-based options include Flex Credits, charged per action an agent takes, and a per-conversation model. Per-user add-ons give employees unmetered use of Agentforce within Sales, Service, Field Service and industry products. Agentforce 1 Editions bundle the add-on with an annual allocation of Flex Credits. Salesforce also lists a way to start at no cost through Salesforce Foundations, and offers pre-purchase, pre-commit and pay-as-you-go purchasing.
Which model fits depends on who the agent serves. Customer-facing agents with unpredictable volume point toward consumption pricing; agents that help your own staff all day may suit a per-user add-on. Confirm current terms with Salesforce before you budget, and estimate how many actions or conversations your first use case will generate.
How to start
Pick one use case with steady volume, clear right answers and low risk if the agent gets one wrong, such as drafting replies to common service questions or summarizing records for staff. Start internal-facing where you can, so employees review output before customers see it. Build the test set before the agent, widen scope only when the results justify it, and name a business owner who decides what the agent may do. Our Agentforce implementation playbook lays out that sequence step by step.
