Agentforce use cases, trust, and permissions

An Agentforce agent understands a request in plain language, picks a topic, and runs actions such as flows and prompt templates to complete it. Use an agent where requests vary and need language skills; use Flow where every step must run the same way every time. Agents only work as well as the data and permissions behind them.

How an agent works

• Agent: An AI assistant that can reason over a request and take actions in Salesforce. Agents can serve customers (for example a service agent on a website) or employees (inside Lightning Experience).
• Topics: The jobs an agent can do, such as order status or appointment booking. Each has a classification description, a scope, and instructions.
• Actions: What the agent can run within a topic: flows, Apex, prompt templates, and standard actions such as querying records.
• Reasoning engine: Reads the request, chooses the best topic, plans which actions to run, and builds the response.
• Channels: Customer-facing agents are deployed to channels such as Messaging for web or an Experience Cloud site.

When AI is the right tool

• Good fits: Answering questions in natural language, summarizing records or conversations, drafting emails, and handling requests that arrive in many different wordings.
• Better with Flow: Rules that must run the same way every time, such as updating a field when a status changes or a calculation with an exact answer.
• Human in the loop: Keep a person reviewing high-risk outputs, and give agents a clear way to hand off to a human, such as escalating to a service rep.
• Agents work from your data, so incomplete or messy records produce weaker answers. Data quality is part of an AI rollout.

The Einstein Trust Layer

• Purpose: A set of protections between your data and the large language model.
• Grounding with permissions: Prompts pull in Salesforce data only as the running user (or agent user) is allowed to see it.
• Data masking: Sensitive data can be masked before a prompt is sent to the model.
• Zero data retention: Model providers used through the Trust Layer don't keep your prompts or responses or use them for training.
• Toxicity detection and audit: Responses are scored for harmful content, and prompts and responses can be logged for review.

Access and troubleshooting

• Prerequisites: Einstein generative AI must be turned on in Setup, and many Agentforce features need Data 360 (Data Cloud) set up.
• Service agent user: A customer-facing agent runs as a dedicated agent user. Give that user only the object and field access its actions need.
• Employee agents: Run with the permissions of the employee using them, so they can't see or change what that person can't.
• Building agents: Admins and builders need the permissions that allow managing agents and prompt templates.
• If an action fails or returns nothing, check the running user's permission sets, object and field access, and sharing before changing the agent's instructions.