Agentforce in Production: A Practical Guide to Actions, Permissions, and Testing

Building a Salesforce AI agent is only the beginning.

A prototype may successfully answer questions or execute a simple action, but production systems need stronger controls around permissions, data access, actions, error handling, and testing.

For Salesforce developers and admins, the goal should not be to make an agent simply “work.” The goal is to make it predictable, secure, and useful in real business workflows.

1. Start With a Specific Use Case

Avoid building an agent with dozens of responsibilities on day one.

Start with one measurable workflow.

For example:

Customer asks about an order
↓
Agent identifies the customer
↓
Finds order information
↓
Checks order status
↓
Returns the result

This makes it easier to test whether the agent is actually solving a business problem.

Salesforce's current Agentforce documentation describes actions as the building blocks that allow agents to perform tasks and interact with data.

1. Design Actions Carefully

An agent's actions need clear instructions.

A vague description such as:

Get order information.

doesn't provide much context.

A better description explains what the action does, when it should be used, and what information it expects.

For example:

Retrieve the customer's order status using
the order ID provided in the conversation.
Return the current status and expected
delivery date when available.

Salesforce recommends concise but specific action instructions and testing them iteratively in Agentforce Builder.

1. Apply Least-Privilege Access

One of the biggest mistakes in AI implementation is giving an agent more access than it actually needs.

If an agent only needs to read Cases and update a specific field, it should not automatically receive broad access to unrelated objects and data.

Salesforce recommends starting agent users with minimal access and granting only the permissions required for their tasks. Object permissions, field-level security, Apex access, and Flow access can all affect what an agent is able to do.

Think about permissions like this:

Agent
↓
Required Action
↓
Required Object
↓
Required Fields
↓
Minimum Necessary Access

This makes the security model easier to understand and audit.

1. Ground the Agent With Reliable Data

An AI response is only as useful as the information available to the agent.

If customers are asking about company policies, product information, or support procedures, the agent needs access to reliable sources.

Depending on the use case, grounding can involve Salesforce records, Knowledge, or other supported data sources.

The objective is to give the agent enough relevant context to answer the question without unnecessarily exposing unrelated information. Salesforce documentation also provides specific guidance for grounding service agents in customer context and knowledge.

1. Test Failure Scenarios

Testing only successful conversations is not enough.

Try inputs such as:

Missing record IDs
Incorrect customer information
Unauthorized requests
Empty search results
Unexpected user questions
API failures
Incomplete information
Requests outside the agent's responsibility

For example:

User → "Cancel my order."

Agent → Does the user have permission?
↓
Is the order found?
↓
Is cancellation allowed?
↓
Execute action
↓
Confirm result

The agent should fail safely when any step cannot be completed.

1. Monitor Before Expanding

Once the initial workflow works, monitor how users actually interact with it.

Look for:

Failed actions
Incorrect responses
Unnecessary handoffs
Permission errors
Repeated questions
Unexpected agent behavior
Low-value conversations

Then improve the action instructions, data sources, permissions, and conversation design.

This iterative approach is important because AI behavior depends on the quality of the context and instructions provided to the system.

From Prototype to Production

Moving from an Agentforce prototype to a production implementation often requires more than configuring an agent.

The surrounding Salesforce org may also need data cleanup, automation changes, integration work, security reviews, testing, and ongoing support.

For organizations planning a larger Salesforce AI implementation, iBirds Salesforce AI and Agentforce solutions can be explored alongside its related Salesforce AI adoption guidance.

The broader lesson is simple: build the smallest useful agent, control its permissions, give it reliable context, test failure paths, and measure the results before expanding its responsibilities.

That approach creates a much stronger foundation for production AI inside Salesforce.