Setting Up Your First AI-Powered Service Agent | Automate with Agentforce

The challenge

As a Salesforce Admin, you're at the center of your business, constantly asked to increase efficiency, improve user experience, and drive innovation, often all at once. When ‘learning AI’ gets added to your to-do list, it can feel like another mountain to climb. But these new capabilities are essential, and they build on your existing expertise. My specific challenge was to demonstrate how to build a customer support agent in a way that doesn't require starting from scratch, but instead builds directly upon the core Salesforce knowledge we use every day.

The solution

My goal in building a customer support agent for case management was to uncover the thought process needed to empower admins to build their own tailored agents that understand unique processes, adopt company tone, and integrate seamlessly with existing automation—guiding you from initial setup to activation.

A key foundational step in my process involved creating an autolaunched flow to Find Contact by Email. This autolaunched flow takes an email address as input and queries contact records to retrieve the corresponding contact record. If a contact is found, the flow assigns the relevant contact details. If no contact is found, the flow ends. I'll show you how this autolaunched flow comes into play later on.

Another critical, and sometimes overlooked, foundational step is ensuring your agent is properly enabled and configured with the correct permissions. It’s important to understand that both the Einstein Generative and Agentforce capabilities must be on so that the agent can function correctly.

Crucially linked to activation is the absolute necessity of correctly configuring permission sets for your agent. The agent’s ability to function is entirely dependent on having the correct access to objects and fields. As an admin, you know the importance of access, so ensure you don’t rush this step. Applying the principle of least privilege (only giving the agent the minimal access needed) is key. You’ll notice that, for both permissions, the Edit All Fields permission isn't even displayed. This safeguard is designed to prevent you from accidentally over-provisioning an agent with permissions it doesn't need (that is, least privilege).

After enabling our foundational items, now we need to define the agent’s specific role through topics (in other words, this is the agent’s job to be done). I created a custom topic and gave the agent a clear objective: “I would like to do case management. Find the contact, update the case comments, and look up cases.” Why is this clear definition of the agent’s ‘job’ or ‘topic’ so crucial? Because it dictates the agent’s fundamental logic and ensures it operates exactly as intended, right from the start.

Once a clear topic is defined, the next crucial step is equipping your agent with the necessary actions to fulfill that purpose. My custom action allowed me to extend the agent’s functionality precisely to fit the use case. Likewise, it’s important to build with your unique business processes in mind, often by leveraging existing automations you’ve already built.

Beyond actions, effectively defining the scope (think of these as the agent’s guardrails) is important. Ensuring the instructions are clear and concise is key. This guides the agent on what it can and can’t do, preventing unintended actions, and keeps the agent focused on its specific purpose.

Last but not least, pay attention to the metadata in your action’s inputs and outputs. While this might seem like a technical detail, it’s how the agent truly understands the data it’s working with. Well-defined metadata ensures the agent accurately interprets user requests, processes information, and makes your actions much more reliable and efficient.

Remember, not everything has to be custom. The Asset Library offers several standard and prebuilt actions for expanding your agent’s capabilities. In my case, I leveraged ‘Add Case Comment’ and ‘Get Case by Case Number’ to complete my agent’s capabilities. Custom actions should be used when your business processes are truly unique or require integration with external systems not covered by existing options. Extended standard actions offer the best of both worlds—you can use a prebuilt action as a baseline and then modify its underlying flow or prompt template to add your unique customizations. Combining custom, standard actions, extended standard actions will help to create comprehensive, efficient skills for your agent—which will get you value faster.

No agent is truly production-ready without a strong testing strategy. For me, validating the agent’s logic and responses wasn’t just a final hurdle—it was an iterative process to ensure it behaved as expected and addressed all edge cases. Your testing strategy should focus on thoroughly exercising the conversation flow and anticipating diverse user inputs. Leverage tools like Agentforce Testing Center that test large numbers of scenarios using batch testing so you can quickly assess multiple requests in a single test.

Key benefits and lessons learned

• With Agentforce, your team can stop searching and start solving. It connects directly to your customer information to automatically handle the groundwork for any case, meaning your agents can spend their time delivering excellent service instead of doing data entry.
• The agent’s ability to function is entirely dependent on having the correct access to objects and fields. This isn’t a step to be rushed. Practice the principle of least privilege by only giving the agent minimal access needed to get the job done.
• Every business is unique, and your support agent should be too. With Agentforce, you can build a highly tailored assistant that understands your specific processes, uses your company’s tone, and connects directly to the custom automations you’ve already built.

Takeaways for admins

• Start with a clear definition of your agent’s topic or job (for example, “Case Management”).
• Always configure permissions first—without the right access, no action will work.
• Connect agent actions to existing flows to accelerate development.
• Combine standard actions from the Asset Library with your custom actions for a comprehensive skill set.
• The final step before going live should always be to test the conversation flow thoroughly.

Resources

• Trailhead: Agentforce Builder Basics
• Trailhead: Quick Start: Agent Actions
• Trailhead: Become an Agentblazer Champion
• External site: YouTube: Agentforce Decoded Series
• Salesforce Help: Agentforce Testing Center

Want to see more good stuff? Subscribe to our channel!