How Agentforce Works > How Agentforce Works video

In this demo, I'm going to show you how Agentforce works by building an AI service agent to deflect incoming support cases by answering customer questions. To see how this works, let's expand this agent's capabilities by also enabling it to help with the installation scheduling of these products. I'll do this by adding a new topic. Topics are how you define the jobs your AI agents will and won't do, and they help establish guardrails for your AI agents. I want to expand my agent's guardrails by creating a new topic.

If you've ever built an old-style chatbot, this process will look very different. There are no dialog trees, just natural language descriptions that help the agent understand the job to be done. In this case, the job is appointment management, followed by a description that helps the AI agent understand when it should use this topic. The description includes the scope of the job and the instructions the agent should use when doing this work, such as getting the customer's email address before scheduling an appointment, getting the customer's installation date preference, converting all dates into a specific format, and ensuring the installation date isn't before the delivery date. Finally, the agent will confirm the customer's appointment once it's been booked. These instructions would involve lots of complicated if-then-else logic with a traditional chatbot, but with Agentforce, it's just natural language descriptions that help the AI agent understand the job to be done.

The next step is to add actions, which enable the agent to actually do these jobs. Actions can be based on flows, Apex classes, or prompts. For now, I'm going to keep it simple. This is a flow I created for retrieving available installation times, which will equip the AI agent with the knowledge it needs to do this work. This flow actually books the installation appointment and has a bit more logic to establish a few extra guardrails and be ultra-prescriptive about how this work gets done. With these two flows created, I'm now going to add them as actions to my AI agent. Believe it or not, this AI agent can now schedule installation appointments.

Let's try it. I'll start by asking it to get the status of an order again. Once again, I get an answer, and I can see the topic was classified as order management. Now, I'm going to ask this agent to schedule the installation of this order. As you can see, the agent pivots and reclassifies the topic. Because I gave it an instruction, it's asking me what my installation date preference is. Now, I'm going to throw it a curveball and reply with "tomorrow," which is actually before the delivery date. But this was another instruction I gave the agent, and it prevents this from happening, politely asking me to pick a date after the delivery date.

Now, I'm going to be really difficult and say, "What about that Friday, but not morning?" Let's just pause to consider the context required to understand a statement like this. Not only was I understood, but I got an appropriate response. Let's double-click into what just happened here. Beyond topic classification, Agentforce used the context of this entire conversation to make a dynamic plan, a plan that involved executing the appointment retrieval flow I added as an action. Look closer: that Friday was converted to the right date using the format required by the flow. Look at the output of that flow, which included an AM appointment time. But because I said "not morning," it was automatically filtered from the generated response. This is happening because Agentforce is giving AI agency to reason within the guardrails, which makes these AI agents much more flexible and capable of doing their assigned jobs.

So now you've seen what Agentforce can do within the guardrails. But what happens when I go outside the guardrails by asking it to do something it wasn't designed to do? As you can see, Agentforce classified this as off-topic and redirects me back to the approved topics. Agentforce also has built-in harm and toxicity detection, which means if I ask the AI agent something inappropriate, it will be detected, classified as inappropriate, and the agent will refuse to engage. If I go even further and attempt something malicious, like a prompt injection attack, you'll see Agentforce automatically detects and rejects this reverse engineering attempt.

That is how Agentforce works: by helping you create AI agents and deploy those agents across multiple channels to answer questions, take action, and drive customer success by reliably doing the jobs they're assigned. Thanks for watching.