Building trust in your AI Agent with Testing Center for Agentforce

Imagine having a team of digital employees working for your business, handling customer queries, processing orders, and more. That’s the promise of Agentforce, but like any new employee, these AI Agents need training and testing to ensure they perform their roles accurately and reliably. This is where the Agentforce Testing Center comes in.

What is Agentforce Testing Center?

Think of the Agentforce Testing Center as a quality control lab for your AI Agents. It’s a tool that allows you to rigorously test your Agents before they start interacting with real customers or employees. The testing center ensures your AI Agents are responding to the correct topics and providing the right answers to the questions they have been asked.

Building trust with testing

Like Salesforce, Trust is one of our core values at Nebula. Although AI Agents are powerful, they are not fully trusted by everybody yet because people do not understand how they come up with their answers. While it is great to embed digital employees into your business workflow, they will never get past the starting line without trust. Building trust through rigorous testing can help give your business confidence that your AI Agents are working as they should.

The testing center is an essential part of build on Agentforce as it allows you to:

• Ensure Accuracy: AI Agents are complex and need to understand the nuances of human language. Testing lets you validate that they interpret requests correctly and don’t make mistakes.
• Prevent “Hallucinations”: Generative AI can sometimes fabricate information, which is known as “hallucinating.” The Testing Center helps identify and correct these issues before they impact your business.
• Save Time and Resources: Providing the ability to perform multiple tests and allows you to find potential issues and resolve them quickly. Keeping testing consistent and recorded allows you to see your progress.
• Build Trust: Being able to show consistent outcomes and reflect these over time as you build upon your Agents provides you with metrics to show the business and help build trust in Agentforce.

How does the Agentforce Testing Center work?

You begin by loading the tests you wish to run and the Agent you plan to use to execute them. At present, these are loaded through a spreadsheet. However, in future releases, you will have the option of AI-generated test scenarios.

Once you run your tests, these are then scored against three key metrics:

1) Topic

Topics in Agentforce represent the jobs to be done. Your Agent will determine which topic it deems appropriate for the task it has been assigned. If it selects the wrong topic, this can lead to numerous issues, ultimately failing to deliver the answer the user anticipated. Should your Agent not succeed in this test, there are several factors to consider:

• Do you have too many topics? As with any employee, providing too many options can lead to confusion and incorrect decisions. It is important to ensure that Agents have only the relevant topics available to them.
• Are your topics clear and concise? As you build over time, you may find that your topics grow as well. Writing your topics as if you were writing them for an intern with limited knowledge will help ensure that the Agent understands them and is more likely to pick the correct one.
• Do you have overlapping topics? If you write multiple topics with similar instructions, it is easier for the Agent to pick a topic you were not expecting. After writing your topics, read them and condense them if needed.

2) Action

In Agentforce, the actions are things an Agent can do. These could be calling tools such as Flow or Apex or calling another prompt built in Prompt Builder. If the Agent performs the wrong action, it could be accessing or updating data that it should not be or consuming Einstein Requests unnecessarily. If your Agent is performing the wrong actions, you should consider: 

• Do you have the correct actions assigned to your topics? If your Agent selects a topic and the action you expected is not assigned to it, the Agent will not be able to perform that action. So, it is important to check which actions you have assigned to a topic, including removing ones that should not be there.
• Are the actions set up to do what they are supposed to? Your Agent may be calling an action, but if it has not been set up correctly or changes are required, the action’s output will not be as you expect.
• Are your inputs and outputs set up correctly? Each action requires inputs and outputs, which control the information the Agent can pass to it and the information it will receive back. If you do not have detailed descriptions or have configured something incorrectly, the action will not work as expected.

3) Outcome

Ensuring the Agent does not break any guardrails in place and delivers a satisfactory outcome is the most important test you need to run. This final test is the most important as an Agent could select the wrong topic or action, but it is how it behaves when this happens which is key. If your Agent is not responding, questions you should be asking yourself should be:

• Do you have sufficient guardrails in place? Guardrails are a critical step in the Agent’s process, ensuring that the Agent does not engage in actions it should avoid. Your tests will reveal the guardrails that should be considered to guarantee the Agent remains on the correct path and delivers the expected outcome.
• Have you documented sufficiently at each stage of the process? There are numerous references from the instructions in the topics to the descriptions of the actions, right through to the rules in your prompts. If you are not clear at every stage, you leave yourselves open to the Agent assuming the appropriate response. Each level must possess the correct extent of documentation. For instance, if you require actions to be executed in a specific order, you must explicitly mention that.
• Do you have the correct data available? If your Agent lacks access to the necessary data, it may be necessary to update its permissions. Additional data may also be stored in an unstructured format, which might necessitate an Agentforce Data Library. If your employees possess information that your Agent lacks, they cannot complete the tasks to the same standard. 

Using Testing Center

The Agentforce Testing Center is an essential tool for businesses looking to utilise the Agentforce platform. It enables teams to confidently deploy accurate and reliable Agents that deliver the expected results, fostering confidence and trust within your business.

Interested in discovering more about how Agentforce and the Testing Center can assist your business? Visit our Agentforce Hub to explore additional content and schedule a call with one of our Agentforce experts.