Test your AI agent before customers do

0:00Test before launch

Before your customers meet your agent, give it a few good tests. In this lesson, we'll use Tuolen's Sandbox to check real answers, spot missing knowledge, and build a simple testing habit you can repeat.

0:14Build a four-question test set

Start with four kinds of questions: something your knowledge covers, something it doesn't, an exception to a policy, and a request for a person. Write down what a good response should do before you test. Otherwise, a confident-sounding answer can be surprisingly easy to accept.

0:32Open the right Sandbox

Open your agent and choose Sandbox. Check that you're testing the right agent and that its knowledge is ready. The embedded preview is your place to try the conversation. Treat its temporary session link as private, and don't use real customer details in your test questions.

0:51Check a known answer

Here's an actual recorded test: what is your return policy for unused items? The agent answers with thirty days from delivery, original packaging, and a useful follow-up question. Compare every condition with the approved source. We're checking accuracy, not whether the wording matches a script.

1:12Check an unknown answer

Now ask something the source doesn't cover: do you offer a lifetime warranty? In this recorded response, the agent says it doesn't have that information and asks which product the visitor means. That's more helpful than guessing. Your own tests should check this boundary just as carefully as the easy answers.

1:32Test exceptions and human requests

Add two more test cases to your checklist. Ask for an exception, like approving a refund outside the policy, then ask to speak with a person. These are suggested tests, not completed results shown here. Look for a clear boundary and the right next step, without a made-up approval.

1:51Fix the cause and retest

When a test fails, fix the right thing. Missing facts belong in the knowledge base. Tone, boundaries, and escalation instructions belong in the agent settings. Change one thing, save it deliberately, and rerun the same questions. Keep a simple pass or review note so you can compare results.

A single good Sandbox answer cannot guarantee a successful launch. Repeat the checks after important changes, and test the real channel separately with your team. A reply in the preview doesn't prove that routing, delivery, or a human handoff works in production. Those need their own end-to-end check.

2:31Make testing a habit

Start small, test the uncomfortable questions, and improve what you find. That's how you build confidence before customers are depending on the answer. Visit Tuolen dot com.