Test Your AI

Test every meaningful change before sending live traffic to an agent. Use a repeatable test set so results can be compared after prompt, knowledge, voice, or calling changes.

​

Complete the required setup

The Test Agent area requires a selected knowledge base and complete Speech to Text and Voice & Audio settings. If testing is unavailable, use the shortcut shown on the page to open the missing section. The Test Agent button offers Chat Test for written messages and Voice Test for a live voice conversation:

• Chat Test is disabled when Conversation or LLM Model settings are incomplete.
• Voice Test is disabled when Transcriber or Voice settings are incomplete.
• If no knowledge base is linked, the page shows Knowledge Base Required with Open Knowledge Base.
• If speech settings are incomplete, it shows Configuration Incomplete listing the missing Transcriber and Voice fields with Open Voice & Audio.

Check the speech-to-text provider and model, voice provider and voice, speaking speed, and interruption setting before judging the conversation.

​

Launch checklist

​

Use the available test experience

Open the agent, select Test Agent, and start a voice test. Speak as a real customer would rather than reading the prompt wording. Pause, interrupt, correct a detail, and ask a follow-up question to confirm that context is maintained. Chat Test first asks for privacy consent (Yes / No), then name and email (name up to 30 characters, valid email required), then opens Ask question… with Send and Reset Conversation. Answers stream in the panel. Voice Test shows the agent name, language, and live status (Connecting, Listening, Thinking, Hearing, AI is speaking, Microphone Muted) with a voice orb visualization. Use Mute / Unmute and End Call controls. The session uses barge-in so you can interrupt, and ends automatically if no audio is detected for about 45 seconds. Phone Call Preview demonstrates the call flow without placing a real call. Enter a phone number with country code (for example India +91 plus a 10-digit mobile), select Start Preview, and follow Initiating gateway → Ringing → Call Connected → Call Completed states. Keep the call identifier or conversation record for any failed case. It helps you compare what was heard, what the agent answered, and which setting needs adjustment.

​

Cover three types of scenarios

Run at least two cases per row - six tests minimum - before any launch. Record each as pass or fail with the call identifier so regressions are comparable across versions.

​

Build a reusable test script

Keep a permanent script document with columns: case ID, caller line (exact words), expected agent behavior, expected evidence (transcript line, outcome, transfer record), and pass/fail. Reuse the same script after every prompt, knowledge, voice, or schedule change, and append real caller phrasing harvested from live calls. A script that never grows goes stale within weeks because callers never phrase things the way builders do.

​

Test real conditions

Use expected accents, background noise, names, phone numbers, and a real provider call before launch. For each failed test, record the input, expected result, actual result, and related call ID. Decide whether the fix belongs in the prompt, knowledge, provider settings, or transfer configuration. Retest the failed case and one normal case after the change.

​

Verify after launch

Review early calls for unexpected questions and unclear outcomes. Monitor completion, transfer, failure, and duration trends, then add useful real-world cases to the permanent test set. Check the first 20 live calls in full, not a sample - early volume is low and each call teaches something. From call 21 onward, sample across outcomes: every failure, every transfer, plus a random slice of completions. Promote any caller question asked twice into the knowledge base or the permanent script within the same week it appears.