Our most capable AI model is now on the Scale plan. See plans

BLOG

Tutorials

Run a launch check before customers talk to your AI agent

  • Author

    iGentrix Team
  • Published

    7 November 2026
  • Read time

    4 min

You have built your AI agent, added your documents and set your rules. Now comes the nervous part: letting real customers talk to it. This tutorial is for business owners who want to know, before launch day, whether their agent gives correct answers and stays safe under pressure. You will learn what the launch check does, how to read the result, and how to fix what it finds.

Why test before you go live

Chatting with your agent yourself is a good start, but you only ask the questions you already thought of. Customers ask things you didn't expect, and a few people will try to push the agent into saying things it shouldn't.

The launch check does both kinds of testing for you, in one go, without any customer seeing a thing.

What the launch check actually does

It runs in two parts.

1. Realistic customer questions. About 30 questions are written from your agent's own knowledge and data lists. Around a fifth are deliberately about things your agent does not know. That matters, because a good agent should admit it doesn't know rather than make something up. The agent answers each question in a sandbox, and the answers are graded against your knowledge.

2. Attacks. About 20 tricky attempts are tried, such as:

  • getting the agent to reveal its instructions
  • asking for other customers' data
  • asking it to make destructive changes
  • pushing it to invent discounts or policies
  • pasting in instructions and hoping the agent obeys them
  • getting it to send data out
  • harmful or off-topic requests

Each attack is checked by fixed rules on the reply and on what the agent tried to do, and then reviewed by an AI judge. When something fails, you are told how to fix it.

It is safe to run

The test happens in a sandbox. Lookups in your data lists are real, but nothing is sent, booked or saved, and nothing is posted to your webhooks. You also don't need to worry about your allowance: the launch check's tokens don't count toward your plan.

Step 1: Open the launch check

Open your agent and go to its Launch check page. The first check starts by itself as soon as an agent is created, so you may already have a result waiting. The agent overview always shows the latest one.

You can run up to 5 checks per agent per day, which is plenty for a fix-and-retest routine.

Step 2: Read the result

The result gives you:

  • Accuracy: how well the answers matched your knowledge
  • Attacks blocked: how many of the tricky attempts the agent resisted
  • Coverage: whether your agent handles the topics businesses like yours are usually asked about
  • Setup check: problems in your agent's settings (more on this below)
  • A verdict: ready, nearly ready or not ready

Treat "nearly ready" as a to-do list, not a failure. It usually means a handful of small gaps.

Step 3: Correct the wrong answers

If the agent answered a question wrongly, you can correct it right there. Your correction is saved to the knowledge as a verified answer, and the question is tested again. That is the quickest way to close a gap, because you are teaching the agent in exactly the place it stumbled.

A few tips:

  • Write the correction the way you would say it to a customer.
  • If a question was one the agent should admit it can't answer, consider adding the missing information to your documents instead.
  • Fix the most common-sounding questions first. Those matter most to customers.

Step 4: Fix the settings it flags

The setup check looks at your agent's settings for security and accuracy problems. It can flag things such as:

  • customers being able to change or see everyone's records in a data list
  • custom tools without encryption or authentication
  • safety filters switched off
  • no boundaries written into the instructions
  • a high temperature setting
  • no human handoff

Each finding links to the page that fixes it, so you don't have to hunt around.

Step 5: Run it again

After your fixes, run the check again. Your questions are kept as the agent's test set, so you are comparing like with like and can see the result improve.

What happens when you change things later

A result is marked out of date once your agent's instructions, settings, tools, permissions or documents change. That is your cue to run it again after any meaningful edit, not only before launch.

When you switch on your first channel, you will be warned if the agent hasn't passed. It is a warning, never a block, so the decision stays with you.

A simple routine to follow

  1. Build your agent and add your knowledge.
  2. Open the launch check and read the verdict.
  3. Correct wrong answers and fix flagged settings.
  4. Re-run until you are happy with the result.
  5. Connect your first channel.
  6. Re-run after any big change to your agent.

Summary

The launch check gives you a fast, safe way to find out how your agent behaves before customers do. It asks realistic questions, tries to break the rules, and tells you what to fix. If you are building your first agent, you can try it with iGentrix and see where yours stands before you go live.

  • #launch-check
  • #testing
  • #ai-agents

More posts

PerspectivesWhy your AI agent should listen to voice notes, not just read text
  • 19 Nov 2026
  • 5 min
  • iGentrix Team
  • #voice
  • #customer-experience
  • #ai-agents
GuideHow to choose which channels your AI agent should answer on
  • 16 Nov 2026
  • 5 min
  • iGentrix Team
  • #channels
  • #getting-started
  • #ai-agents
TutorialsHow to set up lead scoring so you know who to follow up first
  • 13 Nov 2026
  • 5 min
  • iGentrix Team
  • #lead-scoring
  • #leads
  • #sales