Clixa Try for 7 days
EN

How do I test a change to the AI before customers see it?

In the AI Lab. You edit a channel's instructions as a draft, chat with the AI while playing the customer and see the automatic checks on the text. Nothing changes for customers until you click Publish, and each publication becomes a version you can roll back.

The risk of switching directly

Changing the AI's instructions is like changing an agent's script mid-shift: the next conversation already comes out different. One ambiguous sentence, one rule that contradicts another, and the customer is the first to see the mistake.

The Lab separates the two things: writing and testing on one side, publishing on the other.

The test conversation

You write as if you were the customer and the AI replies with what it would really use: the catalog, the knowledge base and the channel rules. The conversation isn't saved and nothing is sent to anyone.

Besides the text, the screen shows what the reply would have done in a real conversation: hand the conversation to a person, register an order, make a pre-booking, note down a customer detail. You can test the draft and compare it with the published version.

Each test reply uses AI credits, just like a real reply.

The automatic checks

When you save the draft, the Lab points out what usually causes problems:

  • empty or overly short instructions;
  • text longer than the limit the AI reads, with the count on screen;
  • asking for information the channel's intake already asks for, which would make the AI ask twice;
  • strong promises like "guaranteed" or "100%", which the AI would repeat to the customer;
  • missing hours and address, among the most common questions;
  • repeated passages and a missing message for when the AI can't answer.

An issue marked as an error can only be published with confirmation.

Versions and undo

Each publication saves how things were and how they ended up, with the date and who published. If the change wasn't good, "Undo this publication" brings back the previous version as a draft, to test and publish again.

Frequently asked questions

Which channels does it apply to?

For the website chat and for each WhatsApp number and Instagram account with its own AI. Channels that use another channel's AI are edited in the source channel, and the Lab warns who else is affected.

What can I edit there?

The AI's instructions and the message for when it can't answer. Catalog, knowledge base and channel rules stay on the usual screens, and the test conversation already uses whatever is registered there.

Who has access?

The account administrator, just like the screens that configure each channel's AI.

Updated on

More on AI customer service