AI Supervisor Overrides

If you have a multi-Agent chatbot setup, GPT-trainer uses an AI Supervisor to automatically determine which user-facing Agent is best suited for handling the most recent input query. However, there may be times when you want more “deterministic” control over which Agent handles the job.

In this article, we demonstrate how you can use AI Supervisor Overrides to set up logic conditions that, when satisfied, directly switches to a designated Agent.

In the AI Agents tab, you will see a button that expands “AI Supervisor Settings” drawer:

After you open it, in the “Rules” tab, you will be able to specify conditional rules that get evaluated every time the user inputs a query into your chatbot. These conditions, if satisfied, will bypass the AI Supervisor’s own judgment and force a designated Agent to handle the current input query.

To add a new rule, click the “Create new rule” button. To edit rules, go to the three-dot-menu on the top right of the rule, then select “Edit” in the dropdown.

After you are satisfied with the rules, make sure to “Connect” them so they become active. Rules are placed in an evaluation queue. This means that as soon as a rule evaluates to TRUE, all subsequent rule evaluations will be skipped and that rule will be immediately executed.

A good example of AI Supervisor Override in action is the “delayed lead collection” setup. Consider the use case where you want to build a chatbot that handles general FAQ, but you want to start collecting user information after 2-3 questions. Furthermore, you want to make it so that if the user insists on not providing their information, your chatbot will continue serving them as an anonymous entity rather than completely deny further meaningful interactions.

First, you need a regular “Lead Collection” setup described in our Lead Collection article. Once you are finished, continue below.

At this point, you should have two user-facing Agents set up:

Then, open AI Supervisor Settings. Click “Create new rule”. In the three-dot-menu on the top right corner, select “Edit”. You will enter edit mode for the rule. For our use case, we set the following conditions:

This means that as soon as the “total # of user input queries” equals exactly 3, the AI Supervisor will set the Agent to be Lead Collection. Click “Save”.

Next, add another rule. This time, we apply the following conditions:

1. If the current Agent is Lead Collection,
2. AND, if the number of queries consecutively handled by the current active Agent is fewer than 4…
3. AND, if any of the user’s information (name, email, phone) is not specified…
4. THEN stay on the Lead Collection Agent.

This means that the Lead Collection Agent will continue to request user information until either it has conversed with the user at least 4 times or it has successfully collected all the requested information.

While you are in edit mode, click “Add Rule” to add a new condition for evaluation. At step 2 above, your rule setup should look like this:

Step 3 requires a “group” evaluation where the conditions are defined as:

• If Variable “user_name” is not specified…
• OR, if Variable “user_ email” is not specified…
• OR, if Variable “user_phone” is not specified…
• Then evaluate to TRUE

So we set up the following “Group”:

Then within the group, click “Add Rule” and set up the appropriate conditional statements:

Finally, as in step 4, make sure that the “Lead Collection” Agent is assigned when the rule is executed:

Connect your rules via the same three-dot-menu option “Connect”. If successful, they should display a green “Active” status:

Now, we can go to “Preview” and test the chatbot under this new set of AI Supervisor Overrides. You should notice the correct behavior being applied. In debug view, you should see an “override” tag next to Active Agent if an override rule is enforced.

You can design many complex workflows using AI Supervisor Overrides. However, this is an advanced feature that directly controls how the AI Supervisor behaves. We highly recommend that you test your configuration thoroughly before deploying it if you wish to make use of this feature.