I recently presented at Microsoft’s monthly Power Hour session on Power Virtual Agents. One popular question that came up was “How can I know what unrecognised questions people are asking the bot?”. This post will take you through how to configure your PVA bot’s system fallback topic, and use Power Automate track unrecognised phrases entered in by your bot users. This kind of feedback is critical to improving your bot’s performance. You can use this feedback to add new topics to your bot or improve triggering on current topics.

The system fallback topic in PVA allows you to customise the actions and conversation that takes place if your bot is unable to recognise input from the user. To find it, click on the Settings cog in PVA.

Click Add for the fallback topic.

Once added, go to fallback topic.

You will see that there is already a conversation set up for the fallback topic. This can be modified however you need, as you do with any other topic (except you can’t add trigger phrases to it).


You can use Power Automate to create a workflow to notify yourself of the unrecognized user input. The method of notification here is up to you- e.g. you may choose to send yourself a Teams message, post it in a Teams channel, add it to a spreadsheet or database, etc. The flow below will simply send yourself a Teams message containing the unrecognized user input. Click + to add a new node > Call an action > Create a flow.

Build your flow. You will need to add a text input in the Power Virtual Agents trigger action as shown below. Then reuse/insert this input throughout the flow in your subsequent steps. You will later link this back to the unrecognised text used in the bot conversation. Save the flow and go back to PVA.

Back in PVA you may need to save your topic and refresh the page so that you can see the flow you just created listed.

Set the gets value from value to UnrecognizedTriggerPhrase.

When you are done with your topic, save it. It’s now ready to test. Remember to also publish your bot to make your changes live.

Gracias !