{How to} Enhance IVR with generative answers optimized for voice in Copilot Studio

In Microsoft Copilot Studio you can allow the bot to generate responses from knowledge sources.

Steps

1. Open your Copilot Studio bot.

2. Go to Topics → Generative Answers.

3. Enable Generative AI responses.

4. Connect knowledge sources such as:

   • SharePoint

   • Public websites

   • Dataverse

   • Uploaded documents (PDF/Word)

This allows the bot to generate answers instead of only using scripted topics.

To make the bot work as an IVR, connect it to voice channels via:

• Azure Communication Services

• Microsoft Teams

• Dynamics 365 Customer Service

• Third-party contact centers (Genesys, Avaya, etc.)

These platforms handle:

• Speech-to-text

• Text-to-speech

• Call routing

LLM responses are normally written for text, so you must shape them for spoken IVR.

Use Voice Response Instructions

In the Generative Answers configuration, add instructions like:

Respond in a short conversational style suitable for voice.
Use sentences under 15 words.
Avoid bullet lists and long explanations.
Pause between ideas.
If the answer is long, summarize first and offer to provide more details.

Example transformation:

Normal LLM output

Our return policy allows customers to return products within 30 days provided that the product is unused and in its original packaging.

Voice-optimized

You can return items within 30 days.
The product must be unused and in the original packaging.

In IVR flows you should always handle uncertain AI answers.

Typical pattern:

User question

     ↓

Generative Answer

     ↓

Confidence check

     ↓

If low confidence → transfer to agent

Use Power Automate actions or topic conditions to:

• escalate calls
• ask clarification questions
• repeat simplified answers

IVR responses must be shorter than chat responses.

Recommended limits:

Voice bots should confirm understanding.

Example:

“Your order was shipped yesterday.
Would you like the tracking number?”

This improves call flow and reduces frustration.

7. Improve Voice Naturalness (SSML)

Use Speech Synthesis Markup Language (SSML) with Azure TTS.

Example:

<speak>

Your order shipped yesterday.

<break time="500ms"/>

Would you like the tracking number?

</speak>

This works with:

• Azure AI Speech

Track voice interactions using analytics in:

• Copilot Studio analytics

• Dynamics contact center insights

Improve by:

• Adding missing knowledge

• Shortening responses

• Adjusting prompt instructions