How we scaled our customer interview process with ElevenLabs Agents

We used ElevenLabs Agents to interview over 230 users of our ElevenReader app. 

In this post, we share how we built the voice agent, the results of this production test, and how you can use these tools for your own product improvement.

The challenge of scale

We value customer interviews, but scaling them is difficult. A typical 15-minute live interview provides great insights, but scheduling more than a few per day is a logistical challenge.

Calendars rarely align, and supporting a global user base across dozens of languages is nearly impossible for a small team - having in-depth conversations around the clock is a physical limitation.

While surveys are easier to scale, they are often lossy. They flatten feedback into multiple-choice options and miss the emotion and nuance of a 1-1 conversation. With advances in voice AI and LLMs, we can now bridge that gap. 

We built an AI Interviewer with ElevenLabs Agents to collect feedback from users through real conversations. We hosted over 230 interviews in under 24 hours, and we’ve already shipped improvements to the app using these customer insights.

Creating the AI Interviewer

We used the ElevenLabs Agents platform to create a conversational researcher. Our goal was to understand ElevenReader app users' perceptions in four main areas:

• Feature requests and improvements
• Primary use cases
• Competitor comparisons
• Pricing and brand value

We chose the voice “Hope - The podcaster” for its friendly, conversational tone  - it feels like sitting across from an empathetic researcher. For the logic, we selected Gemini 2.5 Flash to balance low latency with high intelligence.

Prompting and Guardrails

We crafted a system prompt that instructed the agent to ask follow-up questions for deeper insights while keeping the conversation on track. If a user responded with vague or one-word answers, the agent would prompt them for more detailed feedback. Before launching, we used ElevenLabs simulated tests to ensure the agent handled edge cases, such as vague responses or inappropriate language.

See here for the system prompt we used.

Data Collection

We used the Analysis feature in ElevenLabs Agents to evaluate each call. This tool extracts structured data from transcripts, allowing us to turn open-ended conversations into tangible insights. For example, we could automatically track the responses to:

• "How are you primarily using ElevenReader today?"
• "What two ways would you improve the app?"

The agent used the end_call tool to close conversations after ten minutes and politely thank the user for their time.

The Results

Within 24 hours, we collected over 36 hours of total conversation time.

• Success Rate: 85% of calls were successful and on-topic.
• Engagement: The median call lasted 10 minutes (the maximum allowed).
• Depth: One conversation lasted 87 messages. The median was 25 messages. 
• Cost: Each 10-minute call cost just under $1.00, or 9 cents per minute to complete, all hosted through ElevenLabs Agents

Analyzing the Data

We used Claude Opus 4.5 to analyze the 36 hours of transcripts for trends and insights based on UX research principles.

While the model provided high-level themes, we refined the analysis with additional prompts to surface granular insights  such as user segmentation, navigation feedback, and price sensitivity by region.

To share these results internally, we built an interactive artifact with Claude. Our team can now click into specific data points to see the exact user quotes that informed the trend.

Key Findings

Users were comfortable speaking with an AI -  almost 95% of respondents engaged directly with the interviewer without acknowledging it was a conversational agent. One user said:

“Well, this customer service interview is the most remarkable AI experience I think I've ever had. I wish all questionnaires were like this and I wish all customer service digital services were like this.”

We learned:

• Segmented needs: 21% of ElevenReader’s fiction readers requested multi-character dialogue, a much higher rate than other user segments.
• Brand value: Users associate ElevenReader with the freedom to listen to any book, anywhere, with the most natural-sounding voice quality.
• Languages needs: Users identified areas where the app’s Text to Speech (TTS) model confuses languages and accents  - clear areas for improvement. 
• Bug Reports: The interviews highlighted specific issues that our engineering team fixed the following day.

Future Implications

The future of user research is conversational - AI voice agents let you listen to and speak with your users globally on their own time.

This test showed how realistically and reliably AI agents can conduct in-depth interviews at scale. Paired with text analysis by LLMs, these conversations revealed patterns across hundreds of responses - insights that would have been difficult to uncover manually.

You can build a similar AI interviewer with ElevenLabs Agents - start building today or contact our team to learn more.