The challenge
Spring break 2025 was approaching and Museon-Omniversum had a problem: not enough phone operators for the expected rush. Hiring and training extra people? That takes weeks. The holiday started in ten days.
The solution: an AI phone assistant that's faster to train than humans. In just over a week - weekend, work week, weekend - from zero to live.
Speed over perfection
This project was about one thing: going live on time. No months of preparation, but a working solution that we improved along the way.
The core approach:
We launched with a basic assistant and optimized daily based on real conversations. Every day the assistant got better. Visitors who didn't want AI were immediately transferred to a human - no forced menus or frustration.
Two content sources covered 90% of questions:
Daily agenda sync via scraping
A Playwright script fetched current shows, movie times, and exhibitions daily. Including automatically clicking "load more" until all content was in. Output to Google Docs, from there copy-paste to Bonnie. Pragmatic? Absolutely. But faster than waiting weeks for an API integration.
Complete FAQ
All frequently asked questions about tickets, parking, accessibility, and practical matters.
The details that make the difference
Building an AI phone assistant is one thing. Making it sound natural is an art. These parameters proved crucial:
Speech rate
Response delay
The sweet spot: 500-700ms delay prevents the AI from talking over the caller, while keeping the conversation flowing. For complex answers, 700-1000ms can still feel fine.
Listening pauses
The assistant waits briefly for signals from the caller before continuing. This prevents the "talking over each other" that makes many phone AIs annoying.
Who did what
This project was a joint effort:
The numbers
From first conversation to live in production: a few days. Then 10 days running during spring break - exactly the peak period we were solving for.
What stood out: visitors who explicitly asked for a human were transferred immediately. No queues, no "I don't understand you". This happened surprisingly rarely - most callers got their answer from the AI and hung up satisfied.
What this delivers
Without weeks of recruiting and training.
Staff available for complex questions.
Daily content sync means never outdated answers.
Direct escape to human, no detours.
Daily optimization based on real conversations.
Bonnie (award-winning telephony AI)
ASN telephony platform
Playwright scraping → Google Docs → Bonnie (daily)
FAQ + show information
Direct transfer to staff
What we learned
A working 80% solution today is better than a perfect solution in three months. The holiday won't wait.
A Playwright script to Google Docs to copy-paste doesn't sound sexy. But it worked within a day, while a "real" API integration would take weeks. Pragmatism wins.
By launching with a basic version and improving daily, the assistant got better while already delivering value.
Knowing they can always reach a human, callers accept the AI more easily. Paradoxically, the exit ensures people use it less.
Speech rate and response delay sound like details, but determine whether an AI conversation feels natural or annoying.
Need more control?
For many telephony challenges, a platform like Bonnie is the fastest path to results. But sometimes the situation calls for more customization: a specific voice, deeper integrations with existing systems, or unique conversation flows.
In those cases, we build a custom AI Assistant together with the client using ElevenLabs - from voice selection to system integrations. The same conversational intelligence, but fully tailored.
Check out these cases
Similar challenge?
Need to handle peak demand without months of preparation? Curious what AI telephony could mean for your organization?