🤖 HeyWowza · Live Analytics

We let an AI loose in a science museum.
5,012 people walked up and talked to it.

For nine days, HeyWowza stood in Technopolis and answered anyone who pressed the button. Then we read every single conversation. Here’s what 5,012 of them taught us, with the numbers to prove it.

Scroll and discover
💬 The headline
5,012
conversations · 14–22 Feb 2026 · kiosk “HeyWowza”
≈557
chats / day
a busy little robot
97%
spoke out loud
4,862 talked · only 150 typed
3.2s
avg reply
down from 7.6s at launch
💬 What they asked

It wasn’t toilets, opening hours or parking.

On average, each visitor asked 3.5 questions. And the single biggest category wasn’t the practical stuff. It was visitors trying to understand what they were looking at.

✦ Exhibit questions32.3%
➤ Wayfinding14.6%
➤ Questions about Wowza13.0%
➤ Age recommendations4.5%
★ Kids saying “67” 😄1.1%
“How does Maxwell’s Wheel work?”exhibit question
“How does a thermal camera work?”exhibit question
“How do clouds form?”exhibit question
“What should I know about the plasma ball?”exhibit question
“Where are the toilets?”wayfinding
“Where is the maze?”wayfinding
“Are you a robot? How old are you?”questions about Wowza
“Why do you look like a donut?”questions about Wowza
“I’m here with a 4-year-old, what do you recommend?”age recommendation
🙋 What it means

They weren’t just looking things up.

Visitors aren’t only using the assistant for practical information. They’re using it to understand what they’re experiencing. A few other categories stood out:

WayfindingWhere are the toilets? Where is this exhibit? How do I get there?14.6%
Questions about WowzaWho are you? Are you a robot? How old are you?13.0%
Age recommendations“I’m here with a 4-year-old, what do you recommend?” “What can I do with a 13-year-old?”4.5%

A small note on the data: every question was placed in one category only. Real visitor messages often mix things together. “You’re stupid, where is the toilet?” becomes wayfinding, not abuse, because the real intent wins. That keeps the categories clean and easy to compare.

AI in visitor environments isn’t just about answering FAQs. It’s about understanding what visitors need in the moment: orientation, explanation, recommendations, reassurance, play, and sometimes just a little bit of chaos.
🌍 Tongues

Three languages, one clear favourite.

This is simply the mix of languages visitors picked on the kiosk before they started. Dutch ran away with it, while French and English split the rest.

🇳🇱 Dutch (nl-NL) · 4,185 chats83.5%
🇧🇪 French (fr-BE) · 536 chats10.7%
🇬🇧 English UK (en-GB) · 291 chats5.8%
🧠 Memory

And then it started to remember.

When two people ask near-identical things, HeyWowza recognises it and replays the answer from memory instead of rebuilding it from scratch. Skip the heavy thinking, skip re-recording the voice, just answer.

Data flowing into a brain
Day 1: a total stranger.Only 2.3% of answers came from memory. Everything was brand new.
Day 7: nearly 1 in 4.The cache hit rate climbed to 23.4% as the same questions kept coming back.
Every remembered answer saved 32.3 credits……and shaved 4.0 seconds off the reply. Same answer, a fraction of the effort.
≈15,800 credits saved in 9 days.Purely from recognising “oh, I’ve heard this one before.”
📈 Watch it learn

Memory, day by day.

Each bar is the share of answers served straight from memory. The trend only goes one way: up and to the right.

2.3%
14 Feb
1.9%
15 Feb
2.6%
16 Feb
6.7%
17 Feb
14.9%
18 Feb
18.8%
19 Feb
23.4%
20 Feb
23.4%
21 Feb
26%
22 Feb
🏆 Cache ROI

Remembering is 6× cheaper and 3× faster.

A fresh answer runs the full pipeline. A remembered one barely breaks a sweat. Same visitor, same question, wildly different bill.

Fresh answer
38.4
credits / reply
⏱ 6.0s to respond
From memory
6.2
credits / reply
⏱ 2.0s to respond
32.3
credits saved / hit
mostly the voice never re-recorded
4.0s
faster every time
6.0s → 2.0s
⚡ Follow the seconds

And where does the time go?

Half of every reply is HeyWowza looking things up, searching its knowledge base (RAG) for the right facts before it ever opens its mouth. That’s the bottleneck, and the next thing we’re optimising.

🔎 Looking up the facts (RAG) · the bottleneck4.2s
🔊 Speaking the answer1.1s
👂 Understanding the audio1.0s
🧠 Checking its memory0.8s
💭 Forming the thought (LLM)0.8s
🚨 The boring (good) bit
99.80%
of 5,012 conversations went off without a hitch
10
errors total
in nine days of non-stop chatter
0.20%
error rate
you’d never notice
100%
uptime
online every minute the museum was open
🎤 The takeaway

5,012 hellos, and it’s only getting smarter.

People treat a good AI like a person, so we built one worth talking to, that learns from every conversation and gets cheaper and faster as it goes. This is what HeyWowza does in the wild.

Give people something genuinely curious to talk to, and they’ll keep coming back to understand the world around them.

Source: HeyWowza interaction logs in BigQuery · kiosk “HeyWowza” · 14–22 Feb 2026 · 5,012 turns · 13.1% overall cache hit rate · 0.20% error rate. Numbers rounded for readability.