Three campaigns, three holidays, three shopping-mall kiosks. Christmas at Mammut, International Women's Day at the same mall, Easter at Aréna Pláza. A voice AI gift advisor ran through all three, built into a kiosk with an array microphone: the shopper walks up, says who they are buying for, the agent suggests something and points them to the right store.
In total, more than 1000 real interactions passed through these three campaigns between December 2025 and April 2026. This was not a demo or a lab pilot but a live campaign run for Hello Agency.
So what do three campaigns teach you about what a shopping-mall voice AI kiosk is good for, what it is not, and what separates a demo that sounds good from a campaign that runs reliably? Here are the numbers and the lessons.
The three campaigns, briefly
- Mammut KARI-AI (Christmas): Dec 11-24, 2025 — 14 days — 433 interactions
- Mammut Women's Day: Mar 3-11, 2026 — 9 days — 323 interactions
- Aréna Pláza Easter: Mar 23 - Apr 2, 2026 — 11 days — 244 interactions, of which 123 were substantive gift advice
All three ran continuously with zero system failures. The one technical problem that did occur, during KARI-AI, was not the agent.
What a kiosk agent is actually good for
When most people hear "AI gift advisor" they picture a Netflix-style recommendation engine: collect data, analyze, suggest. This is not that. This is a spoken, persona-based advisor that talks to a shopper standing at the kiosk for 60-120 seconds and sends them off with a specific store to visit.
In the Easter campaign at Aréna Pláza, out of 123 substantive conversations the agent successfully identified who the gift was for in 64 cases. The top five:
- A child — 13 conversations (20%)
- My mother — 10 (16%)
- My girlfriend — 9 (14%)
- My partner — 9 (14%)
- Other (grandmother, colleague) — 2
These are not theoretical categories but the personas that genuinely exist in the mind of someone walking into a mall to buy a gift. A kiosk agent supports exactly that moment of the decision: the shopper is already inside, already intends to buy, and just does not know which store to head for.
In the Easter campaign, 46 substantive conversations ended with a specific store recommendation (RÉGIÓ Játék, GUESS Kids, Lovisa, Libri, Crystal Nails — from the Aréna Pláza portfolio). That is the real value: not an abstract suggestion but "go up to the second floor, Libri is on the left, you will find a book for the child there".
What it is not good for
This is where projects usually go wrong. A voice AI gift advisor kiosk is not suited to:
- Live pricing: if the knowledge base has no current price, the agent should not state one. Better to say "go into the store, they will tell you exactly".
- Live store availability and stock: the kiosk is not connected to the stores' tills. If a store is closed that day or something has sold out, the agent does not know.
- Complaints, lost property, security: sensitive, human responsibilities. We keep them out of scope.
- Complex wayfinding: "which exit is closer to the metro?" — without structured map data, do not force it.
- Running continuously, all year: the logic is campaign-based. At Mammut, Women's Day is a 9-day window and Christmas a 14-day one. A kiosk running permanently loses its value between campaigns.
Push these in anyway and the system will sound good but fail in operation.
How it behaves at the peak of a campaign
One of the most important questions: what happens on the critical shopping day? That is where you find out whether the system was only ever a demo.
Mammut Women's Day 2026 — March 7 (the eve of the holiday): 82 interactions in a single day. That is close to 25% of the entire 9-day campaign concentrated into one day — served steadily, with no downtime.
Aréna Pláza Easter 2026 — March 25-28, the peak stretch: 29-32 interactions a day across four consecutive days, close to 49% of the whole campaign's traffic. Busiest day: Friday March 27, with 32 interactions, arriving in a wave between 18:00 and 20:00.
Neither campaign had a failure or an outage of our own. That is not luck: platform-level stability is a baseline requirement, not an extra.
What it takes to make this work
After three live campaigns, these are the real operating conditions — from log files, not from theory:
1. Good hardware, not just a good model
On the first campaign, at Christmas, a subcontractor's array microphone was not loud enough for the acoustics around the kiosk. The agent knew its job, but the audio did not come through cleanly. The next campaign went in with a better microphone — and that produced one of our most important lessons: in a kiosk installation, the quality of the microphone and the speaker matters as much as the choice of model.
2. A structured knowledge base, with weighting
It is not enough to know that "this is a toy store". The knowledge base records which persona typically shops there, which age group it suits, and where it sits inside the mall. Hello Agency delivered this fresh for every campaign: scrape, personas, weighting. Without the knowledge base the agent just chats in generalities instead of recommending anything.
3. A kiosk conversation style
Across the 123 substantive conversations of the Easter campaign, the audit found 100% tone consistency and 91% on-topic. That does not happen by itself: the scope, the way it opens, and topic discipline all have to be trained in. The moment the system starts wandering into open topics, the shopper walks away after 20 seconds.
4. Campaign-based rollout
The Mammut Women's Day agent was not a new build but a seasonal retune of the KARI-AI Christmas base — cheaper and faster, and just as stable as a new build. The Aréna Pláza Easter version, by contrast, was built from scratch for a new mall and a new store portfolio. Whether the base can be reused is an important design decision: a kiosk agent is not a one-off project if it is built properly.
A real conversation, or just trying it out?
Of the 244 interactions in the Easter campaign, 123 were substantive gift advice — longer than 10 seconds, with at least two exchanges. In the other 121 the shoppers "just tried it": walked up, said a word, had a look, moved on. Those are not failures but the natural pattern of kiosk use.
Distribution by length:
- 22% — under 10 seconds ("just a glance")
- 21% — 10-30 seconds ("one attempt")
- 30% — 30-60 seconds (a conversation getting started)
- 16% — 60-120 seconds (genuine advice)
- 11% — over 2 minutes (a deeper look at preferences)
Substantive conversations ran 48.8 seconds on average, with a median of 38 seconds and a longest of 231 seconds (about 4 minutes). Over the campaign's 11 days, cumulative live shopper interaction came to 176 minutes — close to 3 hours.
That pattern matters, because a kiosk is not built on "everything for everyone". If someone gets useful advice after 30 seconds and heads off to the store to buy, that is the success — not several minutes of engagement.
What to do if you were starting now
If a shopping mall, a retail chain, a bank, or any campaign-driven organization were starting a kiosk project today, the shortest route to live operation is these four steps:
- Design for a campaign window, not continuous operation. Six to fourteen days per campaign, then it rests. Cheaper, more flexible, and you learn faster.
- Hand over the knowledge base in a structured form — store, product, persona, weighting; not as a PDF. Without a scraper or a structured source, the first week goes on this.
- Spend on hardware; do not save on the models. The quality of the array microphone, the speaker, and where the kiosk stands matter as much as the choice of model. Without those, the best AI still speaks quietly into an acoustically tiring space.
- Measure at audit level, not just volume. Tone, staying on topic, identifying the recipient, recommending a store. Those are what show whether it "works well" rather than just "sounds good".
In short
Three campaigns, one client group, 1000+ real interactions. The main lesson is not that "voice AI works" — we knew that. It is that a shopping-mall kiosk agent is a well-bounded use case you can roll out per campaign, where the narrow scope (gift advice plus in-mall directions) and the kiosk conversation style together produce the stability.
The right first step is not choosing a platform. The right first step is a 6-14 day campaign window, a structured knowledge base, a well-chosen microphone, and a tightly bounded scope.
If a similar campaign-based retail or in-store AI project is on your table and you would like to work out which campaign window makes the best first pilot, we are happy to talk it through — book a short call.