Hello, Jana.
Jana is an AI avatar for everyday shopping questions. For two weeks, the AI avatar helped customers at EDEKA Müller in Babelsberg find their way around the store, learn about current offers and get answers to general questions about the store and their shopping.
01 / The starting point
Big ideas can start just around the corner.
Jana is an AI avatar for everyday shopping questions. Together with store owner Randy Müller, unframe brought the digital assistant to EDEKA Müller in Potsdam-Babelsberg for two weeks. Positioned at the entrance, the AI avatar helped customers find their way around the store, learn about current offers and get answers to general questions about the store and their shopping.
The guiding question: Can a conversational avatar provide directions, handle recurring questions and meaningfully complement the shopping experience?
02 / The implementation
Answers through direct conversation.
The AI avatar was available at the store entrance. The Holobox displayed it with a spatial 3D effect. Its knowledge base covered the store’s products and offers.
Just ask
Information about products and offers through a direct conversation with the AI avatar.
Finding your way
Help with locating products and the right area of the store.
50+ languages
Customers could ask questions in different languages, making conversations in store easier.
03 / In-store moments
Right there. In store.
The AI avatar at the entrance, in conversation with customers and alongside the project team on site.
The AI avatar in practice.
The setup, deployment and idea behind the pilot project in Babelsberg.
04 / Pilot usage
Two weeks. Many conversations.
Customers and staff welcomed the AI avatar, which enabled many interactions on site. We were able to respond to feedback quickly and precisely, helping the pilot run smoothly in store.
1,500
~392*
1,145
(+192 % ↑)
The pilot, day by day
17–30 August 2026 · 14 calendar days · interactions per day
Daily values as a table
| Day | Recorded interactions |
|---|---|
| Week 1 · Mon | 49* |
| Week 1 · Tue | 28* |
| Week 1 · Wed | 35* |
| Week 1 · Thu | Holobox out of service |
| Week 1 · Fri | 103 |
| Week 1 · Sat | 177 |
| Week 1 · Sun | Closed |
| Week 2 · Mon | 245 |
| Week 2 · Tue | 105 |
| Week 2 · Wed | 161 |
| Week 2 · Thu | 153 |
| Week 2 · Fri | 129 |
| Week 2 · Sat | 352 |
| Week 2 · Sun | Closed |
Counts represent interactions, not individual people or sessions.
*Counted interactions; tracking did not yet capture all enquiries during the first three days. The Holobox was out of service on Thursday.
Source: supplied daily pilot figures.
05 / Pilot-period analysis
Categories
Topics, product interests and conversation starters at EDEKA Müller
17–30 August 2026
At the entrance · Open questions without introductory or suggested prompts
The Holobox stood at the store entrance. This placement shapes the context of use: questions about products and departments meet the start of the shopping trip.
- Analysed contributions
- 1,011
- Finding products
- 31%
- Shopping-related topics
- 45%
- Interest in the avatar
- 9%
The pilot period’s topic profile
Share of responses · Nonlinear scale (power 0.75)
| Finding products | 31.1% |
|---|---|
| Avatar | 8.9% |
| Offers | 6.3% |
| Product info | 4.7% |
| Buying advice | 2.6% |
Selected topics by day
Topic references including search queries · August · Original chart in German
Directions shape usage
Around 31% of the analysed responses concern finding products. Practical questions come first at the entrance: where is the bottle-return machine, where are the milk or eggs? The Holobox can handle recurring requests for directions and support the store team with routine questions.
The topic mix broadens
The share of offers, product information and buying advice rises from 8.6% to 14.7% in the analysed week-to-week comparison. Product selection therefore becomes more prominent. Current offers and suitable alternatives are key topics for advice.
Capture specific product interests
Club-Mate, Knabe Cola and Pokémon cards show how specific product requests can be. Questions about toys and dinosaurs are concentrated on 29 August. These topic clusters give the store team a basis for checking the range and improving product information and directions.
Curiosity starts the conversation
Around 9% of the analysed responses concern the avatar itself, its appearance and how it works. The Holobox is both a source of information and an attention-grabber. A simple question such as “What are you looking for?” connects this curiosity with a specific shopping need.
The 25 most frequent topics and terms
Aggregated tag mentions
- Finding products314
- Avatar101
- Offers64
- Greetings60
- Product information48
- Small talk42
- Fruit28
- Buying advice26
- Bottle returns24
- Eggs18
- Ice cream18
- Vegetables17
- Drinks16
- Gummy sweets15
- Milk15
- Salad bar13
- Meat12
- Haribo12
- Cola10
- Randy10
- Baked goods9
- Coffee9
- Entertainment9
- Opening hours9
- Baking ingredients8
How to interpret the data: Radar and headline figures: 1,011 response contributions. Topic bars include search queries; a contribution may cover multiple topics. Shopping-related: directions, offers, product information and buying advice. Follow-up questions may be counted more than once.
Privacy: During regular operation, analysis is performed directly by categories and tags, without assigning data to people or conversations and without storing conversation content. Conversations are not tracked. The pilot analysis uses a one-off evaluation of response texts.
Source: EDEKA Müller pilot data · Analysis dated 15 September 2026
07 / Our conclusion
What worked well in store.
In natural conversations, the AI avatar answered general questions, helped customers find their way around the store and provided information about products and offers. This worked smoothly in more than 50 languages.
Starting a conversation
For many customers, interacting with the AI avatar was an exciting experience. It led to many interesting conversations and lively exchanges right in the store.
Answering everyday questions
The AI avatar provided useful answers to simple questions about products and offers. This made the value of an additional point of advice tangible in daily use.
Promoting inclusion. Reaching more people.
Multilingual support promoted inclusion and made access easier for visitors with different language skills. They could describe what they needed in a familiar language.
Identifying challenges. Improving directly.
Valuable feedback from the trial highlighted clear potential for improvement: tracking was unreliable during the first three days, and some promotional products were not recognised in follow-up questions. The flexible setup allowed us to resolve both issues during the pilot.
The pilot showed us what already works well in everyday use and where we can make targeted improvements. Customers’ questions and the store’s feedback give us a concrete basis for that work. One focus is the avatar’s voice output, which is being optimised step by step. Another frequently mentioned issue was the lack of subtitles. We have since resolved this: subtitles are available in the current version and help people follow conversations even with loud background noise.
Potential for software and hardware extensions
Questions and feedback from the store are informing further improvements. The pilot also offers considerable potential for software and hardware extensions: product images could complement advice, while enhanced wayfinding could improve orientation. The modular architecture enables connections to a wide range of services. Flexible placement, for example on tablets as specialised points of advice for specific food categories, opens up further possibilities.
Qualitative assessment by the project team, not a measured satisfaction or efficiency metric.
06 / Social media response
Experienced in store. Shared online.
Jana was also a popular topic online. Posts sparked ideas and lively conversations and encouraged people to visit the store. Feedback was consistently positive. With impressions well above earlier posts, the pilot gave the store a real boost on social media.
LinkedIn and Instagram, as of September 2026. Supplied lower bounds.
Impressions and interactions are reported as lower bounds. Interactions combine reactions, likes and comments.
The approximate 101% increase refers to the supplied average impressions (1,547 versus 3,102).