It started in a client meeting at GPA in 2023, with the AI hype cresting. The retailer arrived with an open question: how can we use this? No spec, no wireframe, no predetermined use case.
Wine wasn't a hunch. It surfaced from research and working sessions with GPA's stakeholders and our team: a large, high-consideration category where customers hesitate and floor expertise is scarce. It was also the retailer's first serious AI product bet. Gimmicky would set the whole AI strategy back. Useful would unlock a roadmap. I shaped the answer end to end, from that first room to the kiosk on the floor.
One framing decision shaped everything downstream, and it was the one I had to make stick across the room: treat Paola as a service, not a feature. Prove the concept on a physical kiosk in a real store before touching the mobile app. Let people talk to her out loud, in the actual wine aisle, before betting on a screen.
Easy to say, hard to hold. Behind the framing sat a coordination problem: one consistent Paola across four areas of GPA at once. Marketing (brand voice, avatar build, naming poll), IT (data, integrations, agent orchestration on Azure), Legal (AI governance and customer-interaction policy), and the wine category and store ops (assortment logic, pilot store, floor-staff training). Products at this scale usually fail at the coordination layer, not the model layer.
Discovery ran from inside the client's building: interviews with category leads, store ops and the wine team, then customer research in the aisles. The problem showed its shape early. Unless you already know what you want, choosing a label by grape, region, body and price is easy and boring, and it isn't how anyone actually decides.
So we stopped asking what wine someone likes and started asking what they need today. A taste profile is fixed and fits in a database; an occasion is not, and has to be asked for out loud. That one word is the argument for a conversation instead of a filter, and it set the four needs the product had to meet: one-click purchase, easier search, contextual recommendation, purchase by occasion.
“Tonight I'm having dinner alone. Tomorrow I have a date.” Context, in a customer's own words
Three of these the customer never says out loud, one only a conversation can reach, and one belongs to the retailer. Getting them into a single reply is the actual product.
The MVP. Conversational wine sommelier on the in-store kiosk.
Live stock awareness so recommendations match what's on the shelf.
Pairings. Recipes, dishes, sides. Alongside the wine pick.
Identification carries the conversation across the kiosk and the app.
Personalised offers surfaced inline, not buried in a separate area.
Cart + order intake. The conversation closes with a transaction.
The avatar moved from references through hair and wardrobe studies to three finalist outfits and a final pick, with GPA's marketing team as the validation loop at every step. Each round came back annotated in their own words: "mais simpática (sorriso)", "escurecer cabelo", "levantar um pouco a manga". We dropped the off-the-shelf option early. Paola had to read as part of the company, not a vendor's mascot. Final direction: green sweater, plum wide pants. Reads as Pão de Açúcar on any surface without a logo doing the work.
The name came from inside the company too. The pitch went in as Club des Sommeliers. GPA ran an internal employee poll and Paola won. The voice was negotiated the same way, across marketing, IT and legal.
I directed the pitch video that carried the strategic call into the C-level room. Same team, same week, different language: less deck, more film.
With the strategy bought, I worked alongside the engineering leads on the system that would carry her across surfaces. One agent, multiple endpoints: kiosk in store, tablet at the counter, mobile app for browsing at home. Each surface tuned to its context, all of them speaking with one voice.
The MVP was built on Capgemini's Olivia asset. Re-skinned, re-named, re-trained. Orchestration on Azure OpenAI, cognitive search against GPA's Whitelabel Wine API, and phoneme correction on the top grape varieties so "Tempranillo" comes out right. I sat between design, GPA's IT and the engineering pod to keep experience and stack in step.
55" screen + 4K cam + motion sensor. Mobile companion. Front-end in Unity.
Prompt-tuned per surface. Phoneme correction on grape types. One personality, three contexts.
700+ labels, stock, recommendations. Cognitive search + custom feeder per surface.
Before the AI gets a turn, the totem has real UX work to do: age check, legal disclaimer, press-and-hold mechanic, none of which can feel like a gate. From there Paola listens and recommends, and two buttons carry the whole decision loop: "Quero esse" / "Quero outra sugestão".
The MVP shipped in store in eight weeks and proved the concept in front of real customers. On the strength of it we mapped a full roadmap: new capabilities layered onto the same agent, expansion from the kiosk to the mobile app, and a bigger slice of GPA's catalog for Paola to advise on.
The work didn't stop at the avatar. It became a playbook. Magic search, contextual upsell, recipe pairing, post-purchase advisory, entry-level wines. Each one a tile in the roadmap, each one designed to ship on the substrate the MVP proved.
Client · Pão de Açúcar (GPA)
Team · AIE Innovation Lab · Capgemini Brazil
Strategy, service & product design · Davi Serrano, with the AIE team