Work / Nuvemshop

Teaching an AI assistant to sell

Nuvemchat answers a store's WhatsApp customers with AI. Setting it up depended on the sales team sitting next to each merchant. I designed a setup merchants could do on their own, in five steps, testing the assistant before it talked to a real customer.

Role: Product Design Specialist, AI
When: 2024 to 2025
Product: Nuvemchat, an AI assistant for store chats
Focus: Research, onboarding, AI setup

7merchants interviewed, six in Argentina, in Spanish
18opportunities mapped for setup and for managing conversations
5steps from zero to a tested assistant
Livethe new setup went live in Nuvemshop's admin

Context

Nuvemshop is Latin America's largest e-commerce platform. I joined as Product Design Specialist for AI, working with a team in Argentina. Nuvemchat was in test: an AI assistant that answers a store's customers on WhatsApp and Instagram, using what it knows about the store and its products.

Merchants liked the demo. The assistant was smarter than the step-by-step chatbots they had used before, and it already knew their store. The trouble started after the sale.

What the research said

I ran semi-structured interviews in Spanish with seven merchants: one close to canceling, three who had paused the assistant, two still setting it up, and one who recommended it to others.

Setup felt easy, but only because someone from sales did it with them.
Merchants expected to tune the assistant little by little, as real customers asked real questions.
After turning it on, most of them opened every conversation to check what the assistant was saying. Seeing was believing.
When something was off, they called sales again, or paused the assistant.
They couldn't find settings on their own, so sales ended up as tech support. That doesn't scale.

A setup so easy that merchants could do it alone, and see the answers before their customers did.

What I designed

A five-step setup inside the store admin, with a short tip on each step for first-timers:

Personality. Traits like friendly or formal, other details in the merchant's own words (with a "generate with AI" shortcut), emojis on or off, and how long answers should be.
Product information. Permission for the assistant to use the store's catalog.
Extra information. Store hours, shipping, payment, returns. An example card shows what good information looks like.
Test. The merchant chats with the assistant as if they were a customer, with suggested questions, and goes back to adjust anything.
Connect. Only then, WhatsApp.

Chat page before setup: 'Answer your customers automatically with AI', with a button to start the setup. — Where it starts: the chat page explains what the assistant does and opens the setup.

Step 1 of 5, response personalization: personality traits, other characteristics with 'Generate with AI', use emojis, and short, medium or long answers. A tip explains the personality field. — Steps 1, 3 and 4: personality, extra information, and testing as a customer. Screens in Portuguese, with test data.

Step 3 of 5, extra information: an example card and a dashed card to add more information about the store, with a tip pointing to the example. — Steps 1, 3 and 4: personality, extra information, and testing as a customer. Screens in Portuguese, with test data.

After setup

The research showed that tuning never stops. So everything from setup lives on in settings, in four tabs: preferences, library, test and channels.

The library holds what the assistant knows about the store, topic by topic. Each topic has a human attention switch: the merchant decides which subjects, like returns, should always go to a person.

Settings, library tab: a table of topics like store information, shipping policies, payment and returns, each with status, a human attention switch and an edit button. — The library in settings. Each topic can be sent to a person instead of the assistant.

Merchants set up their assistant in their own words, and saw how it would answer before any customer did.

The same research also mapped what happens once the assistant is running: knowing which conversations need a person, and fixing answers without calling sales. Those opportunities went to the team to prioritize for 2025.