Leo Abreu / Selected Work
Magalu · AI Commerce
Conversational UX Designer / Magalu · Lu
From conversationto commerce.
I designed how Lu helps shoppers find products in WhatsApp and keeps the purchase moving when they leave the chat to view a product or check out.
Conversational UX Designer
DEC 2025 – FEB 2026
Product outcomes / Reported by Bain & Company
3×
Conversion versus the Magalu app
4.5M
Unique visitors in the first six months
85
Net Promoter Score (NPS)
02 / Project at a glance
I owned the conversation flows.
People
Magalu customers shopping or looking for help in a channel they already used for order tracking.
My scope
I designed how Lu recognized intent, decided when to ask or act, and kept context between messages, product details, account and cart.
Collaboration
I worked with product and engineering on blueprints, alternate paths and the rules behind them. The wider team built and launched the shopping experience.
03 / The challenge
Show value before asking for effort.
Lu was already a familiar face for a Brazilian retailer that spans stores, an app, a site and a marketplace. On WhatsApp, she helped with orders. Shopping meant adding discovery and purchase without getting in the way of that existing task.
In early flows, shoppers could encounter an explanation, a request for details or account verification before seeing anything useful. Moving those requests later made the start lighter, but pushed more responsibility into the handoffs between chat and commerce.
In conversation
Recognize intent → Refine → Show products
In dedicated views
Product details → Account and cart → Checkout
Back in conversation
Keep comparing, resume a cart or get order help
Journey summary / Shoppers could enter, leave and resume at different points.
04 / Key design decisions
Four decisions behind Lu’s next move.
01 / Entry and onboarding
A campaign is context, not intent.
A coupon invite told Lu why someone might be there, not what they wanted. In the Sandbox I explored an offer-led opening and a neutral welcome: pressing “claim the coupon” called for a different response than typing “Hi.”
The trade-off was between explaining everything upfront and getting to the task sooner. We designed a short opening for ambiguous messages and let a clear shopping or tracking request skip the generic introduction.
Greeting only
“Hi, Lu”
Give a short welcome and ask what the person needs. Don’t apply an offer by default.
Shopping intent
“I’m looking for a TV”
Acknowledge the request and enter discovery, without repeating onboarding.
Order support
“Where’s my order?”
Route to tracking rather than pushing a purchase.
UX Blueprint / First session without a CRM invite. Examples translated and condensed.

Campaign prototype / “Claim the coupon” signals intent to use the offer.

Campaign prototype / “Explain how it works” asks for guidance instead.
02 / Conversational discovery
Show options early. Ask first only if the answer changes the result.
Instead of making everyone complete the same questionnaire, I designed two paths. If the catalog could already produce useful options, show them and use the conversation to refine. If a missing detail would change the product or variant, ask one question before proceeding.
In the Blueprint, storage can matter before selecting an iPhone variant. In the Sandbox notebook exploration, Lu shows an initial carousel, then asks whether the laptop is for work, study or gaming. The examples demonstrate different timings, not an A/B test.
When information is sufficient
Show → ask → refine
Give shoppers options to react to; use one follow-up at a time.
When a detail blocks the next step
Ask → show the right options
Clarify the one attribute that materially changes what Lu can recommend or add.
Proposed guardrail: 0–1 questions for simple choices, up to 2 for medium complexity and up to 3 for complex choices. These were ceilings, not targets. Stop once the shopper can make a useful choice.

01 / Intent
The shopper asks for a TV.

02 / Blocking detail
Lu asks where it will go.

03 / Discover
Lu moves to product options.
TV prototype / The screenshots show the interaction sequence, not recommendation accuracy.
03 / Location, identity and checkout
Ask for information where it earns its place.
Location and identity solved different problems. A postal code improved shipping and availability estimates, but wasn’t needed to understand a first product request. An account mattered when a purchase action needed it.
That changed the design work at checkout. In one Sandbox exploration, an account was found but the phone number did not match. Lu explained what happened and offered to retry with another identifier or check the account details, rather than ending the conversation.
Some coupon entries explored asking for identity earlier. That was a separate branch, not a rule that applied to every session.
Discover
Let a shopper explore while signed out when the flow allows it.
Estimate
Ask for the postal code when local shipping or availability matters.
Commit
Verify identity at the action that needs it and resume the shopper’s task.
Design rules / Synthesis of the chat notes, authentication blueprint and campaign explorations.
04 / Product details and cart recovery
Carry the shopping context across screens and sessions.
The product view could show specifications, delivery and seller details without turning the chat into a long message. On return, Lu needed to know what the person had just seen so she could answer a question, compare options or continue toward purchase.
The cart blueprint turns the same idea into a concrete handoff. If someone comes back with an unfinished cart, Lu acknowledges it and offers to finish, keep shopping or clear it. We also mapped what to do when price, stock or shipping changed instead of promising that the cart state was always current.
State
Shopper returns with items in their cart
Lu’s response
Name the unfinished cart and offer a choice
Next action
Finish checkout, continue searching, or clear and confirm
UX Blueprint / WhatsApp cart abandonment. Simplified from the original flow.

UX Blueprint / WebView exploration. Use the dedicated view for detail; keep questions and comparisons in chat.

UX Blueprint / Cart recovery. The suggested actions reflect the shopper’s unfinished cart.
IMPLEMENTATION HANDOFF
What Lu needed to know before making the next move.
My work with developers turned a journey map into response and transition rules. The important unit was not a polished message in isolation; it was a known situation, what Lu could carry forward and what action should follow. These examples synthesize the design rationale; they are not a reproduction of production logic.
SITUATION
CONTEXT
DESIGN RESPONSE
An offer opens the chat
The campaign identifies the offer. It does not prove the shopper wants to use it.
Acknowledge the offer without treating it as intent; let the shopper claim it, ask how buying works or describe another need.
A broad product request arrives
Category and any constraints are already known. A missing detail only matters if it changes the result.
Show options when they are useful. Ask one short question first when the answer would materially change the recommendation.
The product page closes
A product detail page was open, but closing it is not a purchase decision.
Keep the viewed product in context. Make it possible to ask about it, compare alternatives or continue shopping without restating the request.
The shopper returns to a cart
An existing cart gives a useful starting point, but its contents or availability may have changed.
Offer actions that match the cart state. If the terms changed, make that visible before asking the shopper to commit.
This is why I separated discovery from transactional requirements. A ZIP code can improve delivery and availability information, but a first search should not have to stop for it. Identity belongs at the action that needs an account; a coupon-specific exploration can introduce it earlier. That exception matters when briefing implementation, because “always ask at checkout” would erase a real branch in the design.
05 / Outcomes & business impact
From blueprints to a shipped shopping experience.
The team launched Lu’s WhatsApp shopping experience and measured and tested it with thousands of users. My contribution was to make the behavior explicit for engineering: what Lu knows, what she needs to ask, and what action can follow.
Bain’s public case study reports the conversion, reach and NPS figures at the top. They describe the whole product, not the lift from an individual flow or decision.
06 / What I learned
Design the next action, not just the next message.
Asking less only works if the system knows when it can take a useful action and how to recover when it can’t. Next time, I’d track turns to the first useful result, unnecessary follow-ups and successful cart resumptions separately. NPS tells me how the product was received; those signals would tell me where the conversation needs work.