Magalu's WhatsApp AI agent converts 3x its app. It's 0.2% of sales

Brazil's Magalu sells inside WhatsApp through an AI agent. Its Q2 2026 release: 18m chats, 3x app-search conversion, R$100m in sales since launch.

Magazine Luiza, the Brazilian retailer known as Magalu, says its AI shopping agent on WhatsApp converts three times better than traditional search in its own app. In its second-quarter 2026 earnings release, published 6 August, Magalu reports more than 18 million conversations with the agent since its launch in late October 2025, more than R$100 million in sales by July, and about 20% of buyers coming back to buy again.

That is the best-documented result in AI shopping we can find from a retailer that is not a technology platform. It is also, by our arithmetic, about a fifth of one per cent of Magalu's sales.

What Magalu actually built

The agent is called WhatsApp da Lu, after Lu, the virtual brand persona Magalu has used for years in its marketing. A customer messages the same WhatsApp number Magalu already uses for customer service, or taps a link from the website or an ad, and describes what they want in text, by voice or with a photo.

Behind the persona sit several AI agents, each with one job, such as comparing prices or making small talk, according to Brazil Journal's December 2025 report, which also says Magalu uses both paid closed models and open-source ones. Magalu's chief AI officer told Mercado&Consumo that Lu shows at most five products per search, picked by the customer's profile and location.

Checkout happens inside the chat. The customer's account is matched to their phone number, and the agent produces a Pix payment code or takes a card without sending the shopper to the app. In the six-week window Meta measured for its case study, 75% of payments used the in-chat Pix code. After the sale, the same thread carries order tracking.

The agent did not replace a channel. It sits on the WhatsApp number Magalu's customers were already messaging.

What do the 3x and the R$100 million actually measure?

The 3x is a conversion comparison against search inside the Magalu app, per the earnings release. Meta's case study states the same multiple against the app and the website combined, over 21 October to 3 December 2025, and marks it self-reported. Neither source says how conversion is defined, whether the same customers are compared, or whether the people who choose to chat were already more likely to buy.

That last point matters most. A shopper who opens a conversation and describes a product is further along than one browsing a search page. Some of the 3x is likely to be who turns up, not what the agent does. Magalu has not published a holdout test.

The R$100 million is cumulative sales since launch, reached in July. Magalu's total sales were R$29.7 billion in the first half of 2026 alone. By our arithmetic, nine months of sales at that pace is around R$43 billion, which puts the agent at roughly 0.2%. Spread across 18 million conversations, it is about R$5.50 of sales per conversation, a rough figure because the two totals are dated a few weeks apart.

The satisfaction figure has also moved with each telling: an NPS of 90 in the chief executive's December 2025 interview, 85 in Meta's case study, 84 in the August release.

What did it cost?

Magalu has not disclosed the cost. The chief executive told Brazil Journal in December 2025 that early data suggested better unit economics than app sales, but that this still depended on assumptions being proven, and Magalu has not published a margin since.

The running costs it has described are of two kinds. Model costs, charged per token for the paid models, and Meta's fees for business messages. Meta now charges per delivered message, at a rate set by message type and the recipient's country. From 1 October 2026 it also resumed charging for replies inside the 24-hour service window beyond 1,000 free messages a month per number. A chat-led store pays for the conversation in a way a website search box does not.

Does this apply to a 200-person retailer?

Yes, in the markets where customers already shop by message, and only if you measure it properly. WhatsApp is the default channel in Brazil, India, Mexico and much of Southeast Asia, and a mid-sized retailer there probably already answers product questions on WhatsApp by hand. What Magalu has shown is that a structured chat with five products, a payment code and order tracking can close the sale in the same thread.

What a mid-sized retailer will not have is Magalu's in-house AI team, a brand persona with millions of followers, or Pix built into the chat. Expect to buy this from a WhatsApp solution provider and to start with a narrower job, such as answering stock and price questions and sending a payment link.

What to do on Monday

Count what already arrives. For two weeks, log every inbound WhatsApp message and tag the ones asking about a product, a price or stock. That share is your ceiling.

Measure your current close rate on those chats when staff answer them. Without that figure, a vendor's 3x has nothing to be compared to.

Price the conversation. Ask your provider for Meta's per-message rate in your country, its own markup, and what the 1 October 2026 service-message change adds at your volume.

Pilot with a holdout. Send half of product enquiries to the agent and half to staff for a month, and compare orders, returns and complaints, not chat volume.

What it does not mean

It does not mean chat will replace your website. After nine months at the most advanced end of this channel, Magalu's agent accounts for a fraction of one per cent of its sales. It does not mean the 3x is caused by the AI either: Magalu compares chatters with app searchers, and has not said they are the same kind of shopper.