Women's fashion brand · Retail · Costa Rica
A WhatsApp agent that never promises a size that doesn't exist.
For a women's fashion brand, we built an agent that serves customers on WhatsApp across the full catalog. Every size, price and stock level is checked against the catalog before it answers, and anything it shouldn't handle goes to a person. We deployed it and the brand's sales team tested it.
- Sector
- Fashion retail
- Channel
- Catalog
- 1,026 products, 2,697 variants
- Status
- Deployed and tested by the sales team
How it works
From the customer's message to the answer, with a check in between.
01
The customer writes
A question about a size or a price, or a screenshot of something she saw on Instagram.
02
The agent searches the catalog
Across 2,697 variants. If a photo came in, it recognizes the item by its visual fingerprint.
03
Checked before sending
Size, price and stock are compared against the catalog. If a fact isn't there, it isn't stated.
04
Answers or escalates
It replies with the real data, or hands the case to the team with the full conversation.
Architecture diagram. The live demo happens on the call.
The starting point
In fashion, the question is almost always the same: "do you have my size?"
Over a thousand products, each in several sizes and colors, and the questions arrive on WhatsApp at any hour. Answering well means someone has to look it up in the system; answering fast tempts people to reply from memory.
A generic assistant makes that worse: it fills what it doesn't know with a guess, and a promised size that doesn't exist is a lost sale and an upset customer. What was needed was the opposite: speed without making anything up.
What we built
Four pieces, one rule: if it isn't in the catalog, it isn't said.
01
Answers from the catalog.
Every size, price and stock level comes from the brand's catalog and is checked before it's sent. If the fact isn't there, the agent doesn't state it: it asks or escalates.
02
Recognizes the item from a photo.
The customer sends a screenshot and the agent identifies the item by its visual fingerprint among the catalog's 1,026 photos. If two items look too alike, it asks instead of guessing.
03
Orders checked before confirming.
It builds the order with reference, size and quantity, and checks it against the catalog before confirming. Anything that doesn't add up isn't confirmed: it gets corrected or goes to a person.
04
A person, when it's needed.
Discounts, exchanges or anything outside the rules go to the team with the full conversation. From the panel, the team sees every conversation, pauses the agent and replies in person.
2,697
catalog variants every size and price is checked against.
1,026
catalog photos in the visual index that recognizes an item from a screenshot.
891
automated tests the agent passes before every change.
System figures, not sales figures: we don't publish what we haven't measured.