← Back to Case Studies
Design Marketplace · Collector Events Live since October 2026

AI Shopping Companion for Collectors

How we gave a design marketplace a shopping companion that knows the shopper's own orders, the piece on screen, and the whole business, so every visitor gets the answers a collector would get from a personal advisor.

5 days

First commit to production

12

Tools across the shopper, the page, and the catalog

Every page

Knows the piece on screen

CLIENT OVERVIEW

A curated marketplace for collectible design, art, and fashion from independent sellers, which also runs in-person collector shows in major cities. Its most valuable buyers are collectors and interior designers who expect a high-touch experience, and those relationships were handled one at a time by a small sales team. Thousands of pieces were live, many of them large, fragile, or shipped across borders. The team wanted every visitor to get the kind of answers a collector gets from a person, without adding headcount.


THE PROBLEM

What was broken

01

Collector service didn't scale past the sales team

The company's revenue plan leaned on its top collectors, and a small sales team served each one by hand. Every other visitor got a product page and a support inbox.

02

The best context lived in decks and press, not on the page

Event recaps, press coverage, and show details sat in decks and articles that shoppers never saw. Questions like "what's the next show" or "who is this seller" had no fast answer on the site.

03

Buyers wouldn't commit without a shipping price

Large, one-of-a-kind pieces ship from sellers and warehouses in different countries. Shoppers asked what shipping would cost before they'd buy, and the sales lead waited on a logistics partner for each custom quote. One quote came back higher than the chair itself.


THE GOALS

What success looks like

Give every visitor, signed in or not, a companion that knows the piece on screen and the whole catalog

Let signed-in shoppers ask about their own orders, tracking, and shipping address, and nobody else's

Answer only from the company's own listings, press, pages, and event decks, with private details kept out

Quote the shipping price checkout charges, and measure whether shoppers who chat go on to buy


OUR APPROACH & SOLUTION

Building a companion that knows the shopper, the page, and the business

We designed it as a companion, not a support bot: something that knows who is shopping, what they're looking at, and everything the company has published, and answers the way the sales team would. We started from the question the sales lead heard most, shipping cost, and built the companion as a side panel on the existing storefront so nothing about the shopping flow changed, demoed it on staging after one day, folded in the CEO's feedback over a weekend, and soft-launched it with no announcement ahead of a traffic spike. Web deploys meant we could keep tuning it daily after launch.

Phase 1

Shipping Quotes That Match Checkout

Built a quote engine that returns the exact price checkout charges for each piece, from its real origin, with carrier delivery times on top. Along the way we found and fixed a checkout bug that priced cross-border pieces at domestic rates.

Phase 2

A Companion That Knows the Shopper and the Page

Gave the AI 12 typed tools over listings, sellers, events, site pages, and the shopper's own orders, plus a snapshot of the page on screen. It suggests the next best question for each page and searches the full catalog by category, seller, and price.

Phase 3

Company Knowledge Base

Indexed every press article, the site's own pages, and event decks into a searchable knowledge base. The team uploads decks in the admin, the AI drafts reference notes with pricing and contacts removed, and nothing goes live until someone publishes it.

Phase 4

Production Hardening and Measurement

Added a one-command kill switch, per-shopper and global rate limits, private-page exclusions, link restrictions, and 90-day transcript retention. Every chat is tracked in product analytics and tied to later purchases, without sending message text.


WHAT WE BUILT

Features and capabilities

01

A Panel Beside Every Page

The companion sits in a narrow column on the right of every storefront page on desktop and opens from a pill at the bottom of the screen on phones. Any shopper can use it, signed in or not, and a shopper who collapses it keeps it closed for the rest of the visit.

02

Answers About the Piece on Screen

The panel sends a capped snapshot of the visible page, so a question about "this piece" or "this event" gets an answer from that page first. Each page also gets three quick questions written from its own content, cached for a day so a busy page costs one call.

03

Knows the Shopper's Own Orders

Signed-in shoppers can ask where an order is, what was in it, how it's tracking, or which address it ships to. Every account tool is locked to the signed-in shopper, so no one can see another customer's data. The companion can't change orders, issue refunds, or apply discounts, and it points those requests to the support team.

04

Checkout-Accurate Shipping Quotes

Shoppers get the exact shipping price checkout charges for the piece from where it ships, plus carrier delivery estimates. A US ZIP code is enough to price it, international quotes say that duties and taxes aren't included, and freight or white-glove pieces are routed to the team for a quote after purchase.

05

Search Across the Full Catalog

The companion searches every published listing, sold-out pieces included, by keyword, category, seller, price range, and availability. Results sort by relevance, newest, or price, 12 at a time.

06

Answers From the Company's Own Knowledge

Questions about shows, sellers, press, shipping and returns, or membership are answered from the company's press coverage, site pages, event list, seller profiles, and event decks, with no web search. Press and site pages are re-crawled every night, and invitation-only events can be mentioned but are never linked.

07

Deck Uploads With Human Review

Marketing and admins upload event decks in the admin. The AI reads each deck and writes reference notes with sponsorship pricing and contacts removed, the uploaded PDF is deleted from the AI provider after reading, and each note stays a draft until someone publishes it.

08

Guardrails and a Kill Switch

The companion stays off private pages, links only to the site and known press outlets, and treats page text as information, never as instructions. Rate limits cap each signed-in shopper at 30 messages an hour, each visitor at 20, and all signed-out visitors together at 1,500, and one command turns the companion off across every server.

09

Purchase-Linked Analytics

Opens, messages, replies, link clicks, and errors are tracked in product analytics, and a shopper's first message tags their later purchases. Message text is never sent, and stored transcripts are deleted after 90 days.

BUILT WITH Elixir Phoenix LiveView PostgreSQL Anthropic Claude EasyPost Mixpanel

KEY TAKEAWAYS & IMPACT

What changed for this business

5 days

From first commit to production

The companion went from first commit to live on the storefront in 5 days, 19 days after the team first named the shipping-quote bottleneck. It shipped as a soft launch so real shoppers could find it on their own.

1 price

Shipping quote and checkout always agree

The companion quotes from the same rate tables checkout uses, so a shopper never sees one number in chat and another in the cart. Building it also caught cross-border pieces that were being priced at domestic rates.

Day 1

Measurement built in from launch

Every conversation has been tied to later purchases in product analytics since the first day, and message text is never sent to analytics. The team can see whether shoppers who chat go on to buy, not just whether they chat.

THE OUTCOME

The wins

  • Every visitor gets answers about the piece, the seller, or the next show without leaving the page they're on, whether or not they're signed in.

  • Signed-in shoppers can check on their own orders and tracking in the same conversation, without opening a support ticket.

  • Show details and press coverage that sat in decks and articles reach shoppers in plain answers, with pricing and contacts stripped from decks before anything goes live.

  • Shoppers can get the shipping price for a carrier-shipped piece on the product page, before they reach checkout or have to ask the sales team.

  • Cross-border pieces are charged international shipping rates at checkout, after the build exposed that they were being priced at domestic rates.

  • When the team said the panel took too much of the screen in the first week, the fix shipped within 3 days.

  • Cost and abuse are capped from launch by hourly rate limits and a one-command kill switch, and the team can judge the companion on whether shoppers who chat go on to buy.

Want a shopping companion like this on your storefront?

Let's talk about what an AI shopping companion that knows your customers, your catalog, and your business could look like on your site.

Schedule a Call →

Get dev leadership insights

Tips on optimizing your dev team, shipping faster, and building products that scale.

TenEx
Privacy Terms SMS Program © 10ex.dev 2026