AI agent case study
Building Tagmade: an AI listing agent that lives in your shop’s phone.
How Changeable built an AI agent for WooCommerce that takes a phone photo of a product and its price tag, cleans the image, reads the tag, writes the listing and hands back a finished draft, starting as a staff tool for The Collective in Hāwera and becoming a product any store can install.
Project overview
Staff take two photos. The agent does the listing.
The Collective is a craft consignment shop in Hāwera, selling the work of many local makers. Every item that comes through the door needs to go online, and every online listing used to mean someone at a desk: uploading photos, fixing backgrounds, typing out the price, writing a title and description, and remembering which maker it belonged to. Small, one-off handmade items made that work slow and easy to put off.
Changeable built an AI agent that does that job inside a phone app. A staff member photographs the product and its price tag. The agent cleans up the photo, reads the tag, works out the price, writes the title, short description and full description, and returns a draft listing for a person to check. One tap saves it as a draft or publishes it to the shop. Once it was working in a real shop, the same agent was rebuilt as Tagmade, a WooCommerce plugin any store can install.
Two photos in, a listing out
A product photo and a price-tag photo are all the agent needs to produce a complete draft listing.
Built in a real shop first
The agent was proven on The Collective’s live store before any of it was turned into a product.
People stay in charge
The agent drafts; staff check. Only people with publishing rights can put a listing live.
Works for any store
Currency, listing language, maker labels and category paths all follow the store’s own settings.
The problem
The shop was full. The website was behind.
A consignment shop lives on variety: dozens of makers, hundreds of one-off items, each with its own handwritten tag. In store, that variety is the appeal. Online, it is a data entry job. Every item needed a clean photo, a correct price, a decent description and the right maker attached, and the person best placed to do it was usually busy serving customers.
Existing AI tools for WooCommerce sit inside the product editor on a desktop and help rewrite text or edit images. Nothing started where the work actually starts: a staff member holding the item, on the shop floor, with a phone. That was a process problem before it was a technology one.
What the listing job involved before
- Photographing items and cleaning up backgrounds by hand
- Reading handwritten tags and typing prices into the shop
- Writing a title, short description and full description for each item
- Assigning the right maker so sales could be reconciled
- Entering packed size and weight so shipping could be calculated
- Doing all of it at a desk, away from the shop floor
How the agent works
One agent, six steps, from the shop floor to a finished draft
Staff do the two things only a person can do: hold the item and check the result. Everything in between is handled by the agent inside the app.
Capture
StaffA staff member opens the app on their phone, photographs the product, then photographs its price tag. Step dots and on-screen tips guide the shot, so anyone on the team can do it without training.
Clean the photo
AgentThe agent sends the product photo to a background removal service and gets back a clean, shop-ready image. Every upload is checked to make sure it really is an image before anything else happens.
Read the tag
AgentUsing Anthropic’s Claude, the agent reads the price tag, extracts the price and any product details, and converts the price into the store’s own currency format. If a price is ambiguous, it refuses to guess and asks staff to type it.
Write the listing
AgentThe agent writes a product title, a one or two sentence short description and a full description, in the store’s chosen listing language, then suggests the maker and category. Staff can pick an existing maker or add a new one on the spot.
Check its own work
AgentBefore handing back, the agent checks that the photo and the tag actually match. If they don’t, it raises a warning for staff only and keeps that note out of anything a customer will see, with a server-side check as a backstop.
Hand back for review
StaffThe finished draft appears on a check screen. Staff confirm the details, add packed size and weight for shipping, then save it as a draft or, if they have publishing rights, publish it straight to the shop. Drafts from listers wait in a review queue.
How it was built
From one shop’s staff tool to a product any store can install
The agent was built for a real business first, then deliberately rebuilt for every business. Each stage was tested before the next began.
Get the shop’s foundations right before adding AI
Before any agent was built, The Collective’s online operation was rebuilt: a new WooCommerce site, a simple maker taxonomy instead of a complex multi-vendor plugin, standalone in-store EFTPOS, and sales flowing to Xero so commission is managed where the bookkeeping already happens. Getting that structure right meant the agent would be writing into a shop that was organised to receive its work.
A staff app on the shop’s own website
The agent was built as part of a staff-only phone app that lives on the shop’s own site and opens with the normal shop login. It was developed in stages with automated tests at every step, proven on a staging copy of the store, then installed on the live site. Real feedback shaped it: packed size and weight became required fields because shipping is calculated from them, a separate AI-written short description was added, and a virtual option was added for workshops and event tickets.
Turn one shop’s settings into choices
To become a product, everything that was specific to The Collective had to become a setting. The agent learned to follow any store’s currency and number format, write in the store’s chosen language, use the store’s own word for a maker (brand, vendor, supplier, artist or a custom term), work with nested categories, and make packed size required, optional or switched off. A two-layer permission model and a review queue were added so larger teams could let staff list without letting them publish.
Setup wizard, theming and a redesigned app
A six-step setup wizard connects the AI and background removal services, tests the connection, shows a QR code to open the app on a phone, and gives a cost estimate before the first listing. Stores can match the app to their brand with their own colours, logo, font and corner style. The app itself was redesigned with a home screen showing photos, prices and status, a To review filter, and a working screen that shows the agent’s real progress.
Licensing, quality checks and a public demo
Licensing and automatic updates were added, and the code was taken through WordPress coding standards, static analysis and plugin checks, along with a security pass. A public demo store was set up with its own spending limits and a nightly reset, so anyone can try the agent on a real WooCommerce shop. The product, its website, documentation and checkout went live in October 2026, built to WooCommerce Marketplace standards.
Guardrails
An agent that does the work, and knows where to stop.
An AI agent writing into a live shop needs the same kind of controls Changeable designs into client AI governance work: clear permissions, human review, cost limits and careful handling of data.
Results
What the build delivered
A working agent in a real shop, and a tested, supported product built from it.
All staff need to take
The agent turns a product photo and a tag photo into a complete draft, including price, title, descriptions and maker.
At The Collective, Hāwera
The original staff app runs on The Collective’s live WooCommerce site, alongside a custom gift voucher system built in the same app.
Automated checks on every build
Every stage of the product was built with automated tests, and every release passes them before it ships.
Coding standards and security
WordPress coding standards, static analysis and plugin checks all pass, with a dedicated security review of uploads and outbound requests.
Currency, language and labels
The agent adapts to each store’s currency format, listing language, maker wording and category structure.
Public demo store
A public WooCommerce demo with its own spending caps and a nightly reset lets anyone watch the agent work.
Questions
Questions about the Tagmade AI agent?
Common questions about how the agent works, how it was built, and how the same approach applies to your business.
What does the Tagmade AI agent actually do?
It takes two phone photos, one of the product and one of its price tag, and produces a finished WooCommerce draft listing. It cleans the product photo, reads the price from the tag, writes the title and descriptions, suggests the maker and category, checks the photo and tag match, then hands the draft back for a person to check and save or publish.
Why start with a single shop instead of building a product straight away?
Building for The Collective first meant the agent was shaped by real items, real tags and real staff on a real shop floor. Only once it worked there was it rebuilt for other stores, with everything shop-specific turned into a setting.
Can the agent publish listings without anyone checking?
Only if the store lets it. Staff in the Lister role can only save drafts, and the server refuses any publish request from them. Drafts wait in a review queue for a Publisher, who can check and publish with one tap.
What happens if the price on the tag is hard to read?
The agent does not guess. If a price is ambiguous, it refuses to fill it in and asks the staff member to type the correct figure, so a misread tag never becomes a wrong price in the shop.
What if the photo and the tag don’t match?
The agent checks for this and raises a warning that only staff can see. That warning is kept out of the title and descriptions, and a server-side check removes any such note that slips through, so customers never see it.
Which AI does Tagmade use, and who pays for it?
Tagmade uses Anthropic’s Claude to read tags and write listings, and a separate background removal service for photos. Each store connects its own accounts, so data and costs stay under the store’s control. The app shows an estimated cost per product and a running monthly total.
Does it work outside New Zealand?
Yes. The agent follows the store’s own currency and number format, writes in the store’s chosen listing language, and uses whatever word the store prefers for a maker, such as brand, vendor, supplier or artist.
Can Changeable build an AI agent like this for our business?
Yes. The same pattern, where staff capture something in the field and an agent turns it into structured records in your existing system, applies to stock intake, asset registers, inspections and job notes. We start by understanding the work, then design the agent around it. See our AI agents service or book a discovery session.
What job could an AI agent take off your team?
Try Tagmade on the live demo store, or talk to Changeable about an agent built around your own work and systems.