Listings written from photographs
The slowest part of an auction is the typing. We built a listing platform for Frank Zoller Authentic History, a US militaria auction house, that reads photographs and fills the catalogue template: title, description, start price and estimate for 1,250 lots a sale, trained on 32,000 past items. It has saved the house 400+ hours. Around it we run the website, the retail shop and the catalogue email.
In detail
How does an AI write an auction listing from a photograph?
It starts with the house's own history. We loaded 32,000 past items, with their titles, descriptions, estimates and realised prices, as the reference set. A new photograph goes in; the system identifies the item, finds its nearest relatives in that history, and writes the row the catalogue needs in the house's own voice. Repeat items pull their own record. Each row carries a confidence score, so a cataloguer reviews only the lots the system is unsure about and the rest pass straight through. The filled template uploads to the bidding platforms the house already uses.
From photographs to a live sale
- 01
Photograph as you already do
No new camera, no new process on the floor. The lots are shot, edited and numbered the way the house has always done it.
- 02
Drop the folder in
The platform reads every image and writes a complete listing row for each: title, description, start price and estimate range, in the house's style.
- 03
Review the flagged lots only
Uncertain items are marked. A cataloguer checks those, corrects where needed, and the correction feeds back into the reference set.
- 04
Upload the template
The finished sheet goes straight to the house's bidding platform, and the existing automation cross-posts it. The first live sale on the platform ran on 16 August 2026.
The website, the shop and the catalogue email
A listing platform fixes the slowest week in the cycle. It does not bring a new bidder to the sale. So the second half of the work is the house's own front door: a website rebuilt in an archival, institutional register, with the accession card as its signature, so a collector arriving from a search sees a house that keeps records rather than a storefront with a countdown.
Beside it sit the retail shop, for the pieces that sell better at a fixed price than under the hammer, and the catalogue email, sent to the house's own list before each sale, which is the cheapest bidder the house will ever reach. Each of the three is measured on its own number: visits, shop orders, and the bids that trace back to the email.
A custom tool, or an off-the-shelf one?
| Off the shelf | Built for the house | |
|---|---|---|
| Knows your past sales | No. It writes from a general model and sounds like everyone. | Yes. It is trained on your own records and sounds like you. |
| Knows your template | You reformat the output every time. | It fills the bidding platform's sheet directly. |
| Confidence | Every lot needs a human read. | Only the flagged lots do. |
| Price | A subscription that grows with every user. | Scoped against the hours it saves, then yours. |
Can this work for a dealer or a shop rather than a house?
Yes, and it already does. Oorlogsspullen, a militaria shop in the Netherlands, uses an AI writer we built that drafts its product listings. A shop has no catalogue deadline, but it has the same bottleneck: a back room full of stock that cannot be sold until it has been described. The writer drafts, the dealer corrects, and the corrections make the next draft closer. We publish no figure for it because none has been measured and cleared, which is the same rule we apply to every name on this site.
What we found in the data, and why it matters
- With 32,000 records in one place, realised prices sat consistently below estimates across the history. Systematic estimate inflation, invisible one sale at a time.
- A house that knows that can set estimates that hold, which bidders learn to trust.
- Software without a price is a hobby with a repo. Every build we scope is priced against the hours or the sales it changes, before a line is written.
- The platform is the house's. It is hosted for them, and they keep it if they leave us.
The questions that follow.
Does the platform replace the cataloguer?
No. It replaces the typing. The cataloguer's knowledge goes into reviewing the flagged lots and correcting the drafts, which is the part only a person can do.
What does a build like this cost?
It is priced to the number it changes. We scope it against the hours per sale and the sales per year, and we put the price in writing before anything is built. There is a journal piece on what a custom AI app costs.
Do you only work with militaria?
No. The method is the same for coins, watches, wine or furniture: a back catalogue to learn from, a template to fill, and a review step for the lots that need one.
Can you run our ads as well?
Yes, through the Tailored tier, where paid media sits beside the build. For most houses the catalogue email and the search work come first, because they are cheaper per bidder.
Who owns the code and the data?
You do. The reference set is your history, the platform is hosted for you, and the handover is documented so another developer could pick it up.
The price is published.
Where this leads.
No deck. No discovery theatre. One metric, one date, in the contract.