The catalogue writes itself, almost
We build AI listing writers for auction houses and dealers. Photographs go in, catalogue descriptions come out, and a person reads only the lots the writer is unsure of. For Frank Zoller Authentic History, a US militaria auction house, the writer fills 1,250 lots a sale from a 32,000-item reference corpus and has saved 400+ hours. Its first live auction ran on 16 August 2026.
In detail
How does an AI listing writer turn a photograph into a description?
It starts from what the house has already written. Every past listing, with its title, description, estimate and realised price, becomes the reference set, so the writer learns the house's own vocabulary rather than a generic one. A new photograph is matched against that set: a piece the house has sold before pulls its own record, a new piece is identified from the image and described by analogy with its nearest relatives. Each description leaves with a confidence score. High scores pass straight to the upload sheet. Low scores wait for a person. The writer never publishes anything; it drafts, and the house decides.
The build for Frank Zoller Authentic History
Frank Zoller Authentic History sells militaria at auction in the United States. Before the writer, each sale began with 1,250 lots typed by hand, one title, one description, one start price and one estimate at a time, and nothing could go live until the last row was done.
We loaded 32,000 past items into a hosted backend as the reference corpus. Then we built the writer around the house's existing routine: photograph, edit, number, drop the folder in. The writer returns a filled upload template. A cataloguer reads the flagged rows, corrects what needs correcting, and the correction goes back into the corpus so the next sale starts smarter. The first live auction on the platform ran on 16 August 2026, and the client's own count puts the time saved at 400+ hours.
The corpus did one more thing nobody had asked for. With every past sale in one table, realised prices sat visibly below estimates across the set. A tool built to save typing showed the house something about its pricing that no single catalogue could have.
How a writer is built for a dealer
- 01
Gather what you have written
Past listings, past catalogues, the shop's product pages, even the notebook. That is the training material, and most dealers have years of it without knowing it is an asset.
- 02
Match the format you upload
An auction house fills a bidding template; a shop fills a product page. The writer is built to produce the exact sheet or page you already use, so nothing changes downstream.
- 03
Set the confidence line
Together we decide how sure the writer must be before a row passes unread. Early on the line sits high and a person reads most rows. It drops as the corrections accumulate.
- 04
Run it on a real sale
The first live use is a real catalogue, not a demo set. We sit beside it, and the rows a person had to fix tell us what to improve next.
- 05
Hand it over
The code, the hosting account and the corpus are in your name. You can run the next sale without us, and we would rather you could.
Does the writer work for a shop with no auction calendar?
Yes. Oorlogsspullen, a militaria shop in the Netherlands, has no sale dates and no bidding template; it has stock that needs words. Its writer drafts product listings in the shop's own register from the same kind of reference set. The engine is the same; the output is a product page rather than a catalogue row. We publish no figure for that build, because none has been measured to a standard we would put our name to, and a number without a measurement behind it is the kind of claim this page exists to avoid.
What it costs to try
A listing writer is priced under Tailored, to the number it is meant to produce: the hours it removes from each sale, multiplied by what those hours cost you, set against the build. We write that arithmetic down before anything is built and show it to you. If it does not clear inside a year, we tell you to keep typing, or point you to an off-the-shelf tool that does enough.
The first sale on the writer is the test. You bring a real catalogue; we bring the writer and sit beside it. If the flagged rows outnumber the clean ones by the end, the build has not earned its fee and we say so.
Who owns the output
- The descriptions are yours. They were written from your own past listings, in your voice, for your catalogue, and no licence from us sits on top of them.
- The corpus is yours. It is your sales history, hosted in an account in your name, and it leaves with you.
- The code is yours. The repository and the hosting are in your name from the first week, not transferred at the end.
- The corrections are yours. Every fix a cataloguer makes goes into your corpus, not into a model we then sell to the next dealer.
- The model account is yours too. We say which model we chose and why, and you can change it.
The questions that follow.
How accurate are the descriptions?
Accurate enough that a person reads a minority of rows, and the writer says which. The confidence score is the honest answer: it is wrong sometimes, it knows when it might be, and it never passes a doubtful row unread.
Will it write in our house style?
Yes, because it learns from your past listings and nothing else. A house that writes terse rows gets terse rows; one that writes a paragraph of provenance gets that.
What do we need to start?
Your past listings in any exportable form, a folder of photographs from the next sale, and the template or page you upload to. A few hours of a cataloguer's time in the first fortnight.
Can it work with our bidding platform?
It fills the upload template your platform already accepts, so nothing on the platform side changes. If your platform has no template, it fills your shop's product page instead.
What does it cost?
It is priced under Tailored, to the hours it removes. We write the arithmetic down first and show it to you. If the build cannot pay for itself inside a year, we say so before you spend.
Do we need our own developer afterwards?
No. The writer is documented so a developer could continue it, but running a sale needs a cataloguer, not an engineer. We stay on if you want us; you do not need us.
The price is published.
Where this leads.
No deck. No discovery theatre. One metric, one date, in the contract.
