BRENNEN

Industries / Retail, hospitality and personal services / Retail and e-commerce / Auction houses

Valuation and cataloguing

Valuation and cataloguing is working out what a lot is and what it may fetch, and writing it up in the catalogue: the lot number, title, maker, date, material, size, condition, ownership history and a pre-sale estimate, given as a low-to-high price range.

4

Milestones achieved, from 1599 to 2018.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1599

    Earliest surviving auction catalogue

    The bookseller Christophorus Guyot issued a printed catalogue for the sale of the library of Filips van Marnix, lord of Sint-Aldegonde, held in Leiden on 6 July 1599. The History of Information timeline, citing Breslauer and Folter's bibliography, calls it the earliest surviving catalogue of a book auction; it says the first book auction with a printed catalogue took place in Leiden in 1593, but that catalogue is lost, and that only two copies of the 1599 catalogue survive. The Dutch national library (KB) now holds about 50,000 auction and booksellers' catalogues, only some of them described in its catalogue.

    historyofinformation.comkb.nl

    Credit

    Christophorus Guyot, Leiden

  2. 1986

    Coins graded by outsiders

    The Professional Coin Grading Service (PCGS) opened on 3 February 1986 and graded coins to one standard, set by a party other than the buyer or seller. Its co-founder David Hall said the company "created sight-seen and sight-unseen markets for trading certified coins", so that graded coins could change hands without being handled. By 2011 PCGS had graded more than 20 million coins worth over $20 billion, by its own count.

    coinweek.com

    Credit

    David Hall and the other founders of PCGS

  3. 2013

    Nazi-era catalogues made searchable

    The Getty Research Institute, the Kunstbibliothek in Berlin and Heidelberg University Library added more than 2,000 sale catalogues published in Germany, Austria, Switzerland and occupied territories between 1930 and 1945 to the Getty Provenance Index, a free database of who owned which artworks. Heidelberg scanned about 150,000 pages, and Getty editors spent two years correcting the machine-read text and adding sale prices from art journals, taking the index past 1.5 million records. The project later covered sales from 1900 to 1945, and each record links to a scan of its catalogue.

    getty.edugetty.edu

    Credit

    Getty Research Institute; Kunstbibliothek, Staatliche Museen zu Berlin; Universitätsbibliothek Heidelberg

  4. 2018

    An auction house buys AI

    Sotheby's bought the Mei Moses Art Indices, a price database that compares art with other investments, in 2016 and opened it only to its own specialists and clients. In January 2018 it bought Thread Genius, a start-up of two engineers whose software identified objects in images and recommended similar ones, to build a recommendation engine using the Mei Moses database of 50,000 images.

    LessonA price index that others had used became a tool for one house's clients.

    techcrunch.comtheartnewspaper.com

    Credit

    Sotheby's; Andrew Shum and Ahmad Qamar of Thread Genius

Watch for

  1. AI-written lot descriptions

    Gavelist, which sells AI cataloguing for $0.15 a lot, says that 96.3% of the 58,134 lots its 41 client accounts exported between 20 March and 22 September 2026 went out with no edit to the AI-written title or description; one client supplied 70% of those lots, and across the rest the figure is 88.2%. AuctionWriter, AScribe and Circuit Auction sell similar tools that write titles, descriptions and, in some cases, estimates from photos, and none of their pages read for this file gives an error rate. A description sent out unedited is not the same as a correct one.

    Achieved whenAn outside party checks a sample of AI-written entries against the objects and publishes the error rate.

    gavelist.comauctionwriter.comcircuitauction.com

    Claimed by

    Gavelist, September 2026

  2. Lots read from old catalogues

    Mathias Zinnen and five co-authors tested vision-language models, AI that reads images and text together, on pages of the German Sales catalogues, and report that they can turn printed lot entries into structured records; commercial services did best, and models run on an institution's own computers worked if their output format was enforced. The authors say "varying degrees of human-in-the-loop correction are still necessary" and released their code and test data.

    Achieved whenA library or archive uses the method to publish lot records for a whole catalogue series with a measured error rate.

    arxiv.org

    Claimed by

    Mathias Zinnen and co-authors, arXiv preprint, August 2026, accepted at the VISART workshop of ECCV 2026

Open

  1. 01

    Audit AI catalogue entries

    A blind test of AI cataloguing tools on 200 objects whose facts are already settled: each tool writes an entry from the photos, and two checkers count factual errors in maker, date, material, size and stated condition.

    The brief
    Credit · unclaimed
  2. 02

    Read handwritten sale prices

    A tool that reads the prices and buyers' names written by hand in the margins of digitised sale catalogues, links each note to its lot, and checks the prices against those already in the Getty Provenance Index.

    The brief
    Credit · unclaimed

Open 01 · the brief

Audit AI catalogue entries

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A blind test of AI cataloguing tools on 200 objects whose facts are already settled: each tool writes an entry from the photos, and two checkers count factual errors in maker, date, material, size and stated condition.

Why it is still open

Vendors report speed and edit rates, not accuracy. Gavelist's published figure is the share of descriptions sent out unedited, and the AuctionWriter, AScribe and Circuit Auction pages give no error figures.

gavelist.com

Built with what exists

Objects with photos and checked records from museum open-access collections; the commercial tools themselves (Gavelist charges $0.15 a lot); Claude Fable 5.2 as a baseline (see the Note "Three working devices"), with each tool run several times on the same lot to show how much its answers vary (see the Note "All twenty times"); a scoring sheet with two checkers per entry.

  1. First step

    Pick 50 objects from the Metropolitan Museum's open-access collection across furniture, ceramics, silver and prints, and write down the facts a correct entry for each must contain.

  2. Done looks like

    Errors per 100 fields for at least three tools on the same 200 objects, the two checkers agreeing on at least 90% of fields, and the photos, entries and scores published.

Open 02 · the brief

Read handwritten sale prices

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A tool that reads the prices and buyers' names written by hand in the margins of digitised sale catalogues, links each note to its lot, and checks the prices against those already in the Getty Provenance Index.

Why it is still open

The Getty says many of the German Sales catalogues carry "handwritten annotations about sale prices and buyers' names", but its editors entered sale prices from art journals such as Weltkunst. The Lot Machine paper describes extracting printed lot entries and does not report reading handwritten notes.

getty.eduarxiv.org

Built with what exists

Heidelberg University Library's scans of the German Sales catalogues, which the Getty credits under a CC BY-NC-SA 3.0 DE licence; Claude Fable 5.2 (see the Note "Three working devices") or a small open model fine-tuned on a few hundred hand-labelled notes (see the Note "One GPU to fine-tune"); the code and test data released with the Lot Machine paper.

  1. First step

    Find 10 annotated catalogues in the Heidelberg collection and transcribe one catalogue's notes by hand to serve as a test set.

  2. Done looks like

    Notes read from 50 annotated catalogues, with prices matching the Getty's journal-derived prices for at least 90% of the lots where both exist, buyers' names checked by hand on a 10% sample, and the records published under the same licence.