Industries / Retail, hospitality and personal services / Retail and e-commerce / Auction houses
The sale
The sale is the auction itself: the auctioneer takes bids from the room, the telephones and the internet, in steps set by the house, until one bidder is left and the hammer falls.
6
Milestones achieved, from 1766 to 2020.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1766
Christie's first sale
James Christie, born in Perth in 1730, worked for a London auctioneer before setting up on his own in Pall Mall, and held his first auction on 5 December 1766, by the company's records, selling "diverse property" from houses whose owners had died or moved. His first sale to include pictures followed four months later.
Credit
James Christie
1958
The evening sale begins
On 15 October 1958 Peter Wilson, Sotheby's chairman, sold seven Impressionist and Post-Impressionist paintings from the collection of the New York banker Jakob Goldschmidt at a black-tie evening sale in London, with television crews and film stars in the room. The auction record for a work of art was broken twice, and Sotheby's says the sale took 21 minutes. Sotheby's calls it the first black-tie gala auction and credits it with the evening sale, still the main event of the art market's year.
Credit
Peter Wilson, Sotheby's
1975
Buyers start paying fees
In 1975 Christie's and Sotheby's in London began charging buyers a premium of 10% on top of the hammer price, and cut their commission to dealers selling through them from 10% to 6%, by the Spectator's account in 1981. Dealers' associations alleged collusion, and in 1979 the UK's Director General of Fair Trading used his powers under the Restrictive Trade Practices Act 1976 to ask both houses whether they had an agreement. The buyer's premium is now a major source of the houses' revenue: from February 2026 Sotheby's charges 28% on hammer prices up to $2 million in New York.
Credit
Christie's; Sotheby's
1993
Property auctions run poorly
The Resolution Trust Corporation, the US government agency that sold the assets of failed savings banks, used auctions to sell real estate. In July 1993 the US General Accounting Office reported that its 1992 Washington and Baltimore auctions were "planned and managed poorly", and asked the agency to see that "book value reduction goals do not cause staff to disregard established policies and procedures".
LessonTargets for how much property to clear can lead staff to skip the agency's own procedures.
Credit
US General Accounting Office
2001
Commissions fixed in secret
In May 2001 the US Department of Justice charged A. Alfred Taubman, former chairman of Sotheby's, and Anthony Tennant, former chairman of Christie's, with fixing the commissions charged to sellers from 1993 to 1999; during that time the two houses charged sellers in the US at least $400 million in commissions. Christie's had cooperated under the department's leniency programme, and Sotheby's had pleaded guilty in October 2000 and been fined $45 million.
LessonTwo houses that reportedly held more than 90% of the world's live art auctions could stop competing on price.
Credit
US Department of Justice, Antitrust Division
2020
Bids from screens worldwide
On 29 June 2020, with salerooms shut by the pandemic, Oliver Barker stood alone in a curved booth in London and took bids from screens showing Sotheby's telephone specialists in New York, Hong Kong and London, and from online bidders, in three linked sales streamed live. They made $363.2 million with fees, and a Francis Bacon triptych sold for $84.5 million. The Art Newspaper called it the first major evening sale held entirely online.
Credit
Sotheby's; Oliver Barker
Watch for
Bids read from sale videos
Two economists built a record of the bids on more than 3,500 lots sold by Christie's and Sotheby's from 2020 to 2024 by reading the on-screen bid display in the houses' YouTube livestreams with the Tesseract text-recognition program, and by transcribing the auctioneers' speech. They write that the houses publish prices but not "historical bid sequences, bidder identities, and the number of active bidders", and that they could set only bounds on how many people bid.
Achieved whenThe paper passes peer review, or another team rebuilds the bid records from the same videos and gets the same sequences.
Claimed by
Kirill Ponomarev (University of Chicago) and Bruce Wen (Yale University), working paper presented at cemmap in London, November 2025
Open
01
Count who actually bid
For each lot in a public sale video, a count of the separate bidders by channel (room, telephone, online, absentee) and the price at which each dropped out, checked against careful hand counts.
The briefCredit · unclaimed02
Test set of auctioneer speech
Ten hours of auctioneer audio from 20 public sale videos in at least three languages, every word checked by hand, with a separate score for bid amounts and lot numbers, and the error rates of current speech-to-text models on it.
The briefCredit · unclaimed
Open 01 · the brief
Count who actually bid
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For each lot in a public sale video, a count of the separate bidders by channel (room, telephone, online, absentee) and the price at which each dropped out, checked against careful hand counts.
Why it is still open
Ponomarev and Wen write that "the exact number of bidders in each auction cannot be reliably identified from the auction videos", so they bounded it from transcripts, using the names auctioneers call out for telephone specialists and words such as "online bidder", "sir" and "madam".
Built with what exists
The houses' public livestream recordings; ElevenLabs Scribe v3 or another speech-to-text model that tells speakers apart (see the Note "Fewer word errors"); the Tesseract text-recognition program for the on-screen bid display; Claude Fable 5.2 to join speech and screen into one timeline (see the Note "Three working devices").
First step
Hand-count the bidders on 20 lots of one recorded evening sale, noting how each bidder was told apart.
Done looks like
500 lots from at least 20 sales with exact bidder counts for at least 80% of them, matching two independent hand counts on a 100-lot sample, published with timestamps so anyone can check them against the video.
Open 02 · the brief
Test set of auctioneer speech
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Ten hours of auctioneer audio from 20 public sale videos in at least three languages, every word checked by hand, with a separate score for bid amounts and lot numbers, and the error rates of current speech-to-text models on it.
Why it is still open
The bid study above used machine transcripts only to set bounds on bidder numbers, and published speech-to-text scores come from other kinds of speech: the site's Note "Fewer word errors" reports that ElevenLabs measured Scribe v3 on read speech. Auctioneers speak fast, repeat numbers and switch currencies, and no test set of their speech was found in the search for this file.
Built with what exists
Public livestream recordings; a spreadsheet of timed segments for the hand transcripts; Scribe v3 (see the Note "Fewer word errors") and at least two other speech-to-text models.
First step
Transcribe 30 minutes of one sale by hand and measure one model's errors on it.
Done looks like
Word error rate and bid-amount error rate for at least three models over the full ten hours, with the hand transcripts released for reuse as far as the videos' terms allow.