Vision · 16 September 2026 · 6 min read
Artificial intelligence & auctions
A manifesto on artificial intelligence in auction houses: the uses worth building – finding the right buyers for each lot, cataloguing, translating, searching the archive – and the red lines that protect trust.
By Romain Kohn
Improving the craft. Preserving trust.
Building the uses together with the people who run the sales.
01
Why now?
Auction houses already have what matters most: expertise, client relationships, archives, images, catalogues, results and a knowledge of the market built up over the years. What artificial intelligence mainly changes is our ability to connect these elements.
Until now, much of this information was recorded, filed and then consulted. It could be found if you knew exactly where to look, or if someone remembered it. AI now makes it possible to link an object to old archives, to understand a search phrased in different words, to compare images, to surface potential buyers or to prepare a first draft from multiple sources.
So the question is no longer just "What can AI do?" It becomes: "Which uses do we want to build, and which limits do we want to set?"
02
Finding the right buyers for each lot
One of the most immediate opportunities AI offers is to start from a specific lot and identify the people most likely to be interested in it. The lot's public data — description, images, artist or maker, period, materials, dimensions — can be turned into mathematical representations that make it possible to measure similarity. These signals can then be matched against verified information held by the house: purchase history, bids, interests or past interactions.
The system is therefore not looking for a generic category of buyers. For this particular object, it looks for the people whose history and interests make this lot especially relevant — whether they are active buyers or clients who have not been in touch with the house for a long time.
What is at stake is as much the relationship as the sale: putting the right object in front of the right person, at the right time, can be far more useful than sending more and more messages to an entire database.
Fewer messages. More relevance. And the chance to revive relationships that no one would have thought to pursue manually.
03
Describing a lot from the first encounter
A photo and a voice note can become the raw material for cataloguing work, while preserving the memory of the first look. The specialist photographs the object, dictates freely what they see and, if needed, adds a few notes. The audio is transcribed; the images can be analysed; visually similar objects can be found; internal archives and external sources can be consulted.
From these elements, AI can suggest a title, a structured description, characteristics, leads for identification and possibly an estimate range. The benefit is not only the time saved. The photographs, the voice note and the initial observations are kept: a record remains of the thinking at the exact moment of the first encounter with the lot.
AI prepares. The specialist validates. The original source always remains accessible.
04
Campaigns built around the objects
Traditional segmentation starts from people. AI makes it possible to start from the lots and look, for each one, for the most relevant people. Across an entire catalogue, every lot can be compared with buyers' observed interests. A dynamic relevance threshold then keeps the number of recipients under control.
Two clients may receive the same campaign for entirely different reasons: one because three lots match an artist they follow; the other because a group of objects resembles their previous purchases.
The right measure is not the number of emails sent, but the quality of the connections made.
A campaign is no longer simply aimed at a population. It becomes a connection between a catalogue and the people for whom it genuinely contains something of interest.
05
Truly personalising a message
When a member of staff writes to a single person, AI can suggest a message suited to the real context of that relationship. The tone, the language, the history of the relationship with the house, the relevant objects or the exact reason for getting in touch can all be taken into account to prepare something far more personal than a generic template.
This use does, however, mark a clear boundary with anonymised or pseudonymised recommendation mechanisms: personalising a message for one person requires, by definition, more personal context. This difference must be explicit and controlled. The system should use the minimum information necessary and let the house decide when this kind of personalisation is appropriate.
Personalisation is only valuable if it remains controlled, transparent and proportionate.
06
Translating the catalogue without losing the expertise
The market is international; editorial work remains local. Language models can bring the two together. Descriptions, catalogue notes, communications or editorial content can be translated quickly, while taking into account specialist vocabulary and the house's own glossary.
But translation must not smooth away careful wording, attribution caveats or phrasing that carries legal or commercial weight. The right approach is therefore not blind machine translation, but translation that is generated, traceable and verifiable.
The goal is not just to translate the words, but to preserve the intent and the level of certainty.
07
Beyond generation
The most far-reaching uses are sometimes the least spectacular. They concern memory, search and data quality.
-
The house's memory
Finding similar objects, past descriptions, estimates and results, even when the vocabulary used at the time was different.
-
Semantic search
Searching by idea rather than by exact words: "a 1970s sports watch", "a piece of furniture in the spirit of…", "lots similar to this one".
-
Visual search
Comparing a photograph with hundreds of thousands of images to identify precedents, duplicates, matches or visually similar objects.
-
Data quality
Detecting duplicates, inconsistencies, name variants, missing fields and information present in a description but absent from the structured data.
-
Estimate support
Gathering comparables, summarising historical results and proposing a reasoned range without turning a suggestion into a verdict.
-
Administration
Summarising a request, filing an attachment, finding a file, preparing a reply or extracting repetitive information to give time back to the teams.
08
The red lines
Auctions rest on trust. Uses of AI must therefore be judged as much on what they enable as on what they refuse to do.
-
01
Never invent to fill a gap
An unknown provenance remains unknown. An uncertain attribution remains uncertain. A hypothesis must be presented as such.
-
02
Always be able to go back to the sources
When information comes from a catalogue, an image, a note, an external database or a historical result, the source must remain identifiable.
-
03
Never erase the original
Generated content never replaces the photo, the voice note, the description or the human observation that preceded it.
-
04
Keep sensitive decisions human
Attribution, final approval of a sensitive description, the definitive estimate or any decision that engages the house's responsibility remain under human control.
-
05
Design for privacy from the start
Client data is not a free raw material. When a task can be done without identity, or from pseudonymised representations, it must be.
-
06
Be able to say "I don't know"
A credible system must be able to express its level of uncertainty, ask for validation or decline to conclude when the available evidence does not allow it.
09
Building this transformation with the profession
Some applications will become obvious. Others will prove useless. And the most important ones have probably not yet been imagined.
That is why this change should not be defined solely by software vendors or technology companies. It must be built with auction houses, specialists, auctioneers, cataloguers and the staff who look after the relationship with sellers and buyers every day.
Technology can offer new possibilities. The profession must decide which ones are genuinely valuable. Artificial intelligence will not replace what makes a good auction house. It has neither its history, nor its network, nor its expertise, nor its reputation, nor the trust built with its clients.
But it can help that house make better use of everything it already has: its knowledge, its archives, its images, its data, its relationships and, above all, its teams' time.
10
This manifesto is an invitation.
-
01
Which uses seem genuinely useful to you?
-
02
Which limits should we set together?
-
03
Which applications of AI in auctions have we not yet imagined?