Short answer: accurate enough that review replaces writing, not accurate enough that you skip review. According to Gavelist production data through September 22, 2026, 96.3 percent of 58,134 lots exported by Gavelist clients went out with no client edit to the AI-written title or description. One client supplied about 70 percent of those lots at 99.8 percent; across every other Gavelist client the figure is 88.2 percent, and in September alone 97.7 percent of 11,040 exported lots went out unedited. That number is a measurement, not a survey: a lot counts as edited if the client saved a change to its title or description in Gavelist before exporting; lots the AI regenerated count as unedited. The total includes 6,290 lots backfilled from jobs created March 20 to April 12.
Published August 28, 2026 by the Gavelist team. Updated September 22, 2026: the production figure now runs through September 22, states the single-client concentration, and describes the method correctly.
What the unedited rate does and does not mean
It means that across Gavelist clients other than the largest, about 88 of every 100 lots went out with the AI's title and description as written (about 96 of 100 including that client). It does not mean the AI is an appraiser. Identification is pattern recognition from your photos: maker's marks, materials, form, condition. It is very good when the photos show the evidence and honest about uncertainty when they do not. Attribution and authentication - is this Rookwood or a reproduction, is the signature real - remain human calls, the same way they always were.
The biggest accuracy lever is photo coverage, not the model. Gavelist reads every photo attached to a lot together in one pass, with no per-lot cap - so the backstamp on the bottom and the chip on the rim make it into the description because they were in the input. A single hero photo cannot carry that information no matter whose AI reads it.
What happens when it gets something wrong
You are reviewing before export, so a miss is an edit, not a published error. Corrections take seconds in the review view, and they are not wasted work: Gavelist learns your title and description preferences from the edits you make, so the same correction gets rarer over time. There is no retraining project and nothing to configure - the learning is a byproduct of the review you were doing anyway.
The failure modes worth knowing in advance: lookalike materials (plate vs sterling when the hallmark is not photographed), reproductions with faithful marks, and items whose value lives in provenance the photos cannot show. Those are exactly the lots you were always going to look at twice.
How much time review actually takes
Manual cataloging runs about 8 to 15 lots per hour when photography, writing, and upload are all counted - according to Sound Auction Service in Washington state, full lot preparation is priced at $3.00 per lot. A review pass over AI descriptions is a different activity: skimming a title and description you did not have to write, at reading speed. Operations processing 100 lots a day report the review pass in minutes-per-batch, not minutes-per-lot, with the feature lots getting the real attention. The honest statement is that the writing time disappears and a smaller checking time replaces it.
Will the catalog lose your voice?
Only if you let it ship generic. Voice settings set your house style directly - phrasing, length, what to lead with - and the edit-learning pulls the output toward how you actually write. The unedited rate is measured across clients using exactly those controls.
FAQ
Do I need to review every single lot? You need a review pass; you do not need a writing pass. Most users read every feature lot and skim standard and box lots. The measured edit rate - about 12 percent of lots across Gavelist clients excluding the largest, 3.7 percent including it - is a reasonable estimate of how often you will touch anything.
Can it detect condition issues like cracks and chips? If the damage is visible in a photo, it belongs in the description - which is an argument for photographing damage deliberately, the way you already do for disclosure.
Does accuracy depend on category? Yes. Marked, mass-produced, and well-documented items identify best. Unmarked art, unsigned furniture, and anything whose value is attribution-driven need the human eye. For how value estimates layer on top of identification, see AI that identifies an item from a photo and tells you what it's worth.
How does this compare to hiring a cataloger? At market labor rates you are paying about $3.00 per lot for manual cataloging. Gavelist descriptions are $0.15 per lot, and the review time is yours either way - you were proofing the cataloger's work too.
Sources
- Gavelist production data: export snapshots across 58,134 client-exported lots, jobs from March 20 to September 22, 2026, including 6,290 lots backfilled from jobs created March 20 to April 12 (first-party).
- Sound Auction Service, published $3.00/lot full-preparation rate. soundauctionservice.com