An auction house can use AI to turn verified lot records into catalogue drafts and send clearer consignor updates, provided a specialist checks every attribution, provenance statement, condition description, photograph and sale term. Keep estimates, reserves and authenticity decisions outside the model. Use event-based email templates only after the authoritative lot status is confirmed by a person.
The lot record is the source of truth, not the generated paragraph. A model can combine fields fluently while quietly changing “possibly” to “by”, treating a repaired surface as original, or calling a scheduled sale a completed one. Separate facts, specialist opinions and operational status before asking AI to write anything a consignor or buyer will see.
Build a source packet for each lot
Create a structured record with lot number, object type, dimensions, materials, inscriptions, photographs, consignor-provided history, specialist observations and the approved sale terms. Mark each field as measured, inspected, documented, supplied by consignor or unverified. Those labels matter. “Signed on base” may be an observation; “made by the named artist” is an attribution. A transcription of an old receipt is not proof of every statement printed on it. If the catalogue system already stores these fields, export only the reviewed fields needed for drafting.
Photographs need their own check. Record which image is the front, reverse, detail of signature, damage and scale, and which image is approved for the public listing. An AI model may misread a shadow as a crack or miss a repair visible in a close-up. Do not ask it to infer condition or authenticity from the image. Use it to map the photographer's file names to the lot record, then have a specialist confirm that the published gallery shows the correct object and the relevant details.
An illustrative ceramics lot might have a measured height of 24 cm, a stamped mark under the base, two photographs of a repaired rim and a consignor note saying it was bought decades ago. The source packet should keep those as separate entries. It must not promote the stamp into a maker guarantee or the recollection into a continuous provenance chain. The antique provenance tutorial explains why an attractive attribution should remain a question until evidence supports it.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowMake provenance an evidence trail
Provenance means an object's ownership history. List each claimed transfer with date, prior owner, evidence and uncertainty. A gap stays a gap. Do not tell AI to “fill in a likely history”; it will produce a narrative that looks catalogued but has no documentary basis. A specialist may decide that a partial history can be stated with careful wording, or that it should be omitted. Preserve the consignor's original statement and the approved public version so the team can see the difference.
Suppose a consignor says an item belonged to a relative and supplies a receipt dated 1986. A safe internal record says: “Consignor reports family ownership; receipt dated 1986 supplied; earlier ownership unverified.” An unsafe AI draft says: “From the family's collection since its creation, with impeccable provenance.” The receipt supports one date, not an uninterrupted history. The specialist should ask whether the receipt can be matched to this exact object, then approve a precise statement or no provenance line at all.
Apply the same discipline to estimates. The auctioneer or specialist determines and approves an estimate under the house's process; AI may copy that approved range into a draft, but should not calculate a new valuation from a photograph or pull one from unrelated online sales. A lot can be similar in appearance and different in condition, date, attribution or market. If the estimate changes, update the structured record first, then regenerate dependent copy. Avoid editing the catalogue and email separately, which leaves different figures in circulation.
Draft a catalogue entry with explicit unknowns
Give AI a compact style brief and only approved fields. Request a neutral description, not a sales pitch. Include a rule that “unknown” remains unknown and that condition sits in a separate reviewed report. The prompt can ask for a table of statements that need a specialist check, which gives the reviewer a practical queue. Keep the output short enough to compare with the source packet line by line. A long imaginative paragraph is harder to audit and may bury the one invented claim that matters.
Write one catalogue draft for lot [number]. Use only these approved fields:
[object, measured dimensions, materials, inscriptions, approved attribution].
Do not infer maker, age, authenticity, provenance, condition or value.
Keep qualifiers exactly as supplied. If a field is unverified, omit it.
After the draft, list every statement that needs specialist confirmation.
Do not write a condition report or a sale estimate.Given those fields, the draft could come back (illustrative) as “A glazed ceramic vessel, 24 cm high, with a stamped mark beneath the base.” If the source packet says only “mark resembles X”, the output must not name X as maker. A reviewer can improve the entry with approved period or school wording, but must preserve the evidence status. For an unusual lot, the best description may be only two factual sentences plus an invitation to request a condition report. More adjectives do not add verified information.
Catalogue copy and condition reports serve different jobs. The catalogue entry identifies and describes the lot; a condition report, where the house offers one, records the object's physical state in more detail and is written from an inspection. In a small house, that suggests separate approval fields and separate versions. If a buyer asks about restoration, staff should refer to the reviewed condition material or ask a specialist; AI should not paraphrase a catalogue sentence into a condition assurance. The listing accuracy guide offers a wider check for wording that might mislead a buyer.
Attach a condition note to observations and images
Condition language should distinguish what was observed from what was not inspected. A specialist might note a repaired rim, surface wear and a visible crack. AI can format those notes into consistent sentences, but it must not add “otherwise perfect” or “fully restored” without an inspection supporting those claims. If the object has inaccessible areas, use the house's approved qualified wording. Where a condition report includes images, check that each annotation points to the correct location. A mistaken image reference can be worse than a clumsy sentence.
Suppose a specialist's notes say, “small chip at foot; old repair at handle; interior not examined.” A bad draft says, “Minor age-related wear, in excellent condition.” The correction is a plain report of the noted defects and inspection limit, subject to the specialist's final wording. Have a second person compare the condition note with the image set for high-value or fragile lots. The same second person can confirm dimensions and lot number, two simple fields that can cause expensive confusion when accidentally copied from the adjacent record.
Keep a version history after publication. If a new inspection reveals a defect, the specialist decides the revised description and the auction team follows its established notification procedure. Do not silently change only the website text while a downloadable catalogue or pre-sale email still carries the old statement. Record when the change was made and who was notified. AI can draft consistent update wording from the approved correction, but it should not decide whether the correction is material under the house's conditions.
Use a sale-status board for consignor updates
Consignors need to know what has happened, what remains open and when they will hear next. Create a status board with received, photography complete, specialist review pending, catalogue approved, sale scheduled, sold, unsold, payment pending and settled, adapted to the house's real process. Each status has a named owner and approved message. Do not let AI infer status from a partial email thread. The lot system or staff confirmation should supply the current state, and an exception flag should block automatic sending.
An illustrative pre-sale update reads: “Hello [first name], photography for lot 42 is complete. The catalogue description is awaiting specialist approval. We expect to send the approved listing for your review after that step. No estimate or sale outcome has changed.” This tells the consignor what is known and what is pending. A weaker AI draft, “Your item is ready to sell and attracting strong interest,” invents readiness and market response. If the sale date is not final, the template should not insert one.
Use a separate message after the auction. A sold lot may still have payment, collection and settlement steps; a hammer result is not the same as money available to the consignor. For an unsold lot, the next message should explain the house's actual options without implying a guaranteed private sale. For a withdrawn lot, suppress the ordinary sale-result email completely. These branches need tested triggers. The buyer and vendor follow-up tutorial covers broader sequence design, but an auction's status and financial gates require extra care.
Reply to buyers from approved sale facts
Buyer emails often ask several questions at once: whether a lot is still available, whether a report exists, how to view it and what the charges are. AI can identify those questions and draft a response using approved conditions of sale and the current lot status. Staff should check that the answer refers to the right lot and sale. Do not let a model promise shipping, guarantee authenticity, interpret a condition report or negotiate on behalf of the house unless a person has specifically approved the response.
A typical buyer email (illustrative) reads, “Is lot 42 original, can you guarantee the mark, and can it arrive by Friday?” A safe draft for staff review says: “Thank you for asking about lot 42. We can send the available photographs and arrange for our specialist to answer your attribution and condition questions. Our team will confirm shipping options separately.” It does not evade the questions; it routes them to the people with evidence and authority. An unsupported “Yes, it is original and delivery is guaranteed” must never leave the inbox.
Keep an answer bank for routine questions: viewing times, registration steps, where the conditions of sale are published and how to request a condition report. Check each answer against the current sale. A phrase copied from a previous auction may give the wrong deadline or location. When a buyer asks for a comparison between two lots, have the specialist review any claim about relative condition or rarity. AI can organise the questions, but a fluent comparison should not be mistaken for an inspection.
Count corrections in a 60-lot pilot
Imagine a small auction house preparing 60 lots. These numbers are illustrative. Staff choose ten lower-risk lots for an initial pilot. The source packets take six minutes each to assemble, or 60 minutes. AI produces a first draft in one minute per lot, or ten minutes. Specialist review takes four minutes per draft, or 40 minutes, and photography matching takes three minutes per lot, or 30 minutes. The pilot therefore uses about 140 minutes of recorded work. Compare that with the house's previous end-to-end time for ten similar lots; do not compare only the minute spent generating text.
Suppose the reviewer makes 17 corrections across those ten drafts: five dimension or material edits, four qualifier changes, three image mismatches, three unsupported condition claims and two invented provenance phrases. That pattern shows the model may help with sentence structure but cannot be trusted to decide evidential claims. Move the qualifier and provenance fields into a protected source sheet, tighten the prompt, then repeat on a different ten lots. If corrections remain concentrated in those fields, keep them out of generated copy entirely.
The house can also test 15 consignor updates. Count wrong statuses, confusing next steps and messages held for exceptions. If two “sold” messages would have gone to withdrawn lots, stop automatic sending until the trigger is fixed. A successful pilot has fewer corrections and clearer communications, not merely more words produced. Keep buyers and consignors out of the test until staff have reviewed the drafts in a shadow run. The shadow pilot guide gives a general way to observe an automation before it reaches customers.
Release each lot through three approvals
Before publication, the cataloguer checks the lot record and copy against measurements and approved fields; the specialist checks attribution, provenance and condition wording; the sale manager checks estimate, reserve handling, schedule and terms. Assign names and dates to each approval. A single “looks fine” tick makes it hard to discover who checked which risk. A simple table can track lot number, photo match, evidence status, condition version, estimate approval, email status and publication version.
A filled illustrative release entry might say: “Lot 42; six images matched; stamped mark described but maker unverified; condition report version 2 approved; estimate copied from specialist record; consignor update held pending catalogue approval.” That entry gives the next staff member a useful hand-off. It also prevents an AI-generated update from implying that the lot is approved when a material decision remains open. Buyers should receive only the version and status that the team has released.
Handle split lots, late changes and duplicate photographs
Real catalogues rarely follow a neat one-object-one-record pattern. A box of twelve pieces may be catalogued as one lot, while a consignor's collection may be split after inspection. The source packet should state the count and identify which photographs belong to the final lot. If a split changes lot numbers, regenerate descriptions from the new records and check every saved email draft. Otherwise a buyer can be shown a piece that has moved to another lot, or a consignor can receive an update about the wrong group of objects.
Consider an illustrative lot of six silver spoons, photographed together. After the specialist review, two are separated because their marks differ. The original AI description, “Six matching spoons”, must be withdrawn. New records should say exactly which four remain and which two moved, with each photo reassigned. A search-and-replace of the number six is insufficient: the word “matching” may now be wrong, the measurement may describe only one spoon and the estimate may need another approval. Treat the split as a new release, not a minor copy edit.
Late changes need a channel list. A revised condition note can affect the public lot page, a downloadable catalogue, printed notes, staff answer bank and queued buyer messages. The sale manager should record which channels were updated and which people received the correction. AI can draft a consistent notice from the approved change, but it cannot determine who must be told or whether bids need special handling. Use the house's conditions and specialist judgement for that decision. In the release table, mark the old version superseded and keep its text for the audit trail.
Duplicate photographs are another useful test. Two similar vases on a grey background may be assigned the same image filename pattern. Compare an identifier visible in the image, the reverse view and the measured dimensions against the lot record. If the match remains uncertain, hold publication. A beautiful AI caption attached to the wrong vase is still a wrong listing. A second person checking the gallery is often a better investment than another round of prompting.
Keep private sale information out of public drafts
A consignor record can contain reserve discussions, personal contact details, acquisition documents and internal valuations. Give the drafting tool a public-field export, not the entire record. This reduces accidental disclosure and makes a reviewer more confident about what the model could have used. A lot-number token can connect the draft to the internal system without including the consignor's identity. If a tool offers shared workspaces, check who can open prompts and files before uploading any private material.
For example, an internal note may say, “Consignor will accept a lower reserve if another lot sells.” That sentence is useful to the sale manager but has no place in a catalogue description or buyer response. A public-field export should exclude it by design. If a generated draft mentions a reserve or private negotiation, stop and inspect the input path: either the wrong data was exported or the model invented the statement. Correcting the sentence alone leaves the underlying process problem in place.
Ask the person approving each email to compare its claims with the record as it stands at sending time. An approved template does not make a stale status correct. This is particularly important after a withdrawal, a postponed sale or a condition update. The team should be able to pause a queued sequence with one action and see every pending recipient. If the messaging tool cannot show that queue, keep sending manually until it can be audited reliably.
Questions about AI-assisted auction copy
Can AI identify an object's maker from a photograph?
It may suggest details to investigate, but a photograph alone is not evidence of maker, age or authenticity. Record the suggestion as an unverified research lead. A specialist should inspect the object and supporting records, decide what can be stated, and approve any attribution before the catalogue goes live.
Should a condition report be generated from the catalogue description?
No. A condition report concerns the object's physical state and needs a specialist's inspection, with supporting photographs and qualified wording. AI can help format verified inspection notes, but should not infer wear or restoration from a general listing. Keep condition reports linked to their own reviewed source notes and version.
Can consignor updates and buyer emails send automatically?
Use approved templates for routine event updates only when they pull dates and status from the authoritative sale record. Hold messages about reserve, estimate, condition, authenticity, payment disputes and shipping exceptions for a person. Test each template with sample lots, including withdrawn and unsold lots, before enabling any sending rule.
Further reads
- AI for Used Car Dealerships: Honest Listings and Photo Clean-Up — See another inventory-led photo and description check.
- AI Email Triage for Professional Firms: Sort, Summarise, Draft — Route complex inbound questions to a person.
- How to Write a Customer Newsletter With AI in 30 Minutes — Plan a separate opt-in sale announcement.
- How to Write Product Descriptions With AI That Actually Sell — Improve clarity while retaining factual constraints.
- What to Do When AI Gets Something Wrong With a Customer — Prepare a correction route for a bad message.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: No third-party figures or rules are quoted. Follow the auction house's own conditions of sale, condition-report practice and specialist review procedure.