Superintelligent Agents for Auction Houses: Consignments and Lots
Superintelligent agents for auction houses: consignment intake, lot description drafts and buyer questions, with provenance and valuation left to specialists.

Superintelligent agents for auction houses are AI agents that take on the repetitive, document-heavy work around consignment intake, lot description drafts and buyer questions, while specialists keep authority over provenance claims and valuation statements. The practical starting point is one narrow role with a defined document set, explicit approval points and a success measure the operations lead can check every month. This guide explains how to scope that role, where the agent must stop and how a managed setup keeps it under control.
Key takeaways
- Start with one role, such as consignment intake triage, and keep valuation and provenance judgments with specialists.
- Treat every consignor document as untrusted input, including text that appears inside scans, emails and photographs.
- Let the agent produce drafts and flags, never published estimates, attributions or authenticity opinions.
- Tie each factual sentence in a lot description to a document or specialist note so review takes minutes instead of an hour.
- Judge success by specialist review time, missed gaps and real discrepancies caught, not by how many items the agent touched.
Where an agent fits in consignment intake
Consignment intake is the stage where a potential seller offers an item or a collection to the house. The work includes receiving the offer, collecting photographs and measurements, asking for ownership and acquisition records, noting known damage and passing the file to the right department. A large share of the effort goes to chasing missing items and retyping details from emails into catalogue or client systems. That repetitive layer is where an agent is useful.
The agent should not decide whether the house accepts a consignment, what reserve to propose or what an item might realize at sale. Those are commercial and expert decisions. The agent's job is to make each file complete, consistent and easy for a specialist to judge, and to state clearly what is still missing. Houses that also handle repairs and custom orders face similar custody questions, which are covered in Superintelligent Agents for Jewelry Stores: Repairs and Custom Orders.
Planning the first role, step by step
A first engagement should be small enough to measure within a few weeks. A workable sequence looks like this:
- Name the department and one intake task, for example, triaging new consignment offers by category and completeness.
- List every document type the agent may read, such as photographs, condition notes, invoices, ownership letters and prior sale records, and mark which ones are optional.
- Write the forbidden outputs in plain language: no accept or decline decisions, no reserve suggestions, no valuation ranges, no authenticity or attribution opinions.
- Define the handoff. The agent creates a structured intake summary and a list of missing items, then places both in a specialist queue. It does not discuss pricing with the consignor.
- Set permissions narrowly: read access to the intake folder, write access only to catalogue drafts, no publishing rights and no access to payment or bank details.
- Choose two or three checks, such as whether specialists agree with the missing-items list and how often the agent's discrepancy flags turn out to be real.
- Review permissions and instructions before the role expands to a second category of lots.
Lot descriptions: drafts a specialist can check
Catalogue copy is where small errors become public claims. A lot description drafted from a specialist's notes, a condition report and the consignor's documents saves typing, but only if every statement has a visible source. The agent can organize the facts in the house style: dimensions, materials, marks, a condition summary and a list of the documents it used.
The drafting rules should be strict. The agent uses measured facts from the record, quotes the condition report wording rather than paraphrasing it into something stronger, and marks unverified statements as unverified instead of smoothing them over. It should not use words such as rare, important, original or investment unless a specialist has written them into the file and approved them for publication. Any wording that attributes a maker, period or origin belongs in the specialist's own sentence.
A practical format is a short claims list placed above the draft. Each claim carries its source, such as the condition report, a named document or a specialist note. The specialist reviews the claims first. If a claim has no source, it is removed from the description before anyone reads the prose.
Provenance and valuation: where the agent stops
Provenance is the recorded chain of ownership and history that a buyer may rely on. Valuation is a judgment about likely sale value. Both carry legal and reputational weight. An agent can help by listing which provenance documents were supplied, which were promised but not received, and which statements in the draft depend on a document nobody has seen. It should never assemble a provenance narrative from memory or from a consignor's verbal account and present it as established.
Illustrative scenario: a consignor emails photographs of a pair of ceramic vases and writes that they came from a well-known estate. The intake agent records the email as a claim, notes that no estate documentation is attached, adds the request for documents to the missing-items list and flags the statement for the specialist. The draft description says nothing about an estate until a specialist decides what the file supports.
Consignor documents also need a security mindset. A scanned letter or a PDF may contain text written to steer an AI system, for example an instruction to mark the item as accepted. The agent must treat that text as content to be read, never as a command to follow. The OWASP guidance on prompt injection prevention describes why instructions and untrusted data should be kept separate, and that separation should be built into the intake design from the start.
Buyer questions without new promises
Buyer questions tend to repeat: is an item in the sale, what is its condition, can it be viewed, how does shipping work, how do I register to bid. The agent can answer these from the published catalogue, the approved condition report and the house's posted policies, and it should say when it is quoting from them. Anything about condition beyond the report, provenance, authenticity or likely outcome goes to a specialist with the original question attached. The agent never confirms a reserve, promises a result or offers an exception.
Disputes follow similar patterns in any marketplace where trust depends on accurate listings. The approach described in Superintelligent Agents for Marketplaces: Trust, Disputes, Sellers applies here as well: the agent gathers facts and routes decisions, and people with authority decide.
Controls, monitoring and how Be Human runs it
Be Human designs the agent around the house's own rules: the style guide, the approved phrasing for condition and provenance, the document register, the tools it can call, and the approval points where a specialist must act. In managed operations, monitoring watches the intake queue for stalled files and unusual volumes. Recurring evaluations sample drafts and compare them against specialist corrections, so drift shows up early. Permission reviews remove access that is no longer needed, and incident response handles cases such as a document that should not have been ingested or a draft that reached the wrong channel. Improvements are made in controlled steps and tested on past files before they touch live catalogue work.
The NIST AI Risk Management Framework is a useful reference for deciding who reviews what and how performance is measured, and the OWASP authorization guidance supports the habit of granting each component the least access its task needs. If you want to find the first intake or catalogue task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an agent write the final catalogue description for a lot? No, the agent can prepare a draft with a claims list and open questions, but a specialist approves every statement before publication. The draft should never introduce attributions, period language or condition claims that the file does not support. Publishing rights stay with people who are accountable for the catalogue.
Should the agent have access to consignor bank or payment details? No, those details are not needed for intake triage or description drafts, so the role should not be able to read them. Narrow permissions reduce exposure if a document is mishandled or an integration fails. Permission reviews should confirm this boundary on a regular schedule.
What happens if a consignor's document contains instructions aimed at the agent? The agent treats the text as content to summarize and flag, not as an instruction to act on. Suspicious text is recorded in the intake summary so a specialist can see it. The design should separate the agent's instructions from the documents it reads, following established prompt injection guidance.
How does the house know the agent is working well? Measure it against the specialist's judgment, not against volume. Useful checks include whether missing-items lists match what specialists later find, how often flagged discrepancies prove real and how long specialists spend correcting drafts. Recurring evaluations and a clear incident process make these results visible to the operations lead.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


