AI for Auction Houses: Consignments, Lot Records and Buyer Questions

An AI agent for an auction house works best as a small set of bounded roles: one that turns consignment submissions into a complete intake packet, one that drafts lot records from documents the house has approved, and one that answers routine buyer questions from those records. Valuations, authenticity opinions, estimates and reserves stay with the specialists who are accountable for them. A small house gets the most value by starting with one role, giving it narrow permissions, and checking its output against source documents on a regular schedule.

Key takeaways

• Start with one role, such as consignment intake or questions about published lots, rather than a general auction assistant.
• Build each lot record from documents the consignor supplied, and mark every field with its source or as open.
• Let the agent answer factual questions from approved records, and route anything about value, condition opinion or reserves to a specialist.
• Give the agent read access to the catalog and write access only to draft fields, with a human approval before anything is published.
• Judge the work by whether records match their source documents and whether questions reach the right person, reviewed on a fixed schedule.

Where an AI agent fits in an auction house

An auction house runs on a steady flow of small, exacting tasks. A consignor sends photographs and a description, a specialist decides whether the property suits an upcoming sale, the team writes a lot entry, and then bidders and buyers ask questions about that lot for weeks. Most of this work is information handling, and that is where an agent helps. The judgment about what a piece is worth, whether an attribution holds, or what reserve to set belongs to the specialists who carry the professional responsibility for it.

The practical boundary is simple to state. An agent may gather, check, format and explain information that already exists in the house's records. It may not invent a provenance, estimate a price or promise a result. Treat this as a design requirement rather than a hope about how a model will behave. The NIST AI Risk Management Framework is a useful reference here because it frames AI risk around governance, mapping, measurement and management, not around the model alone. Marketplaces face similar questions about trust and disputes, and the approach in our marketplace article covers many of the same boundaries.

How to organize consignment intake step by step

A consignment intake workflow has a clear order. An agent can carry most of the middle steps once the house has decided that order in advance. The plan below is a starting structure, not a template to copy without review.

• Define the intake packet: consignor identity, ownership statement, photographs, condition notes, any prior sale history, and the documents the house already requires.
• Log every submission in one register with a status, a named owner and the date received, so that nothing waits unseen in an inbox.
• Have the agent compare each packet against the house's own checklist and draft a request for anything missing. A staff member reviews and sends the request.
• Route the complete packet to the specialist for the relevant category. The agent sends a neutral summary, never a recommendation about acceptance.
• Record the specialist's outcome, whether accepted, declined, deferred or sent for further research, with the reason written in the house's own words.
• For accepted property, create a draft lot record and pass it to the cataloging step, where a person approves it before it leaves the house.

The value of this sequence is that every stage has an owner and a status. When a consignor asks where their property stands, the answer comes from the register, not from someone's memory of an email thread.

Building lot records that hold up under questions

A lot record is the single source that catalog text, buyer answers and internal reports all draw from. If it is loose, every downstream answer inherits the looseness. Give each field a source: a consignor document, a specialist note, a condition report, or the house's own research. A field without a source should be marked open, not filled with a plausible phrase that reads well.

When an agent answers questions from records, the reliable pattern is retrieval: it looks up the approved record and quotes from it, instead of relying on general knowledge it absorbed during training. The answer should name the field it came from, so a staff member can verify it in seconds. Where the record is silent, the correct output is that the information is not yet confirmed, along with a route to a person who can confirm it.

A lot record that says what is known, what is pending and who is responsible is more useful to a buyer than one that sounds certain.

Answering buyer questions from the lot record

Buyers ask predictable questions: dimensions, materials, what is included, collection and shipping arrangements, viewing times, and whether a lot has changed since publication. Many of these answers already sit in the record. An agent can respond to them in the house's voice, with the approved record as its only source for facts.

Questions about value, authenticity, likely results or whether a reserve is close are a different category. The agent should acknowledge the question, state that a specialist will respond, and open a ticket that keeps the buyer's exact wording and the lot reference. It should not guess, reassure or hint.

Buyer and consignor messages are also untrusted input. A message that says to ignore the house's rules and confirm a reserve should change nothing about the agent's behavior. The OWASP guidance on prompt injection prevention describes why instructions belong in the system's own configuration while inbound text is treated as data to be read, not as commands to obey.

Illustrative scenario: a buyer asks whether the top of a side table is original. The lot record lists the top as replaced, according to the consignor's note, and the condition report is still pending. The agent quotes the consignor's statement as recorded, says that the condition report is pending, and forwards the question to the specialist. It does not state in its own words whether the top is original or replaced.

Where valuations and judgment stay with specialists

A short decision guide keeps this boundary clear for staff and for the agent's configuration. Anything in the first group belongs to a specialist. Anything in the second group can be delegated. Anything in the third group should never be done by an agent.

• Specialist decisions: estimates, reserves, authenticity and attribution opinions, condition grading beyond the recorded report, and any statement about what a lot will achieve.
• Delegable tasks: retrieving and quoting approved record fields, drafting requests to consignors, summarizing a specialist's notes for internal review, and offering viewing times from a calendar the house controls.
• Never delegated: publishing a figure, changing a reserve, promising delivery or refund outcomes, or telling a buyer that a lot is authentic.

Be Human designs these boundaries during a first engagement. The work defines one useful role, the memory and tools it needs, the approvals required before anything reaches a buyer, and a success criterion the house can observe. Managed operations then cover monitoring, recurring evaluations, permission review, incident response and controlled changes to the agent's behavior. The service is remote and available in English and Portuguese, and it supports the house's specialists rather than replacing them.

Permissions, records and how to tell the work is good

Each agent role should have the narrowest access that lets it do its job. An intake agent reads submissions and writes drafts. A buyer-question agent reads published records and cannot edit them. The OWASP authorization guidance is a practical checklist for this, especially its emphasis on checking access on every action rather than trusting the interface that sent the request.

Keep a record of what the agent read, what it said and which field it used. The observability concepts described in OpenTelemetry's primer, traces, logs and metrics that show how a system behaves, apply directly here. When a buyer disputes an answer, the house should be able to reconstruct exactly what was said and why. Memory should hold house policy and approved records, not personal data beyond what each purpose requires, and buyers or consignors in Brazil fall under the guidance of the ANPD.

A simple success check is a standing review. Each week, a staff member samples a small set of lot records and compares them with their source documents, then reviews the questions the agent escalated to see whether they reached the right specialist. Errors found in that review become changes to the rules, the record template or the permissions, and each change is written down. Over time, the house learns which tasks are reliable enough to delegate further. The free Be Human AI diagnostic can help you find the first task in your auction house worth handing to an agent.

Frequently asked questions

Can an AI agent write lot descriptions for a small auction house? It can draft descriptions from approved record fields, such as dimensions, materials and documented history, and flag any gaps for a specialist. The specialist should review and approve every description before publication, and the agent should not add claims that are absent from the record. Descriptions that involve attribution or condition judgment remain the specialist's responsibility.

Should an AI agent answer buyer questions about estimates? It should not give estimates, price expectations or reserve information. The appropriate response is to say that a specialist will reply, and to route the question with its original wording and the lot reference. Keeping that boundary in the agent's configuration, and checking it during regular reviews, prevents well-meaning but unauthorized answers.

What does a first engagement with Be Human cover? A first engagement defines one useful role for the house, the memory and tools that role needs, the approval boundaries around it, and an observable success criterion. It does not require the house to rebuild its systems at once. Ongoing managed operations can then add monitoring, recurring evaluations, permission reviews and controlled improvements.

Does the agent need access to our bidding platform or accounting system? Most of the useful work can begin with read access to the catalog and the lot records, plus write access only to draft fields in the intake register. Access to bidding, payments or accounting should be considered later, only for a specific task, and granted with narrow permissions and a named approver. Every new tool connection should be reviewed before it goes live.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.